fix(neuron): context-limit observability + qwen3_next profile (#126)

- derive_limit now logs every input and intermediate term (free VRAM,
  prefill rate, vram/throughput ceilings, hard clamp, result) at DEBUG
  so a surprising advertised limit is diagnosable from the journal —
  needed right now: beast's 27B advertises context=0 post-reboot and
  no by-hand input combination reproduces it.
- profile_from_qwen3_5_config had the same too-narrow model_type gate
  as #124 and a plain serde parse that rejects the flat qwen3_next
  layout — Coder-Next has never had an advertised limit or the #65
  length-aware preflight. Parse via Config::from_config_json and
  accept both types.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TczcGF7JSjJs8r15RSSGpx

.gitea/workflows/build-prerelease.yml

@@ -66,6 +66,7 @@ jobs:
       build_cortex: ${{ steps.changes.outputs.build_cortex }}
       build_neuron: ${{ steps.changes.outputs.build_neuron }}
       build_bench: ${{ steps.changes.outputs.build_bench }}
+      build_upstream: ${{ steps.changes.outputs.build_upstream }}
       check_rust: ${{ steps.changes.outputs.check_rust }}
     steps:
       - uses: actions/checkout@v4
@@ -86,7 +87,14 @@ jobs:
           # rpmvercmp ranks digit-prefixed segments above alpha ones.
           # The SHA stays only as a debug identifier; sort order is
           # decided entirely by the timestamp.
-          COMMIT_TIMESTAMP=$(git log -1 --format=%cd --date=format:%Y%m%d%H%M%S HEAD)
+          # format-local + TZ=UTC0 renders the stamp in UTC regardless
+          # of the committer's recorded timezone. Plain `format:` uses
+          # each commit's own TZ offset — local commits (UTC+3) and
+          # Gitea server-side merge commits (UTC) interleaved
+          # non-monotonically, letting an older build out-rank newer
+          # ones in RPM EVR comparison (the 2026-07-01 "235959" stamp
+          # that froze the fleet on a stale build).
+          COMMIT_TIMESTAMP=$(TZ=UTC0 git log -1 --format=%cd --date=format-local:%Y%m%d%H%M%S HEAD)
           RELEASE="0.1.${COMMIT_TIMESTAMP}.git${SHORT_SHA}"
           echo "version=${VERSION}" >> "$GITHUB_OUTPUT"
           echo "release=${RELEASE}" >> "$GITHUB_OUTPUT"
@@ -104,6 +112,7 @@ jobs:
           BUILD_CORTEX=true
           BUILD_NEURON=true
           BUILD_BENCH=true
+          BUILD_UPSTREAM=true
           CHECK_RUST=true
 
           if [ "${GITHUB_EVENT_NAME}" = "push" ]; then
@@ -149,6 +158,7 @@ jobs:
               NEURON_RE='^crates/neuron/|^crates/cortex-core/|^Cargo\.toml$|^Cargo\.lock$|^rpm/helexa-neuron-prerelease\.spec$|^data/neuron|^neuron\.example\.toml$|^\.gitea/workflows/build-prerelease\.yml$'
               CORTEX_RE='^crates/cortex-gateway/|^crates/cortex-cli/|^crates/cortex-core/|^Cargo\.toml$|^Cargo\.lock$|^rpm/cortex-prerelease\.spec$|^data/cortex|^cortex\.example\.toml$|^models\.example\.toml$|^\.gitea/workflows/build-prerelease\.yml$'
               BENCH_RE='^crates/helexa-bench/|^crates/cortex-core/|^Cargo\.toml$|^Cargo\.lock$|^rpm/helexa-bench-prerelease\.spec$|^data/helexa-bench|^helexa-bench\.example\.toml$|^\.gitea/workflows/build-prerelease\.yml$'
+              UPSTREAM_RE='^crates/helexa-upstream/|^crates/cortex-core/|^Cargo\.toml$|^Cargo\.lock$|^rpm/helexa-upstream-prerelease\.spec$|^data/helexa-upstream|^helexa-upstream\.example\.toml$|^\.gitea/workflows/build-prerelease\.yml$'
               # Any Rust change (incl. crates not packaged here, e.g.
               # helexa-acp) still needs lint+test on main.
               RUST_RE='\.rs$|^crates/|Cargo\.toml$|^Cargo\.lock$'
@@ -156,10 +166,12 @@ jobs:
               CORTEX_BASE=$(base_for cortex)
               NEURON_BASE=$(base_for helexa-neuron-blackwell)
               BENCH_BASE=$(base_for helexa-bench)
+              UPSTREAM_BASE=$(base_for helexa-upstream)
               BUILD_CORTEX=$(decide "$CORTEX_BASE" "$CORTEX_RE")
               BUILD_NEURON=$(decide "$NEURON_BASE" "$NEURON_RE")
               BUILD_BENCH=$(decide "$BENCH_BASE" "$BENCH_RE")
-              if [ "$BUILD_CORTEX" = "true" ] || [ "$BUILD_NEURON" = "true" ] || [ "$BUILD_BENCH" = "true" ]; then
+              BUILD_UPSTREAM=$(decide "$UPSTREAM_BASE" "$UPSTREAM_RE")
+              if [ "$BUILD_CORTEX" = "true" ] || [ "$BUILD_NEURON" = "true" ] || [ "$BUILD_BENCH" = "true" ] || [ "$BUILD_UPSTREAM" = "true" ]; then
                 CHECK_RUST=true
               else
                 CHECK_RUST=$(decide "$CORTEX_BASE" "$RUST_RE")
@@ -170,8 +182,9 @@ jobs:
           echo "build_cortex=${BUILD_CORTEX}" >> "$GITHUB_OUTPUT"
           echo "build_neuron=${BUILD_NEURON}" >> "$GITHUB_OUTPUT"
           echo "build_bench=${BUILD_BENCH}" >> "$GITHUB_OUTPUT"
+          echo "build_upstream=${BUILD_UPSTREAM}" >> "$GITHUB_OUTPUT"
           echo "check_rust=${CHECK_RUST}" >> "$GITHUB_OUTPUT"
-          echo "### change detection: build_cortex=${BUILD_CORTEX} build_neuron=${BUILD_NEURON} build_bench=${BUILD_BENCH} check_rust=${CHECK_RUST}"
+          echo "### change detection: build_cortex=${BUILD_CORTEX} build_neuron=${BUILD_NEURON} build_bench=${BUILD_BENCH} build_upstream=${BUILD_UPSTREAM} check_rust=${CHECK_RUST}"
 
   # fmt + clippy + test moved here from ci.yml for main pushes so the
   # two workflows stop queueing against each other (ci.yml's checks
@@ -303,6 +316,45 @@ jobs:
           path: artifacts/helexa-bench
           retention-days: 1
 
+  build-upstream:
+    name: Build helexa-upstream binary
+    timeout-minutes: 25
+    needs: prepare
+    if: needs.prepare.outputs.build_upstream == 'true'
+    # Pure-Rust, non-CUDA binary — same runner as cortex/bench.
+    runs-on: rust
+    env:
+      RUSTC_WRAPPER: sccache
+      SCCACHE_BUCKET: sccache
+      SCCACHE_ENDPOINT: http://caveman.kosherinata.internal:9000
+      SCCACHE_REGION: auto
+      SCCACHE_S3_USE_SSL: "false"
+      AWS_ACCESS_KEY_ID: ${{ secrets.SCCACHE_S3_ACCESS_KEY }}
+      AWS_SECRET_ACCESS_KEY: ${{ secrets.SCCACHE_S3_SECRET_KEY }}
+      # helexa-upstream uses the sqlx runtime query API (no compile-time
+      # query macros), so it builds without a database or a .sqlx cache.
+      # Set OFFLINE defensively so a stray macro can never reach for a DB.
+      SQLX_OFFLINE: "true"
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - name: Build helexa-upstream (release, sccache escalation)
+        run: script/ci-cargo-escalate.sh cargo build --release -p helexa-upstream
+
+      - name: Stage binary
+        run: |
+          mkdir --parents artifacts
+          cp target/release/helexa-upstream artifacts/helexa-upstream
+          ./artifacts/helexa-upstream --version || true
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: upstream-fc43
+          path: artifacts/helexa-upstream
+          retention-days: 1
+
   build-neuron:
     name: Build neuron-${{ matrix.flavour }}
     timeout-minutes: 35
@@ -332,7 +384,10 @@ jobs:
             cuda_home: /usr/local/cuda-13.0
             build_jobs: 8
             nvcc_threads: 4
-            cargo_features: "cuda cudnn"
+            # flash-attn on blackwell first (#95): beast carries the
+            # agentic prefill pain; ada/ampere follow once the win is
+            # measured. NEURON_FLASH_ATTN=0 is the runtime rollback.
+            cargo_features: "cuda cudnn flash-attn"
     runs-on: ${{ matrix.runner }}
     env:
       SCCACHE_BUCKET: sccache
@@ -459,6 +514,44 @@ jobs:
           path: ~/rpmbuild/RPMS/x86_64/*.rpm
           retention-days: 7
 
+  package-upstream:
+    name: Package helexa-upstream RPM
+    timeout-minutes: 20
+    needs: [prepare, build-upstream]
+    runs-on: rpm
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.ref }}
+
+      - uses: actions/download-artifact@v3
+        with:
+          name: upstream-fc43
+          path: artifacts/
+
+      - name: Build RPM
+        run: |
+          set -eux
+          rm -f ~/.rpmmacros
+          rpmdev-setuptree
+          cp artifacts/helexa-upstream ~/rpmbuild/SOURCES/
+          cp data/helexa-upstream.service ~/rpmbuild/SOURCES/
+          cp data/helexa-upstream-sysusers.conf ~/rpmbuild/SOURCES/
+          cp data/helexa-upstream-firewalld.xml ~/rpmbuild/SOURCES/
+          cp helexa-upstream.example.toml ~/rpmbuild/SOURCES/
+          cp LICENSE ~/rpmbuild/SOURCES/
+          rpmbuild -bb rpm/helexa-upstream-prerelease.spec \
+            --define "upstream_version ${{ needs.prepare.outputs.version }}" \
+            --define "upstream_prerelease ${{ needs.prepare.outputs.release }}" \
+            --undefine dist \
+            --define "dist .fc43"
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: rpm-upstream-fc43
+          path: ~/rpmbuild/RPMS/x86_64/*.rpm
+          retention-days: 7
+
   package-neuron:
     name: Package helexa-neuron-${{ matrix.flavour }} RPM
     timeout-minutes: 20
@@ -508,7 +601,7 @@ jobs:
   publish:
     name: Publish to rpm.lair.cafe (unstable)
     timeout-minutes: 25
-    needs: [lint, test, package-cortex, package-neuron, package-bench]
+    needs: [lint, test, package-cortex, package-neuron, package-bench, package-upstream]
     # Runs when at least one package was built and nothing failed.
     # lint/test may be skipped (docs-only refs never get here because
     # no packages build), but a real failure in any blocks the
@@ -518,10 +611,11 @@ jobs:
         !cancelled()
         && (needs.lint.result == 'success' || needs.lint.result == 'skipped')
         && (needs.test.result == 'success' || needs.test.result == 'skipped')
-        && (needs.package-cortex.result == 'success' || needs.package-neuron.result == 'success' || needs.package-bench.result == 'success')
+        && (needs.package-cortex.result == 'success' || needs.package-neuron.result == 'success' || needs.package-bench.result == 'success' || needs.package-upstream.result == 'success')
         && needs.package-cortex.result != 'failure'
         && needs.package-neuron.result != 'failure'
         && needs.package-bench.result != 'failure'
+        && needs.package-upstream.result != 'failure'
       }}
     runs-on: rpm
     concurrency:

.gitea/workflows/ci.yml

@@ -82,7 +82,9 @@ jobs:
   # see commit history).
   cuda-check:
     name: CUDA type-check
-    timeout-minutes: 35
+    # flash-attn kernel compilation dominates the first uncached run;
+    # sccache + the cargo target cache absorb it afterwards.
+    timeout-minutes: 70
     runs-on: cuda-13.0
     # The workflow-level env sets `RUSTC_WRAPPER: sccache`
     # unconditionally, which hard-fails cargo if the CUDA image
@@ -115,7 +117,7 @@ jobs:
           export PATH="/usr/local/cuda-13.0/bin:${PATH}"
           export LD_LIBRARY_PATH="/usr/local/cuda-13.0/targets/x86_64-linux/lib:/usr/local/cuda-13.0/lib64:${LD_LIBRARY_PATH:-}"
           export LIBRARY_PATH="/usr/local/cuda-13.0/targets/x86_64-linux/lib:/usr/local/cuda-13.0/lib64:${LIBRARY_PATH:-}"
-          script/ci-cargo-escalate.sh cargo check -p neuron --features cuda --all-targets
+          script/ci-cargo-escalate.sh cargo check -p neuron --features cuda,flash-attn --all-targets
 
   srpm-cortex:
     name: Build cortex SRPM

.gitea/workflows/deploy.yml

@@ -274,19 +274,52 @@ jobs:
 
           echo "LLM probe against ${model}"
           probe_body=$(printf '{"model":"%s","messages":[{"role":"user","content":"Reply with exactly one word: pineapple"}],"max_tokens":512,"temperature":0}' "${model}")
-          resp=$(curl -fsS --max-time 180 -H "content-type: application/json" \
+          # The probe races real traffic: a deploy publishes a new build
-            -d "${probe_body}" http://localhost:13131/v1/chat/completions) || {
+          # SHA, which is exactly what triggers helexa-bench to re-sweep
-            echo "FAIL: probe request errored"
+          # every scenario against this neuron — long capability-probe
-            exit 1
+          # generations can hold the batch-1 model for minutes. Admission
-          }
+          # control answers concurrent requests with 429/503 +
-          if printf %s "${resp}" | grep -qi pineapple; then
+          # Retry-After (#53/#63); treat those as "busy, try again", not
-            echo "LLM probe passed"
+          # deploy failure. Overall budget ~6 min.
-          else
+          probe_deadline=$(( $(date +%s) + 360 ))
-            echo "FAIL: probe response missing expected token"
+          attempt=0
-            printf %s "${resp}" | head -c 2000
+          while :; do
-            echo
+            attempt=$(( attempt + 1 ))
-            exit 1
+            hdrs=$(mktemp)
-          fi
+            resp=$(curl -sS --max-time 180 -D "${hdrs}" \
+              -H "content-type: application/json" \
+              -d "${probe_body}" http://localhost:13131/v1/chat/completions) || resp=""
+            status=$(awk 'toupper($1) ~ /^HTTP/ {code=$2} END {print code}' "${hdrs}")
+            retry_after=$(awk 'tolower($1) == "retry-after:" {print $2+0; exit}' "${hdrs}")
+            rm -f "${hdrs}"
+            if [ "${status}" = "200" ]; then
+              if printf %s "${resp}" | grep -qi pineapple; then
+                echo "LLM probe passed (attempt ${attempt})"
+                break
+              fi
+              echo "FAIL: probe response missing expected token"
+              printf %s "${resp}" | head -c 2000
+              echo
+              exit 1
+            fi
+            case "${status}" in
+              429|503) ;;  # busy/queue-full — retryable
+              *)
+                echo "FAIL: probe request errored (HTTP ${status:-none})"
+                printf %s "${resp}" | head -c 2000
+                echo
+                exit 1
+                ;;
+            esac
+            if [ "$(date +%s)" -ge "${probe_deadline}" ]; then
+              echo "FAIL: model still busy (HTTP ${status}) after probe budget"
+              exit 1
+            fi
+            wait_s=${retry_after:-15}
+            [ "${wait_s}" -ge 5 ] 2>/dev/null || wait_s=15
+            echo "model busy (HTTP ${status}), retrying in ${wait_s}s (attempt ${attempt})"
+            sleep "${wait_s}"
+          done
           DEPLOY
 
       - name: Ensure firewalld allows helexa-neuron

.gitignore

@@ -1,6 +1,9 @@
 /target
 /bench/node_modules
 /bench/dist
+/helexa.ai/node_modules
+/helexa.ai/dist
+helexa.ai/.env.local
 *.swp
 *.swo
 .idea/

CLAUDE.md

@@ -185,6 +185,32 @@ Run these locally before pushing. `cargo fmt --all` fixes formatting
 automatically. Clippy warnings must be resolved, not suppressed with
 `#[allow(...)]` unless there is a clear rationale.
 
+## Development workflow
+
+Work each change on its own branch; `main` stays releasable.
+
+1. Implement on a feature branch (`fix/<issue>-…`, `feat/<issue>-…`).
+2. Run the CI triad locally (`cargo fmt --check --all`,
+   `cargo clippy --workspace -- -D warnings`, `cargo test --workspace`).
+   Local builds are **CPU-only** — the `#[cfg(feature = "cuda")]` neuron/TP
+   paths do NOT compile locally. The branch CI's **CUDA type-check** job is
+   the only thing that validates them, so for any neuron change the push to
+   Gitea is the real gate, not a rubber stamp.
+3. Push the branch on local-green (no need to ask first), and background-watch
+   its CI run via the gitea-mcp `actions_run_read` tools. Start the next piece
+   of work meanwhile.
+4. Merge to `main` when the four **validation** jobs are green — Format,
+   Clippy, Test, CUDA type-check. The SRPM / COPR / version-bump jobs are the
+   deploy pipeline (they run on `main`), not validation — don't wait on them.
+5. Merging/pushing to `main` triggers the auto-deploy pipeline.
+
+Docs-only changes (no `#[cfg(feature = "cuda")]` impact) can go straight to
+`main` — there's nothing for the CUDA type-check to prove.
+
+SSH note: the gitea remote host offers multiple agent keys and cuts the
+connection before reaching the right one. This repo pins the working key via
+`git config core.sshCommand "ssh -i ~/.ssh/id_grenade -o IdentitiesOnly=yes"`.
+
 ## Environment
 
 - Targets Fedora 43 (systemd, SELinux enforcing)

Cargo.toml

@@ -7,6 +7,9 @@ members = [
     "crates/neuron",
     "crates/helexa-acp",
     "crates/helexa-bench",
+    "crates/helexa-router",
+    "crates/helexa-stream",
+    "crates/helexa-upstream",
 ]
 
 [workspace.package]

asset/helexa-bench/bob.toml

@@ -18,6 +18,35 @@ db_path = "/var/lib/helexa-bench/bench.sqlite"
 [scenarios]
 prompt_sizes = [128, 4096]
 max_tokens = 256
+# Concurrency / agentic-load scenarios (#89), enabled so the 27B baseline
+# carries p95-under-concurrency data before the F3 A/B gate (#94) needs
+# it for comparison. Levels mirror the real a0/hermes/opencode fan-out.
+concurrency_levels = [2, 4, 8]
+concurrency_prompt_tokens = 512
+
+# Capability probes (#91) — the reasoning/planning axis the speed
+# scenarios miss; scored manually via `helexa-bench score` (O7). Enabled
+# for the same reason: the F3 gate compares 80B-A3B variants against the
+# 27B on planning quality, so the 27B needs scored artifacts first.
+[[scenarios.capability_probes]]
+name = "rust-plan"
+max_tokens = 4096
+prompt = """
+Write an implementation plan for adding rate limiting to an Axum service.
+Honor existing conventions, call out trade-offs, and sequence the work.
+"""
+
+[[scenarios.capability_probes]]
+name = "debug-reason"
+max_tokens = 4096
+prompt = """
+A Rust axum server streams SSE responses through a reverse proxy. Clients
+report that streams stall for exactly 60 seconds and then resume, but only
+when response chunks are small and infrequent. Curling the backend directly
+never stalls. List the most likely causes in order of probability, explain
+the mechanism behind each, and describe the smallest experiment that would
+confirm or eliminate each cause.
+"""
 
 # Read-only JSON API consumed by the bench UI (hosted separately) and for
 # programmatic access. Served alongside the sweep loop.

asset/neuron/beast.toml

@@ -16,6 +16,14 @@ name = "candle"
 
 [harness.candle]
 
+# Batched decode engine (#98): up to 8 concurrent text streams
+# multiplex through one lockstep (B,1) forward per decode step.
+# max_in_flight is both the admission bound and the engine's slot
+# count. NEURON_BATCHING=0 (systemd drop-in) is the kill switch;
+# removing this section reverts to batch-1 serialization.
+[harness.candle.admission]
+max_in_flight = 8
+
 [[default_models]]
 model_id = "Qwen/Qwen3.6-27B"
 harness = "candle"

cortex.example.toml

@@ -90,3 +90,20 @@ account_id = "operator"
 key_id = "infra"
 # No hard_cap → uncapped operator infra key (own fleet, own use). Still
 # metered for visibility.
+
+# -- Upstream (helexa mesh) entitlements client (#57) --------------------
+# When enabled, a bearer key NOT found in [[entitlements.keys]] above is
+# validated against the helexa-upstream authority (mesh accounts), and its
+# budget is reserved/settled there. Operator-local keys (incl. the infra
+# key) never leave this process. Fail-closed: if upstream is unreachable a
+# request is refused (503 + Retry-After), never served un-authorized.
+# Disabled by default — a standalone operator runs purely local.
+[upstream]
+enabled = false
+# url = "https://upstream.helexa.ai"
+# Shared client bearer this cortex presents (maps to an operator_id
+# upstream). Override via CORTEX_UPSTREAM__BEARER in prod.
+# bearer = "replace-with-operator-client-secret"
+# timeout_secs = 5
+# How often to flush served-usage counters to upstream for reconciliation (#58).
+# served_usage_report_interval_secs = 60

crates/cortex-core/src/config.rs

@@ -22,6 +22,43 @@ pub struct GatewayConfig {
     /// setups keep working until keys are configured.
     #[serde(default)]
     pub entitlements: EntitlementsConfig,
+    /// helexa-upstream client (#57). When enabled, keys not found in the
+    /// local `[entitlements]` config are validated against the mesh
+    /// authority, and budget is reserved/settled there. Disabled by default
+    /// — a single operator runs purely local.
+    #[serde(default)]
+    pub upstream: UpstreamClientConfig,
+}
+
+/// `[upstream]` — the helexa-upstream authority client (#57). Locally
+/// unrecognised bearer keys are resolved against `url`'s `/authz/v1` surface
+/// (mesh accounts); local keys (operator + infra) never leave the process.
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct UpstreamClientConfig {
+    /// Enable the upstream fallthrough. Off → purely local entitlements.
+    #[serde(default)]
+    pub enabled: bool,
+    /// Base URL of helexa-upstream (e.g. "https://upstream.helexa.ai").
+    #[serde(default)]
+    pub url: String,
+    /// Shared client bearer this cortex presents to `/authz/v1` (maps to an
+    /// operator_id upstream). Sent as `Authorization: Bearer <bearer>`.
+    #[serde(default)]
+    pub bearer: String,
+    /// Per-call timeout (seconds) to upstream.
+    #[serde(default = "default_upstream_timeout")]
+    pub timeout_secs: u64,
+    /// How often (seconds) to flush served-usage counters to upstream for
+    /// reconciliation (#58).
+    #[serde(default = "default_served_usage_interval")]
+    pub served_usage_report_interval_secs: u64,
+}
+
+fn default_upstream_timeout() -> u64 {
+    5
+}
+fn default_served_usage_interval() -> u64 {
+    60
 }
 
 /// `[entitlements]` — the local/static [`crate::entitlements::EntitlementProvider`]
@@ -129,6 +166,7 @@ impl Default for GatewayConfig {
             neurons: vec![],
             models_config: default_models_path(),
             entitlements: EntitlementsConfig::default(),
+            upstream: UpstreamClientConfig::default(),
         }
     }
 }

crates/cortex-core/src/discovery.rs

@@ -68,6 +68,57 @@ pub struct HealthResponse {
     pub devices: Vec<DeviceHealth>,
     #[serde(default)]
     pub activation: ActivationStatus,
+    /// Per-model admission load (#53): how many requests are running vs.
+    /// queued on each loaded model right now. Cortex's load-aware router
+    /// (#55) reads this to spread traffic across replicas and to propagate
+    /// honest backpressure. `#[serde(default)]` keeps older gateways/neurons
+    /// interoperable (absent → empty → treated as no load info).
+    #[serde(default)]
+    pub models: Vec<ModelLoad>,
+}
+
+/// Live admission load for one loaded model (#53).
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ModelLoad {
+    pub id: String,
+    /// Requests currently running (batch-1 → 0 or 1).
+    pub in_flight: usize,
+    /// Requests waiting in the bounded admission queue.
+    pub queue_depth: usize,
+}
+
+#[cfg(test)]
+mod health_load_tests {
+    use super::*;
+
+    #[test]
+    fn health_response_without_models_field_still_deserializes() {
+        // A pre-#53 neuron's /health payload omits `models`; the gateway
+        // must still parse it (serde default → empty).
+        let json = r#"{"uptime_secs":42,"devices":[]}"#;
+        let resp: HealthResponse = serde_json::from_str(json).expect("back-compat parse");
+        assert_eq!(resp.uptime_secs, 42);
+        assert!(resp.models.is_empty());
+    }
+
+    #[test]
+    fn health_response_round_trips_model_load() {
+        let resp = HealthResponse {
+            uptime_secs: 1,
+            devices: vec![],
+            activation: ActivationStatus::default(),
+            models: vec![ModelLoad {
+                id: "Qwen/Qwen3.6-27B".into(),
+                in_flight: 1,
+                queue_depth: 3,
+            }],
+        };
+        let s = serde_json::to_string(&resp).unwrap();
+        let back: HealthResponse = serde_json::from_str(&s).unwrap();
+        assert_eq!(back.models.len(), 1);
+        assert_eq!(back.models[0].in_flight, 1);
+        assert_eq!(back.models[0].queue_depth, 3);
+    }
 }
 
 /// High-level activation state of the neuron daemon. The HTTP listener

crates/cortex-core/src/entitlements.rs

@@ -23,6 +23,14 @@
 use async_trait::async_trait;
 use serde::{Deserialize, Serialize};
 
+/// Internal header carrying the resolved account id from cortex to neuron.
+/// neuron trusts these over the WireGuard link (#54); cortex **strips** any
+/// client-supplied copy before stamping the authoritative value, so a client
+/// can never assert a principal directly.
+pub const HEADER_ACCOUNT_ID: &str = "x-helexa-account-id";
+/// Internal header carrying the resolved key id from cortex to neuron.
+pub const HEADER_KEY_ID: &str = "x-helexa-key-id";
+
 /// Who a request is for. Resolved once at the edge from the bearer key and
 /// carried through the request context. `account_id` is the billable owner
 /// (spendable at any operator, by decision); `key_id` identifies the
@@ -73,12 +81,19 @@ pub struct BudgetSnapshot {
     pub reserved: u64,
 }
 
-/// Authentication failure — the bearer key could not be resolved. Maps to
+/// Authentication failure — the bearer key could not be resolved.
-/// `401 invalid_api_key` (#49/#63).
 #[derive(Debug, thiserror::Error)]
 pub enum AuthError {
+    /// The key is genuinely unknown → `401 invalid_api_key` (#49/#63).
     #[error("invalid or unknown API key")]
     InvalidKey,
+    /// The authority that could resolve the key is unreachable (e.g. the
+    /// helexa-upstream client failed, #57). Fail **closed** but distinctly:
+    /// a transient outage must surface as `503 service_unavailable` +
+    /// `Retry-After`, never `401` — a real key must not be rejected as
+    /// invalid during an upstream blip.
+    #[error("entitlement authority unavailable; retry in {retry_after_secs}s")]
+    Unavailable { retry_after_secs: u64 },
 }
 
 /// Why a reservation was refused. Carries enough for the caller to build the

crates/cortex-core/src/harness.rs

@@ -54,10 +54,26 @@ pub struct ModelLimit {
     pub output: usize,
 }
 
-/// Operator-set pricing in USD per 1M tokens.
+/// Operator-set pricing, **USD per 1,000,000 tokens, as JSON numbers**
+/// (`float`) — the models.dev/opencode `cost` convention, which is what
+/// helexa's primary client reads. NOT per-token, NOT decimal strings (that
+/// is OpenRouter's `pricing` shape, which helexa deliberately does not emit
+/// — see #68). A client must not rescale by 10⁶.
 ///
-/// Self-hosted deployments typically leave both at `0.0`.  Cache fields are
+/// `cost` is sourced from the operator's `models.toml` catalogue profile and
-/// optional — set when the backend supports a prefix-cache discount tier.
+/// surfaced verbatim on `/v1/models`. The *absent* vs *zero* distinction is
+/// intentional and load-bearing (#68):
+/// - **`cost` absent** (the whole object omitted) — the model is **not
+///   priced**: the operator has not declared a rate. Clients should treat
+///   spend as unknown, not free.
+/// - **`cost` present with `input`/`output` = `0.0`** — the model is
+///   **intentionally free** (self-hosted, no charge). opencode renders `$0`.
+///
+/// Cache fields are optional — set them only when the backend supports a
+/// prefix-cache discount tier (relevant once cache-token reporting, #64,
+/// lands). The advertised rate here must equal the rate metering (#51) and
+/// reconciliation (#58/#59) bill against; today both read this catalogue
+/// value.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ModelCost {
     /// USD per 1M input (prompt) tokens.
@@ -98,7 +114,8 @@ pub struct ModelInfo {
     /// `None` when neither the catalogue nor the loaded model can provide it.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub limit: Option<ModelLimit>,
-    /// Operator-set pricing in USD per 1M tokens (0.0 = free/self-hosted).
+    /// Operator-set pricing — see [`ModelCost`] for units and the
+    /// absent (not priced) vs `0.0` (intentionally free) distinction.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cost: Option<ModelCost>,
     /// `true` when the model's tokenizer contains recognised tool-call

crates/cortex-core/src/node.rs

@@ -1,4 +1,4 @@
-use crate::discovery::{ActivationStatus, DiscoveryResponse};
+use crate::discovery::{ActivationStatus, DiscoveryResponse, ModelLoad};
 use crate::harness::{ModelCost, ModelLimit};
 use chrono::{DateTime, Utc};
 use serde::{Deserialize, Serialize};
@@ -27,6 +27,17 @@ pub struct NodeState {
     /// to synthesize `Loading` locations so clients see a catalogued
     /// model that's mid-prewarm as "loading", not "missing".
     pub activation: Option<ActivationStatus>,
+    /// Last-seen per-model admission load from this neuron's `/health`
+    /// (#53), keyed by model id. The router (#55) reads it to pick the
+    /// least-busy replica when a model is loaded on more than one neuron.
+    /// Empty until the first /health poll reports load.
+    pub model_load: HashMap<String, ModelLoad>,
+    /// Consecutive failed `/models` polls. The poller marks a node
+    /// unhealthy only once this crosses a threshold, so a single transient
+    /// miss (e.g. a neuron momentarily slow to answer while busy) doesn't
+    /// yank the node — and all its models — out of routing. Reset to 0 on
+    /// any successful poll.
+    pub consecutive_poll_failures: u32,
 }
 
 /// A model registered on a node, with its runtime status.
@@ -125,7 +136,9 @@ pub struct CortexModelEntry {
     /// at load time. `None` when neither source provides it.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub limit: Option<ModelLimit>,
-    /// Operator-set pricing in USD per 1M tokens (0.0 = free/self-hosted).
+    /// Operator-set pricing from the catalogue profile — see
+    /// [`cortex_core::harness::ModelCost`] for units (USD per 1M tokens) and
+    /// the absent (not priced) vs `0.0` (intentionally free) distinction.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub cost: Option<ModelCost>,
     /// `true` when any neuron reports this model supports tool calls.
@@ -134,6 +147,39 @@ pub struct CortexModelEntry {
     /// `true` when any neuron reports this model supports reasoning tokens.
     #[serde(default)]
     pub reasoning: bool,
+    // ── Flat ecosystem context-window fields (issue #78) ──────
+    // Duplicates of `limit` under the flat, vLLM-convention key names
+    // (`max_model_len` et al.) that OpenAI-ecosystem clients (Hermes
+    // Agent, vLLM tooling) probe for — they cannot see `limit.context`.
+    // Additive: `limit` stays the opencode-oriented source of truth.
+    // Derived, never set directly — call [`sync_flat_limit`] after the
+    // final `limit` value is known. Omitted (not `0`) when the window
+    // is unknown; absent-vs-zero is load-bearing, as with `cost`.
+    //
+    // [`sync_flat_limit`]: CortexModelEntry::sync_flat_limit
+    /// Served max-seq-len in tokens — mirrors `limit.context`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_model_len: Option<usize>,
+    /// Usable input budget — mirrors `limit.input` when present.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_input_tokens: Option<usize>,
+    /// Maximum generation tokens — mirrors `limit.output`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max_output_tokens: Option<usize>,
+}
+
+impl CortexModelEntry {
+    /// Re-derive the flat ecosystem fields (#78) from `limit`.
+    ///
+    /// Must run after the final `limit` is known (post merge/tightening),
+    /// immediately before serialization. Fully overwrites: a `None` limit
+    /// clears the flat fields, so stale values can't survive a merge that
+    /// dropped the limit.
+    pub fn sync_flat_limit(&mut self) {
+        self.max_model_len = self.limit.as_ref().map(|l| l.context);
+        self.max_input_tokens = self.limit.as_ref().and_then(|l| l.input);
+        self.max_output_tokens = self.limit.as_ref().map(|l| l.output);
+    }
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]

crates/cortex-core/src/openai.rs

@@ -116,6 +116,23 @@ pub struct Usage {
     /// prompt caching lands (#11); `None` until then.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub prompt_tokens_details: Option<PromptTokensDetails>,
+    /// helexa extension (non-OpenAI): server-measured prefill/decode
+    /// timing, so the bench harness can compute true prefill vs decode
+    /// tok/s instead of inferring both from client-side SSE arrival
+    /// (#85). Additive and optional — standard OpenAI clients ignore
+    /// it; cortex forwards usage verbatim so it survives proxying.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub helexa_timing: Option<HelexaTiming>,
+}
+
+/// helexa extension carried on [`Usage::helexa_timing`]. Mirrors
+/// neuron's internal `FinishTiming`. All fields are server-measured;
+/// `prefill_tokens` is the prefill-rate denominator.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct HelexaTiming {
+    pub prefill_ms: u64,
+    pub decode_ms: u64,
+    pub prefill_tokens: u64,
 }
 
 /// Sub-counts of `Usage::completion_tokens`.

crates/cortex-core/src/responses.rs

@@ -66,14 +66,48 @@ pub struct ResponsesRequest {
     pub extra: Value,
 }
 
-/// `input` is either a single string or an array of typed items.
+/// `input` is either a single string or an array of items.
 /// `#[serde(untagged)]` so the wire shape `"input": "hi"` and
 /// `"input": [{...}]` both deserialize.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(untagged)]
 pub enum ResponsesInput {
     Text(String),
-    Items(Vec<ResponsesInputItem>),
+    Items(Vec<ResponsesInputElement>),
+}
+
+/// One element of an `input` array.
+///
+/// OpenAI's Responses API accepts three shapes here, and real clients
+/// use all of them — most notably agent-zero (via litellm), which
+/// sends the bare "easy message" form. We must tolerate every shape,
+/// because `input` is an `#[serde(untagged)]` array: a single element
+/// that matches no variant fails the *entire* request with a 422
+/// (`did not match any variant of untagged enum ResponsesInput`).
+///
+/// 1. [`Self::Typed`] — an item carrying an explicit `"type"`
+///    discriminant (`message`, `function_call`, `function_call_output`,
+///    `reasoning`).
+/// 2. [`Self::EasyMessage`] — a bare `{role, content}` with **no**
+///    `type` field. This is OpenAI's `EasyInputMessage` and what
+///    litellm emits for every turn. `content` is optional so an
+///    assistant turn carrying only tool calls (`content: null`) still
+///    parses.
+/// 3. [`Self::Other`] — anything else, captured as raw JSON and
+///    dropped during translation. This is the forward-compat escape
+///    hatch that mirrors [`ResponsesRequest::extra`] at the item
+///    level: an unmodeled item type can never again reject the whole
+///    request.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(untagged)]
+pub enum ResponsesInputElement {
+    Typed(ResponsesInputItem),
+    EasyMessage {
+        role: String,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        content: Option<ResponsesMessageContent>,
+    },
+    Other(Value),
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -91,8 +125,11 @@ pub enum ResponsesInputItem {
         name: String,
         arguments: String,
     },
-    /// User is feeding a tool result back into the model.
+    /// User is feeding a tool result back into the model. `output`
-    FunctionCallOutput { call_id: String, output: String },
+    /// is a `Value` because OpenAI allows it to be either a plain
+    /// string or an array of content parts; the translator renders
+    /// either form to text rather than losing the tool result.
+    FunctionCallOutput { call_id: String, output: Value },
     /// Reasoning items emitted by o-series models. Accepted but
     /// not forwarded to the model — neuron's candle path doesn't
     /// surface reasoning separately yet.
@@ -132,6 +169,11 @@ pub enum ResponsesContentPart {
         #[serde(default, skip_serializing_if = "Vec::is_empty")]
         annotations: Vec<Value>,
     },
+    /// Any content-part type we don't model (e.g. `refusal`, audio).
+    /// Captured as a unit so an unknown part can't reject the whole
+    /// request; dropped during translation.
+    #[serde(other)]
+    Unknown,
 }
 
 // ── Response (non-streaming) ─────────────────────────────────────────
@@ -277,20 +319,116 @@ mod tests {
             ResponsesInput::Items(items) => {
                 assert_eq!(items.len(), 1);
                 match &items[0] {
-                    ResponsesInputItem::Message { role, content } => {
+                    ResponsesInputElement::Typed(ResponsesInputItem::Message { role, content }) => {
                         assert_eq!(role, "user");
                         match content {
                             ResponsesMessageContent::Text(t) => assert_eq!(t, "hi"),
                             other => panic!("expected Text content, got {other:?}"),
                         }
                     }
-                    other => panic!("expected Message item, got {other:?}"),
+                    other => panic!("expected typed Message item, got {other:?}"),
                 }
             }
             other => panic!("expected Items, got {other:?}"),
         }
     }
 
+    #[test]
+    fn deserialises_bare_easy_message_without_type() {
+        // The shape agent-zero (via litellm) actually sends: `input`
+        // items are bare `{role, content}` with NO `type` field. This
+        // is the exact payload that was returning 422.
+        let raw = r#"{
+            "model": "Qwen/Qwen3.6-27B",
+            "store": true,
+            "tools": [{"type": "function", "name": "x", "description": "d", "parameters": {}}],
+            "input": [
+                {"role": "system", "content": "you are helpful"},
+                {"role": "assistant", "content": "{\"tool_name\":\"response\"}"},
+                {"role": "user", "content": "hi"}
+            ]
+        }"#;
+        let req: ResponsesRequest = serde_json::from_str(raw).unwrap();
+        let items = match req.input {
+            ResponsesInput::Items(i) => i,
+            other => panic!("expected Items, got {other:?}"),
+        };
+        assert_eq!(items.len(), 3);
+        for el in &items {
+            assert!(
+                matches!(el, ResponsesInputElement::EasyMessage { .. }),
+                "expected EasyMessage, got {el:?}"
+            );
+        }
+        // `tools` / `store` ride through `extra`, not `input`.
+        assert!(req.extra.get("tools").is_some());
+        assert_eq!(req.extra.get("store"), Some(&Value::Bool(true)));
+    }
+
+    #[test]
+    fn tolerates_null_content_and_unknown_item_types() {
+        // An assistant turn carrying only tool calls has `content: null`;
+        // and a future/unmodeled item type must not 422 the request.
+        let raw = r#"{
+            "model": "m",
+            "input": [
+                {"role": "assistant", "content": null},
+                {"type": "item_reference", "id": "abc"},
+                {"type": "function_call_output", "call_id": "c1",
+                 "output": [{"type": "output_text", "text": "result"}]},
+                {"role": "user", "content": "go"}
+            ]
+        }"#;
+        let req: ResponsesRequest = serde_json::from_str(raw).unwrap();
+        let items = match req.input {
+            ResponsesInput::Items(i) => i,
+            other => panic!("expected Items, got {other:?}"),
+        };
+        assert_eq!(items.len(), 4);
+        assert!(matches!(
+            &items[0],
+            ResponsesInputElement::EasyMessage { content: None, .. }
+        ));
+        assert!(matches!(&items[1], ResponsesInputElement::Other(_)));
+        assert!(matches!(
+            &items[2],
+            ResponsesInputElement::Typed(ResponsesInputItem::FunctionCallOutput { .. })
+        ));
+        assert!(matches!(
+            &items[3],
+            ResponsesInputElement::EasyMessage { .. }
+        ));
+    }
+
+    #[test]
+    fn tolerates_unknown_content_part_type() {
+        // A `refusal` (or any unmodeled) content part must parse, not 422.
+        let raw = r#"{
+            "model": "m",
+            "input": [
+                {"role": "assistant", "content": [
+                    {"type": "refusal", "refusal": "no"},
+                    {"type": "output_text", "text": "ok"}
+                ]}
+            ]
+        }"#;
+        let req: ResponsesRequest = serde_json::from_str(raw).unwrap();
+        let items = match req.input {
+            ResponsesInput::Items(i) => i,
+            other => panic!("expected Items, got {other:?}"),
+        };
+        let parts = match &items[0] {
+            ResponsesInputElement::EasyMessage {
+                content: Some(ResponsesMessageContent::Parts(p)),
+                ..
+            } => p,
+            other => panic!("expected EasyMessage with Parts, got {other:?}"),
+        };
+        assert_eq!(parts.len(), 2);
+        assert!(matches!(&parts[0], ResponsesContentPart::Unknown));
+        assert!(matches!(&parts[1], ResponsesContentPart::OutputText { .. }));
+    }
+
     #[test]
     fn deserialises_input_with_image() {
         let raw = r#"{
@@ -308,10 +446,10 @@ mod tests {
             other => panic!("expected Items, got {other:?}"),
         };
         let parts = match &items[0] {
-            ResponsesInputItem::Message {
+            ResponsesInputElement::Typed(ResponsesInputItem::Message {
                 content: ResponsesMessageContent::Parts(p),
                 ..
-            } => p,
+            }) => p,
             other => panic!("expected Parts, got {other:?}"),
         };
         assert_eq!(parts.len(), 2);

crates/cortex-core/src/translate.rs

@@ -400,6 +400,7 @@ pub fn openai_to_anthropic(resp: ChatCompletionResponse) -> MessagesResponse {
         total_tokens: 0,
         completion_tokens_details: None,
         prompt_tokens_details: None,
+        helexa_timing: None,
     });
 
     MessagesResponse {
@@ -772,6 +773,7 @@ mod stream_tests {
             total_tokens: 267,
             completion_tokens_details: None,
             prompt_tokens_details: None,
+            helexa_timing: None,
         });
         t.on_chunk(&usage_chunk);
         let fin = t.finish();

crates/cortex-gateway/Cargo.toml

@@ -6,6 +6,7 @@ license.workspace = true
 
 [dependencies]
 cortex-core.workspace = true
+helexa-stream = { path = "../helexa-stream" }
 async-trait.workspace = true
 tokio.workspace = true
 axum.workspace = true

crates/cortex-gateway/src/anthropic_sse.rs

@@ -32,6 +32,8 @@ pub async fn stream_translated(
     openai_body: axum::body::Bytes,
     model_id: &str,
     node_name: &str,
+    inbound_headers: &axum::http::HeaderMap,
+    usage_sink: Option<crate::metering::UsageSink>,
 ) -> Response {
     let url = format!("{endpoint}/v1/chat/completions");
     tracing::info!(
@@ -42,13 +44,14 @@ pub async fn stream_translated(
         "proxying streaming request (anthropic SSE translation)"
     );
 
-    let upstream = match client
+    let request = crate::auth::forward_principal_headers(
-        .post(&url)
+        client
-        .header("content-type", "application/json")
+            .post(&url)
-        .body(openai_body)
+            .header("content-type", "application/json")
-        .send()
+            .body(openai_body),
-        .await
+        inbound_headers,
-    {
+    );
+    let upstream = match request.send().await {
         Ok(r) => r,
         Err(e) => {
             tracing::warn!(
@@ -94,6 +97,10 @@ pub async fn stream_translated(
         let mut saw_tool_call = false;
         let mut last_finish: Option<String> = None;
         let mut frames = 0u64;
+        // Engine-truth usage for metering (#51), scanned from the upstream
+        // frames (neuron emits a final `usage` object on the stream, #48).
+        let mut usage_prompt = 0u64;
+        let mut usage_completion = 0u64;
 
         'outer: while let Some(block) = upstream.next().await {
             let block = match block {
@@ -121,6 +128,15 @@ pub async fn stream_translated(
                         continue;
                     }
                     tracing::trace!(node = %node, frame = %data, "anthropic stream: upstream frame");
+                    // Capture usage for metering before translation — the
+                    // usage object rides on a late frame (often after the
+                    // last content delta).
+                    if let Some(p) = crate::proxy::last_count_for(data, "prompt_tokens") {
+                        usage_prompt = p;
+                    }
+                    if let Some(c) = crate::proxy::last_count_for(data, "completion_tokens") {
+                        usage_completion = c;
+                    }
                     let Ok(chunk) = serde_json::from_str::<ChatCompletionChunk>(data) else {
                         tracing::debug!(node = %node, "anthropic stream: unparsable upstream frame skipped");
                         continue;
@@ -162,6 +178,14 @@ pub async fn stream_translated(
             terminated = done,
             "anthropic stream complete"
         );
+
+        // Settle metering with the observed usage (#51). Runs on every exit
+        // path of the pump — clean end, early break, or upstream error — so
+        // the reservation is always resolved. `(0, 0)` when no usage frame
+        // was seen, which releases without recording spend.
+        if let Some(sink) = usage_sink {
+            sink(usage_prompt, usage_completion);
+        }
     });
 
     Response::builder()

crates/cortex-gateway/src/auth.rs

@@ -0,0 +1,144 @@
+//! API-key authentication + principal resolution (#49).
+//!
+//! Identity rides standard bearer auth only — `Authorization: Bearer <key>`
+//! — which is what keeps every tier OpenAI-compatible by construction (no
+//! custom required headers or body fields, per #47). The middleware resolves
+//! the key to a [`Principal`] via the [`EntitlementProvider`], carries it in
+//! the request extensions for cortex-side metering/enforcement (#51/#52), and
+//! stamps it as internal headers on the request so it reaches neuron, which
+//! trusts cortex's assertion over WireGuard (#54).
+//!
+//! Anti-spoofing: any client-supplied principal header is **stripped** before
+//! the authoritative value is stamped, so a client can never assert a
+//! principal it didn't authenticate as.
+//!
+//! Rejection contract (#63): missing key under `require_auth`, or any present
+//! but unresolvable key, yields `401 invalid_api_key` in the #60 envelope.
+
+use crate::error::envelope_response;
+use crate::state::CortexState;
+use axum::extract::{Request, State};
+use axum::http::header::AUTHORIZATION;
+use axum::http::{HeaderMap, HeaderValue};
+use axum::middleware::Next;
+use axum::response::Response;
+use cortex_core::entitlements::{AuthError, HEADER_ACCOUNT_ID, HEADER_KEY_ID};
+use cortex_core::error_envelope::OpenAiError;
+use std::sync::Arc;
+
+/// Endpoints that never require auth: liveness/readiness probes. Everything
+/// else flows through resolution.
+fn is_public(path: &str) -> bool {
+    path == "/health" || path == "/"
+}
+
+/// Extract the bearer token from an `Authorization` header value, if present
+/// and well-formed. Scheme match is case-insensitive per RFC 7235.
+fn parse_bearer(headers: &HeaderMap) -> Option<String> {
+    let raw = headers.get(AUTHORIZATION)?.to_str().ok()?;
+    let (scheme, token) = raw.split_once(' ')?;
+    if scheme.eq_ignore_ascii_case("bearer") {
+        let token = token.trim();
+        (!token.is_empty()).then(|| token.to_string())
+    } else {
+        None
+    }
+}
+
+/// Axum middleware: resolve the bearer key, attach the principal, stamp the
+/// internal headers. Wired in `build_app` via `from_fn_with_state`.
+pub async fn require_principal(
+    State(fleet): State<Arc<CortexState>>,
+    mut req: Request,
+    next: Next,
+) -> Response {
+    if is_public(req.uri().path()) {
+        return next.run(req).await;
+    }
+
+    // Anti-spoof: drop any client-supplied principal headers up front.
+    {
+        let headers = req.headers_mut();
+        headers.remove(HEADER_ACCOUNT_ID);
+        headers.remove(HEADER_KEY_ID);
+    }
+
+    match parse_bearer(req.headers()) {
+        Some(key) => match fleet.entitlements.resolve(&key).await {
+            Ok(principal) => {
+                // Stamp the authoritative principal for neuron. Account/key
+                // ids come from operator config, so they're valid header
+                // values; guard anyway and skip a malformed one rather than
+                // panic.
+                if let (Ok(account), Ok(key_id)) = (
+                    HeaderValue::from_str(&principal.account_id),
+                    HeaderValue::from_str(&principal.key_id),
+                ) {
+                    let headers = req.headers_mut();
+                    headers.insert(HEADER_ACCOUNT_ID, account);
+                    headers.insert(HEADER_KEY_ID, key_id);
+                }
+                // Carry the typed principal for cortex-side metering (#51)
+                // and budget enforcement (#52).
+                req.extensions_mut().insert(principal);
+                next.run(req).await
+            }
+            // The entitlement authority is unreachable (upstream client
+            // blip, #57). Fail **closed but distinct**: a transient outage
+            // must not reject a real key as `401 invalid_api_key` — it's a
+            // retryable `503`. This holds regardless of require_auth: we
+            // can't safely serve a key we couldn't authorize.
+            Err(AuthError::Unavailable { retry_after_secs }) => {
+                envelope_response(OpenAiError::service_unavailable(
+                    "entitlement authority temporarily unavailable",
+                    Some(retry_after_secs),
+                ))
+            }
+            // A genuinely unrecognized key only hard-fails when auth is
+            // *required*. In allow-anonymous mode (the default) we IGNORE it
+            // and serve unauthenticated — otherwise the placeholder keys that
+            // OpenAI-compatible clients send by default (opencode, Open WebUI,
+            // Agent Zero, litellm) would all break though the operator never
+            // opted into auth. Pre-#49 the bearer was never inspected; this
+            // preserves that for require_auth=false.
+            Err(AuthError::InvalidKey) => {
+                if fleet.require_auth {
+                    unauthorized("invalid API key")
+                } else {
+                    tracing::debug!(
+                        "ignoring unrecognized bearer token (require_auth=false): serving anonymously"
+                    );
+                    next.run(req).await
+                }
+            }
+        },
+        None => {
+            if fleet.require_auth {
+                unauthorized("missing API key; supply 'Authorization: Bearer <key>'")
+            } else {
+                next.run(req).await
+            }
+        }
+    }
+}
+
+/// `401 invalid_api_key` in the standard envelope (#63).
+fn unauthorized(message: &str) -> Response {
+    envelope_response(OpenAiError::invalid_api_key(message))
+}
+
+/// Copy the cortex-stamped principal headers from an inbound [`HeaderMap`]
+/// onto an outbound reqwest builder. Used by the Anthropic proxy paths,
+/// which construct their own upstream requests instead of going through
+/// [`crate::proxy::forward_request`] (which forwards all headers verbatim).
+pub fn forward_principal_headers(
+    mut builder: reqwest::RequestBuilder,
+    headers: &HeaderMap,
+) -> reqwest::RequestBuilder {
+    for name in [HEADER_ACCOUNT_ID, HEADER_KEY_ID] {
+        if let Some(value) = headers.get(name) {
+            builder = builder.header(name, value);
+        }
+    }
+    builder
+}

crates/cortex-gateway/src/entitlements_chain.rs

@@ -0,0 +1,112 @@
+//! Chained entitlement provider (#57): operator-local keys first, mesh
+//! upstream for everything else.
+//!
+//! `resolve` tries the [`LocalEntitlementProvider`] (operator + infra keys —
+//! never a network hop); only a locally-unknown key falls through to
+//! [`UpstreamEntitlementProvider`]. Because the local provider treats an
+//! unconfigured principal as uncapped, reserve/settle/release/snapshot must
+//! **not** blindly hit local — they dispatch to whichever backend resolved
+//! that account, remembered in a map keyed by `account_id` (populated at
+//! resolve time).
+
+use crate::entitlements_local::LocalEntitlementProvider;
+use crate::entitlements_upstream::UpstreamEntitlementProvider;
+use async_trait::async_trait;
+use cortex_core::entitlements::{
+    AuthError, BudgetError, BudgetSnapshot, EntitlementProvider, Principal, Reservation,
+};
+use std::collections::HashMap;
+use tokio::sync::RwLock;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum Backend {
+    Local,
+    Upstream,
+}
+
+pub struct ChainedEntitlementProvider {
+    local: LocalEntitlementProvider,
+    upstream: UpstreamEntitlementProvider,
+    /// account_id → which backend owns it, learned at resolve time.
+    backends: RwLock<HashMap<String, Backend>>,
+}
+
+impl ChainedEntitlementProvider {
+    pub fn new(local: LocalEntitlementProvider, upstream: UpstreamEntitlementProvider) -> Self {
+        Self {
+            local,
+            upstream,
+            backends: RwLock::new(HashMap::new()),
+        }
+    }
+
+    async fn record(&self, account_id: &str, backend: Backend) {
+        self.backends
+            .write()
+            .await
+            .insert(account_id.to_string(), backend);
+    }
+
+    /// The backend that owns `account_id`. Defaults to `Upstream` for an
+    /// account never resolved this process-lifetime (a resolve always
+    /// precedes reserve in a request, so this is just a safe fallback —
+    /// upstream fails closed if the account is bogus).
+    async fn backend_for(&self, account_id: &str) -> Backend {
+        self.backends
+            .read()
+            .await
+            .get(account_id)
+            .copied()
+            .unwrap_or(Backend::Upstream)
+    }
+}
+
+#[async_trait]
+impl EntitlementProvider for ChainedEntitlementProvider {
+    async fn resolve(&self, api_key: &str) -> Result<Principal, AuthError> {
+        match self.local.resolve(api_key).await {
+            Ok(p) => {
+                self.record(&p.account_id, Backend::Local).await;
+                Ok(p)
+            }
+            Err(AuthError::InvalidKey) => {
+                let p = self.upstream.resolve(api_key).await?;
+                self.record(&p.account_id, Backend::Upstream).await;
+                Ok(p)
+            }
+            Err(e) => Err(e),
+        }
+    }
+
+    async fn reserve(
+        &self,
+        principal: &Principal,
+        max_tokens: u64,
+    ) -> Result<Reservation, BudgetError> {
+        match self.backend_for(&principal.account_id).await {
+            Backend::Local => self.local.reserve(principal, max_tokens).await,
+            Backend::Upstream => self.upstream.reserve(principal, max_tokens).await,
+        }
+    }
+
+    async fn settle(&self, reservation: Reservation, actual_tokens: u64) {
+        match self.backend_for(&reservation.principal.account_id).await {
+            Backend::Local => self.local.settle(reservation, actual_tokens).await,
+            Backend::Upstream => self.upstream.settle(reservation, actual_tokens).await,
+        }
+    }
+
+    async fn release(&self, reservation: Reservation) {
+        match self.backend_for(&reservation.principal.account_id).await {
+            Backend::Local => self.local.release(reservation).await,
+            Backend::Upstream => self.upstream.release(reservation).await,
+        }
+    }
+
+    async fn snapshot(&self, principal: &Principal) -> Option<BudgetSnapshot> {
+        match self.backend_for(&principal.account_id).await {
+            Backend::Local => self.local.snapshot(principal).await,
+            Backend::Upstream => self.upstream.snapshot(principal).await,
+        }
+    }
+}

crates/cortex-gateway/src/entitlements_upstream.rs

@@ -0,0 +1,246 @@
+//! helexa-upstream client (#57): an [`EntitlementProvider`] that resolves
+//! keys and reserves/settles budget against the mesh authority's
+//! `/authz/v1` surface (B2). It is "just another impl of the trait" — cortex
+//! enforcement (`auth.rs`, `metering.rs`) is unchanged.
+//!
+//! **Fail closed.** When upstream is unreachable, `resolve` returns
+//! [`AuthError::Unavailable`] (→ `503`, never `401`) and `reserve` refuses
+//! with a retryable [`BudgetError::RateLimited`] — a request is never served
+//! on an un-authorized key, and a real key is never rejected as invalid
+//! during a blip.
+
+use async_trait::async_trait;
+use cortex_core::config::UpstreamClientConfig;
+use cortex_core::entitlements::{
+    AuthError, BudgetError, BudgetSnapshot, EntitlementProvider, Principal, Reservation,
+};
+use serde::Deserialize;
+use std::time::Duration;
+
+/// Retry-After (seconds) advertised when we fail closed on an upstream
+/// outage.
+const FAIL_CLOSED_RETRY_SECS: u64 = 5;
+
+pub struct UpstreamEntitlementProvider {
+    client: reqwest::Client,
+    base_url: String,
+    bearer: String,
+}
+
+#[derive(Deserialize)]
+struct PrincipalDto {
+    account_id: String,
+    key_id: String,
+}
+#[derive(Deserialize)]
+struct SnapshotDto {
+    hard_cap: Option<u64>,
+    spent: u64,
+    reserved: u64,
+}
+#[derive(Deserialize)]
+struct ResolveResp {
+    principal: PrincipalDto,
+    #[allow(dead_code)]
+    snapshot: Option<SnapshotDto>,
+}
+#[derive(Deserialize)]
+struct ReserveResp {
+    reservation_id: Option<i64>,
+    rejected: Option<Rejection>,
+}
+#[derive(Deserialize)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+enum Rejection {
+    InsufficientQuota {
+        requested: u64,
+        available: u64,
+    },
+    RateLimited {
+        requested: u64,
+        available: u64,
+        retry_after_secs: u64,
+    },
+}
+
+impl UpstreamEntitlementProvider {
+    pub fn new(cfg: &UpstreamClientConfig) -> Self {
+        let client = reqwest::Client::builder()
+            .timeout(Duration::from_secs(cfg.timeout_secs))
+            .build()
+            .expect("failed to build upstream HTTP client");
+        Self {
+            client,
+            base_url: cfg.url.trim_end_matches('/').to_string(),
+            bearer: cfg.bearer.clone(),
+        }
+    }
+
+    fn url(&self, path: &str) -> String {
+        format!("{}{}", self.base_url, path)
+    }
+}
+
+#[async_trait]
+impl EntitlementProvider for UpstreamEntitlementProvider {
+    async fn resolve(&self, api_key: &str) -> Result<Principal, AuthError> {
+        let resp = self
+            .client
+            .post(self.url("/authz/v1/resolve"))
+            .bearer_auth(&self.bearer)
+            .json(&serde_json::json!({ "api_key": api_key }))
+            .send()
+            .await;
+        let resp = match resp {
+            Ok(r) => r,
+            Err(e) => {
+                tracing::warn!(error = %e, "upstream resolve unreachable; failing closed");
+                return Err(AuthError::Unavailable {
+                    retry_after_secs: FAIL_CLOSED_RETRY_SECS,
+                });
+            }
+        };
+        if resp.status().as_u16() == 401 {
+            return Err(AuthError::InvalidKey);
+        }
+        if !resp.status().is_success() {
+            return Err(AuthError::Unavailable {
+                retry_after_secs: FAIL_CLOSED_RETRY_SECS,
+            });
+        }
+        match resp.json::<ResolveResp>().await {
+            Ok(r) => Ok(Principal {
+                account_id: r.principal.account_id,
+                key_id: r.principal.key_id,
+            }),
+            Err(e) => {
+                tracing::warn!(error = %e, "upstream resolve: bad body; failing closed");
+                Err(AuthError::Unavailable {
+                    retry_after_secs: FAIL_CLOSED_RETRY_SECS,
+                })
+            }
+        }
+    }
+
+    async fn reserve(
+        &self,
+        principal: &Principal,
+        max_tokens: u64,
+    ) -> Result<Reservation, BudgetError> {
+        let fail_closed = || BudgetError::RateLimited {
+            requested: max_tokens,
+            available: 0,
+            retry_after_secs: FAIL_CLOSED_RETRY_SECS,
+        };
+        let resp = self
+            .client
+            .post(self.url("/authz/v1/reserve"))
+            .bearer_auth(&self.bearer)
+            .json(&serde_json::json!({
+                "account_id": principal.account_id,
+                "key_id": principal.key_id,
+                "max_tokens": max_tokens,
+            }))
+            .send()
+            .await;
+        let resp = match resp {
+            Ok(r) if r.status().is_success() => r,
+            Ok(r) => {
+                tracing::warn!(status = %r.status(), "upstream reserve non-2xx; failing closed");
+                return Err(fail_closed());
+            }
+            Err(e) => {
+                tracing::warn!(error = %e, "upstream reserve unreachable; failing closed");
+                return Err(fail_closed());
+            }
+        };
+        match resp.json::<ReserveResp>().await {
+            Ok(ReserveResp {
+                reservation_id: Some(id),
+                ..
+            }) => Ok(Reservation {
+                id: id as u64,
+                principal: principal.clone(),
+                reserved: max_tokens,
+            }),
+            Ok(ReserveResp {
+                rejected:
+                    Some(Rejection::InsufficientQuota {
+                        requested,
+                        available,
+                    }),
+                ..
+            }) => Err(BudgetError::InsufficientQuota {
+                requested,
+                available,
+            }),
+            Ok(ReserveResp {
+                rejected:
+                    Some(Rejection::RateLimited {
+                        requested,
+                        available,
+                        retry_after_secs,
+                    }),
+                ..
+            }) => Err(BudgetError::RateLimited {
+                requested,
+                available,
+                retry_after_secs,
+            }),
+            _ => Err(fail_closed()),
+        }
+    }
+
+    async fn settle(&self, reservation: Reservation, actual_tokens: u64) {
+        // Best-effort; a lost settle is reaped by the upstream sweeper (B2).
+        let _ = self
+            .client
+            .post(self.url("/authz/v1/settle"))
+            .bearer_auth(&self.bearer)
+            .json(&serde_json::json!({
+                "reservation_id": reservation.id as i64,
+                "actual_tokens": actual_tokens,
+            }))
+            .send()
+            .await
+            .inspect_err(
+                |e| tracing::warn!(error = %e, "upstream settle failed (sweeper will reap)"),
+            );
+    }
+
+    async fn release(&self, reservation: Reservation) {
+        let _ = self
+            .client
+            .post(self.url("/authz/v1/release"))
+            .bearer_auth(&self.bearer)
+            .json(&serde_json::json!({ "reservation_id": reservation.id as i64 }))
+            .send()
+            .await
+            .inspect_err(
+                |e| tracing::warn!(error = %e, "upstream release failed (sweeper will reap)"),
+            );
+    }
+
+    async fn snapshot(&self, principal: &Principal) -> Option<BudgetSnapshot> {
+        let resp = self
+            .client
+            .post(self.url("/authz/v1/snapshot"))
+            .bearer_auth(&self.bearer)
+            .json(&serde_json::json!({
+                "account_id": principal.account_id,
+                "key_id": principal.key_id,
+            }))
+            .send()
+            .await
+            .ok()?;
+        if !resp.status().is_success() {
+            return None;
+        }
+        let dto = resp.json::<SnapshotDto>().await.ok()?;
+        Some(BudgetSnapshot {
+            hard_cap: dto.hard_cap,
+            spent: dto.spent,
+            reserved: dto.reserved,
+        })
+    }
+}

crates/cortex-gateway/src/handlers.rs

@@ -190,7 +190,7 @@ async fn completions(
 /// `POST /v1/messages` — accept Anthropic format, translate, proxy, translate back.
 async fn anthropic_messages(
     State(fleet): State<Arc<CortexState>>,
-    _headers: HeaderMap,
+    headers: HeaderMap,
     body: Bytes,
 ) -> Response {
     // Parse as Anthropic request.
@@ -306,6 +306,33 @@ async fn anthropic_messages(
     }
     let start = Instant::now();
 
+    // Per-request metering + budget enforcement (#51/#52), same lifecycle as
+    // the OpenAI paths. Estimate from the translated OpenAI body (what neuron
+    // sees). Refuse over-cap before dispatch via the #63 envelope; otherwise
+    // build the sink consumed by whichever branch runs below.
+    let usage_sink = match crate::metering::principal_from_headers(&headers) {
+        Some(principal) => {
+            let advertised =
+                advertised_output_limit(&fleet, &route.node_name, &route.resolved_model_id).await;
+            let max_tokens = crate::metering::reservation_estimate(&openai_body, advertised);
+            match crate::metering::reserve_or_reject(
+                Arc::clone(&fleet.entitlements),
+                &principal,
+                max_tokens,
+            )
+            .await
+            {
+                Ok(guard) => Some(crate::metering::usage_sink(
+                    principal,
+                    guard,
+                    std::sync::Arc::clone(&fleet.served_usage),
+                )),
+                Err(env) => return crate::error::envelope_response(env),
+            }
+        }
+        None => None,
+    };
+
     if is_streaming {
         // Anthropic SSE translation (#24): upstream speaks OpenAI SSE;
         // re-frame it event-by-event into Anthropic's message_start /
@@ -316,6 +343,8 @@ async fn anthropic_messages(
             openai_body,
             &model_id,
             &route.node_name,
+            &headers,
+            usage_sink,
         )
         .await;
         metrics::histogram!("cortex_request_duration_seconds", &labels)
@@ -335,13 +364,16 @@ async fn anthropic_messages(
             cold_start = route.cold_start,
             "proxying request"
         );
-        let upstream_resp = fleet
+        let upstream_resp = crate::auth::forward_principal_headers(
-            .http_client
+            fleet
-            .post(&target_url)
+                .http_client
-            .body(openai_body)
+                .post(&target_url)
-            .header("content-type", "application/json")
+                .body(openai_body)
-            .send()
+                .header("content-type", "application/json"),
-            .await;
+            &headers,
+        )
+        .send()
+        .await;
 
         let upstream_resp = match upstream_resp {
             Ok(r) => r,
@@ -437,6 +469,15 @@ async fn anthropic_messages(
 
         metrics::histogram!("cortex_request_duration_seconds", &labels)
             .record(start.elapsed().as_secs_f64());
+        // Settle metering with the upstream usage (#51). Scanned from the
+        // raw body — same engine-truth source as the streaming path — so we
+        // don't depend on the typed usage struct's optionality.
+        if let Some(sink) = usage_sink {
+            let tail = String::from_utf8_lossy(&body_bytes);
+            let prompt = proxy::last_count_for(&tail, "prompt_tokens").unwrap_or(0);
+            let completion = proxy::last_count_for(&tail, "completion_tokens").unwrap_or(0);
+            sink(prompt, completion);
+        }
         // Did the model actually produce a structured tool call, or just
         // text? This is the single most useful signal for "is tool
         // calling working end-to-end" — a `false` here alongside a
@@ -540,6 +581,11 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                 // Runtime-detected — will be OR-ed in Pass 2 from neuron data.
                 tool_call: false,
                 reasoning: false,
+                // Flat #78 fields are derived from `limit` in the final
+                // sync pass, once merging is done.
+                max_model_len: None,
+                max_input_tokens: None,
+                max_output_tokens: None,
             },
         );
     }
@@ -597,6 +643,9 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                     cost: None,
                     tool_call: entry.tool_call,
                     reasoning: entry.reasoning,
+                    max_model_len: None,
+                    max_input_tokens: None,
+                    max_output_tokens: None,
                 });
         }
     }
@@ -653,6 +702,9 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                     cost: None,
                     tool_call: false,
                     reasoning: false,
+                    max_model_len: None,
+                    max_input_tokens: None,
+                    max_output_tokens: None,
                 });
         }
     }
@@ -687,11 +739,24 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                 cost: target_entry.cost.clone(),
                 tool_call: target_entry.tool_call,
                 reasoning: target_entry.reasoning,
+                max_model_len: None,
+                max_input_tokens: None,
+                max_output_tokens: None,
             },
         );
     }
 
-    let data: Vec<Value> = entries.values().map(|e| json!(e)).collect();
+    // Final pass: derive the flat ecosystem context-window fields (#78)
+    // from each entry's now-settled `limit`, so vLLM-convention clients
+    // (Hermes Agent et al.) can read the window without knowing helexa's
+    // `limit` schema.
+    let data: Vec<Value> = entries
+        .values_mut()
+        .map(|e| {
+            e.sync_flat_limit();
+            json!(e)
+        })
+        .collect();
     Json(json!({
         "object": "list",
         "data": data,
@@ -724,6 +789,19 @@ async fn proxy_with_metrics(
     body: Bytes,
     model_id: &str,
 ) -> Response {
+    // Fail-fast prompt pre-validation (#56): refuse a prompt that already
+    // exceeds the model's advertised context window *before* dispatching to
+    // neuron — the same `400 context_length_exceeded` neuron would emit on
+    // overflow, just earlier and without burning a cold-load/queue slot.
+    // cortex has no tokenizer, so the estimate under-counts and neuron stays
+    // the exact wall; we only catch gross overages (the A0 failure mode).
+    if let Some(context) = advertised_context(fleet, &route.node_name, model_id).await {
+        let est = estimate_prompt_tokens(&body);
+        if est > context {
+            return context_length_exceeded_response(context, est, &headers);
+        }
+    }
+
     let labels = [
         ("model", model_id.to_string()),
         ("node", route.node_name.clone()),
@@ -734,9 +812,46 @@ async fn proxy_with_metrics(
         metrics::counter!("cortex_cold_starts_total", &labels).increment(1);
     }
 
+    // Per-request metering + budget enforcement (#51/#52): reconstruct the
+    // principal from the middleware-stamped headers, reserve the request's
+    // upper-bound cost (prompt estimate + max output), and build the
+    // completion sink that settles actual spend when the response finishes.
+    // A reservation over the hard cap is refused *before* dispatch with the
+    // #63 envelope. Anonymous requests skip all of this. Must happen before
+    // `headers`/`body` are moved into the proxy.
+    let usage_sink = match crate::metering::principal_from_headers(&headers) {
+        Some(principal) => {
+            let advertised = advertised_output_limit(fleet, &route.node_name, model_id).await;
+            let max_tokens = crate::metering::reservation_estimate(&body, advertised);
+            match crate::metering::reserve_or_reject(
+                Arc::clone(&fleet.entitlements),
+                &principal,
+                max_tokens,
+            )
+            .await
+            {
+                Ok(guard) => Some(crate::metering::usage_sink(
+                    principal,
+                    guard,
+                    std::sync::Arc::clone(&fleet.served_usage),
+                )),
+                Err(env) => return crate::error::envelope_response(env),
+            }
+        }
+        None => None,
+    };
+
     let start = Instant::now();
-    let result =
+    let result = proxy::forward_request(
-        proxy::forward_request(&fleet.http_client, route, path, headers, body, model_id).await;
+        &fleet.http_client,
+        route,
+        path,
+        headers,
+        body,
+        model_id,
+        usage_sink,
+    )
+    .await;
     let duration = start.elapsed();
 
     match result {
@@ -755,6 +870,117 @@ async fn proxy_with_metrics(
     }
 }
 
+/// The model's advertised `limit.output` (#62) on a given node, used as the
+/// default output budget for budget reservations (#52) when the request
+/// omits `max_(completion_)tokens`. `None` when the node/model/limit is
+/// unknown — callers fall back to [`crate::metering::FALLBACK_MAX_OUTPUT`].
+async fn advertised_output_limit(
+    fleet: &CortexState,
+    node_name: &str,
+    model_id: &str,
+) -> Option<u64> {
+    let nodes = fleet.nodes.read().await;
+    nodes
+        .get(node_name)?
+        .models
+        .get(model_id)?
+        .limit
+        .as_ref()
+        .map(|l| l.output as u64)
+}
+
+/// The model's advertised hard context window (`limit.context`, #62/#67) on a
+/// node, used for fail-fast prompt pre-validation (#56). `None` when no limit
+/// is known — pre-validation is then skipped and neuron remains the wall.
+async fn advertised_context(fleet: &CortexState, node_name: &str, model_id: &str) -> Option<u64> {
+    let nodes = fleet.nodes.read().await;
+    nodes
+        .get(node_name)?
+        .models
+        .get(model_id)?
+        .limit
+        .as_ref()
+        .map(|l| l.context as u64)
+}
+
+/// Conservative prompt-token estimate (~4 chars/token over message text).
+/// cortex has no tokenizer; under-counting is the safe direction — we only
+/// pre-reject gross overages (#56), and neuron enforces the exact wall.
+fn estimate_prompt_tokens(body: &[u8]) -> u64 {
+    let Ok(v) = serde_json::from_slice::<Value>(body) else {
+        return (body.len() as u64 / 4).max(1);
+    };
+    let mut chars = 0usize;
+    if let Some(messages) = v.get("messages").and_then(Value::as_array) {
+        for m in messages {
+            match m.get("content") {
+                Some(Value::String(s)) => chars += s.len(),
+                Some(Value::Array(parts)) => {
+                    for p in parts {
+                        if let Some(t) = p.get("text").and_then(Value::as_str) {
+                            chars += t.len();
+                        }
+                    }
+                }
+                _ => {}
+            }
+            chars += 8; // rough per-message role/formatting overhead
+        }
+    } else if let Some(prompt) = v.get("prompt").and_then(Value::as_str) {
+        chars += prompt.len(); // legacy /v1/completions
+    } else {
+        return (body.len() as u64 / 4).max(1);
+    }
+    (chars as u64 / 4).max(1)
+}
+
+/// Client-specific, advisory guidance for an over-long prompt (#56),
+/// fingerprinted from `User-Agent`. Strictly advisory: it rides the
+/// `X-Helexa-Advice` header only, never the error envelope, and behaviour
+/// never depends on it. Unknown clients get nothing.
+fn client_advice(headers: &HeaderMap) -> Option<&'static str> {
+    let ua = headers
+        .get(axum::http::header::USER_AGENT)?
+        .to_str()
+        .ok()?
+        .to_ascii_lowercase();
+    if ua.contains("litellm") {
+        Some(
+            "litellm forwards the full context; lower the configured context window or enable client-side compaction",
+        )
+    } else if ua.contains("agent-zero") || ua.contains("agent zero") {
+        Some("reduce the conversation/context size or summarize earlier turns before resending")
+    } else if ua.contains("zed") {
+        Some("reduce the assistant context window in Zed's settings")
+    } else {
+        None
+    }
+}
+
+/// `400 context_length_exceeded` for an over-long prompt caught at the edge
+/// (#56), in the #60 envelope — the same shape neuron emits on overflow, so
+/// clients (opencode auto-compacts) handle it identically. Attaches the
+/// advisory `X-Helexa-Advice` header for fingerprinted clients.
+fn context_length_exceeded_response(
+    context: u64,
+    prompt_est: u64,
+    headers: &HeaderMap,
+) -> Response {
+    let env = OpenAiError::context_length_exceeded(format!(
+        "This model's maximum context length is {context} tokens. Your request is \
+         estimated at ~{prompt_est} tokens. Please reduce the length of the messages."
+    ))
+    .with_extra("max", json!(context))
+    .with_extra("estimated_prompt_tokens", json!(prompt_est));
+    let mut response = crate::error::envelope_response(env);
+    if let Some(advice) = client_advice(headers)
+        && let Ok(value) = axum::http::HeaderValue::from_str(advice)
+    {
+        response.headers_mut().insert("x-helexa-advice", value);
+    }
+    response
+}
+
 /// Update `last_accessed` timestamp for a model on a node (drives LRU eviction).
 async fn touch_model(fleet: &CortexState, node_name: &str, model_id: &str) {
     let mut nodes = fleet.nodes.write().await;

crates/cortex-gateway/src/lib.rs

@@ -1,25 +1,41 @@
 pub mod anthropic_sse;
+pub mod auth;
+pub mod entitlements_chain;
 pub mod entitlements_local;
+pub mod entitlements_upstream;
 pub mod error;
 pub mod evictor;
 pub mod handlers;
+pub mod metering;
 pub mod metrics;
 pub mod poller;
 pub mod proxy;
 pub mod router;
+pub mod served_usage;
 pub mod state;
 
 use anyhow::Result;
 use axum::Router;
+use axum::middleware::from_fn_with_state;
 use cortex_core::config::GatewayConfig;
 use std::sync::Arc;
 use tower_http::cors::CorsLayer;
 use tower_http::trace::TraceLayer;
 
 /// Build the Axum application router with all routes wired up.
+///
+/// Layer order (outermost first): trace → CORS → auth → handlers. CORS is
+/// outer to auth so preflight `OPTIONS` short-circuits before resolution;
+/// auth (`require_principal`) resolves the bearer key, attaches the
+/// principal, and stamps the internal principal headers before any handler
+/// runs.
 pub fn build_app(fleet: Arc<state::CortexState>) -> Router {
     Router::new()
         .merge(handlers::api_routes())
+        .layer(from_fn_with_state(
+            Arc::clone(&fleet),
+            auth::require_principal,
+        ))
         .layer(CorsLayer::permissive())
         .layer(TraceLayer::new_for_http())
         .with_state(fleet)
@@ -42,6 +58,28 @@ pub async fn run(config: GatewayConfig) -> Result<()> {
         evictor::eviction_loop(evictor_fleet).await;
     });
 
+    // Served-usage reporter (#58): when this operator is part of the mesh,
+    // periodically flush absolute per-principal served-token counters to
+    // upstream for reconciliation.
+    if config.upstream.enabled {
+        let su_fleet = Arc::clone(&fleet);
+        let url = config.upstream.url.clone();
+        let bearer = config.upstream.bearer.clone();
+        let interval =
+            std::time::Duration::from_secs(config.upstream.served_usage_report_interval_secs);
+        tokio::spawn(async move {
+            loop {
+                tokio::time::sleep(interval).await;
+                let rows = su_fleet.served_usage.snapshot();
+                if let Err(e) =
+                    served_usage::report(&su_fleet.http_client, &url, &bearer, &rows).await
+                {
+                    tracing::warn!(error = %e, "served-usage report failed (will retry)");
+                }
+            }
+        });
+    }
+
     let app = build_app(Arc::clone(&fleet));
 
     let listen_addr = config.gateway.listen.parse::<std::net::SocketAddr>()?;

crates/cortex-gateway/src/metering.rs

@@ -0,0 +1,231 @@
+//! Per-request token metering (#51).
+//!
+//! Captures the real `(prompt, completion)` usage of every request and feeds
+//! it to two places: the [`EntitlementProvider`] spend ledger (via
+//! reserve→settle) and per-principal Prometheus counters. The principal is
+//! reconstructed from the internal headers the auth middleware stamped (#49),
+//! so this works uniformly across every proxy path without threading the
+//! typed principal through each handler.
+//!
+//! The reserve→settle lifecycle is established here but, in this phase,
+//! reserves **zero** tokens — metering only, no enforcement. Budget
+//! enforcement (#52) flips the reserved amount to the real
+//! `prompt + max_output` and handles the [`BudgetError`] rejection; the
+//! settle/release plumbing is identical, so that change is localized.
+//!
+//! [`ReservationGuard`] makes leaks impossible: settling records actual
+//! spend and releases the unused remainder; dropping a guard that was never
+//! settled releases the whole reservation. So an early return, error path,
+//! or dropped stream can't strand a reservation.
+
+use axum::http::HeaderMap;
+use cortex_core::entitlements::{
+    BudgetError, EntitlementProvider, HEADER_ACCOUNT_ID, HEADER_KEY_ID, Principal,
+};
+use cortex_core::error_envelope::OpenAiError;
+use std::sync::Arc;
+
+/// Fallback output-token budget when neither the request nor the model's
+/// advertised limit gives one. Bounds the reservation so a capped key is
+/// still gated even on under-specified requests (#52).
+pub const FALLBACK_MAX_OUTPUT: u64 = 4096;
+
+/// Invoked exactly once at request completion with best-effort
+/// `(prompt_tokens, completion_tokens)`. When no usage could be observed
+/// (e.g. a pre-dispatch failure or a dropped stream) it is dropped unused —
+/// which releases the held reservation via [`ReservationGuard`]'s `Drop`.
+pub type UsageSink = Box<dyn FnOnce(u64, u64) + Send>;
+
+/// Reconstruct the principal from the cortex-stamped internal headers. The
+/// auth middleware strips any client copy and stamps the authoritative value,
+/// so these headers are trustworthy within cortex. `None` for anonymous
+/// (unauthenticated) requests.
+pub fn principal_from_headers(headers: &HeaderMap) -> Option<Principal> {
+    let account_id = headers.get(HEADER_ACCOUNT_ID)?.to_str().ok()?.to_string();
+    let key_id = headers.get(HEADER_KEY_ID)?.to_str().ok()?.to_string();
+    Some(Principal { account_id, key_id })
+}
+
+/// Emit per-principal spend counters (#51). Labelled by account/key only —
+/// both are operator-bounded, so cardinality is controlled.
+pub fn record_spend(principal: &Principal, prompt: u64, completion: u64) {
+    let labels = [
+        ("account", principal.account_id.clone()),
+        ("key", principal.key_id.clone()),
+    ];
+    metrics::counter!("cortex_spend_tokens_total", &labels).increment(prompt + completion);
+    metrics::counter!("cortex_spend_prompt_tokens_total", &labels).increment(prompt);
+    metrics::counter!("cortex_spend_completion_tokens_total", &labels).increment(completion);
+}
+
+/// Holds a budget reservation for the life of a request. [`settle`] records
+/// actual spend and releases the remainder; an un-settled guard releases the
+/// whole reservation when dropped. Anonymous requests carry an empty guard,
+/// where every operation is a no-op.
+///
+/// [`settle`]: ReservationGuard::settle
+pub struct ReservationGuard {
+    provider: Arc<dyn EntitlementProvider>,
+    reservation: Option<cortex_core::entitlements::Reservation>,
+}
+
+impl ReservationGuard {
+    /// An empty guard for an anonymous request — no reservation to resolve.
+    pub fn anonymous(provider: Arc<dyn EntitlementProvider>) -> Self {
+        Self {
+            provider,
+            reservation: None,
+        }
+    }
+
+    /// Wrap an already-acquired reservation.
+    fn held(
+        provider: Arc<dyn EntitlementProvider>,
+        reservation: cortex_core::entitlements::Reservation,
+    ) -> Self {
+        Self {
+            provider,
+            reservation: Some(reservation),
+        }
+    }
+
+    /// Settle with the tokens actually consumed, disarming the drop-release.
+    /// Spawns the (fast, in-process for the local provider) settle so the
+    /// caller — which may be a sync stream-completion callback — needn't
+    /// await.
+    pub fn settle(mut self, actual_tokens: u64) {
+        if let Some(reservation) = self.reservation.take() {
+            let provider = Arc::clone(&self.provider);
+            tokio::spawn(async move {
+                provider.settle(reservation, actual_tokens).await;
+            });
+        }
+    }
+}
+
+impl Drop for ReservationGuard {
+    fn drop(&mut self) {
+        if let Some(reservation) = self.reservation.take() {
+            let provider = Arc::clone(&self.provider);
+            tokio::spawn(async move {
+                provider.release(reservation).await;
+            });
+        }
+    }
+}
+
+/// Build the completion sink for an authenticated request: record spend and
+/// settle the reservation with the observed total. Dropping it unused (no
+/// usage observed) releases the reservation via the guard.
+pub fn usage_sink(
+    principal: Principal,
+    guard: ReservationGuard,
+    served_usage: std::sync::Arc<crate::served_usage::ServedUsage>,
+) -> UsageSink {
+    Box::new(move |prompt, completion| {
+        record_spend(&principal, prompt, completion);
+        // Per-principal served-usage tally for #58 reconciliation. Recorded
+        // for every metered (authenticated) request; the flush task reports
+        // it to upstream when the operator is part of the mesh.
+        served_usage.add(
+            &principal.account_id,
+            &principal.key_id,
+            prompt + completion,
+        );
+        guard.settle(prompt + completion);
+    })
+}
+
+/// Reserve the request's upper-bound token cost for the principal, refusing
+/// *before* dispatch if it would exceed the hard cap (#52). On success
+/// returns a guard the caller settles with actual usage; on refusal returns
+/// the #63 envelope (`rate_limit_exceeded` + `Retry-After` for a resetting
+/// window, `insufficient_quota` for a hard balance — never `402`).
+pub async fn reserve_or_reject(
+    provider: Arc<dyn EntitlementProvider>,
+    principal: &Principal,
+    max_tokens: u64,
+) -> Result<ReservationGuard, OpenAiError> {
+    match provider.reserve(principal, max_tokens).await {
+        Ok(reservation) => Ok(ReservationGuard::held(provider, reservation)),
+        Err(err) => Err(budget_error_to_envelope(err)),
+    }
+}
+
+/// Map a [`BudgetError`] to the #63 envelope. The provider chose the window
+/// semantics; this only translates them to HTTP.
+fn budget_error_to_envelope(err: BudgetError) -> OpenAiError {
+    match err {
+        BudgetError::RateLimited {
+            retry_after_secs, ..
+        } => OpenAiError::rate_limit_exceeded(err.to_string(), retry_after_secs),
+        BudgetError::InsufficientQuota { .. } => OpenAiError::insufficient_quota(err.to_string()),
+    }
+}
+
+/// Upper-bound tokens to reserve for a request (#52): an over-estimate of
+/// the prompt plus the maximum output. `advertised_output` is the model's
+/// `limit.output` (#62), used when the request omits `max_(completion_)tokens`.
+/// Over-reserving is safe — settle corrects spend to the actual usage.
+pub fn reservation_estimate(body: &[u8], advertised_output: Option<u64>) -> u64 {
+    let max_output = requested_max_output(body)
+        .or(advertised_output)
+        .unwrap_or(FALLBACK_MAX_OUTPUT);
+    estimate_prompt_tokens(body).saturating_add(max_output)
+}
+
+/// The client's requested output cap, from `max_completion_tokens` (or the
+/// legacy `max_tokens`). `None` when unspecified.
+fn requested_max_output(body: &[u8]) -> Option<u64> {
+    let v: serde_json::Value = serde_json::from_slice(body).ok()?;
+    v.get("max_completion_tokens")
+        .or_else(|| v.get("max_tokens"))
+        .and_then(serde_json::Value::as_u64)
+}
+
+/// Rough prompt-token estimate at ~4 chars/token over the whole body. cortex
+/// has no tokenizer; JSON overhead makes this a conservative over-estimate,
+/// and neuron remains the exact context wall (#56/#60). Settle reconciles to
+/// the real usage afterward.
+fn estimate_prompt_tokens(body: &[u8]) -> u64 {
+    (body.len() as u64 / 4).max(1)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn requested_max_output_prefers_max_completion_tokens() {
+        let body = br#"{"model":"m","max_completion_tokens":256,"max_tokens":99}"#;
+        assert_eq!(requested_max_output(body), Some(256));
+    }
+
+    #[test]
+    fn requested_max_output_falls_back_to_legacy_max_tokens() {
+        let body = br#"{"model":"m","max_tokens":128}"#;
+        assert_eq!(requested_max_output(body), Some(128));
+    }
+
+    #[test]
+    fn estimate_uses_requested_output_when_present() {
+        // Requested output dominates; prompt estimate is small for a tiny body.
+        let body = br#"{"model":"m","max_tokens":1000}"#;
+        let est = reservation_estimate(body, Some(8192));
+        assert!(est >= 1000 && est < 1100, "est was {est}");
+    }
+
+    #[test]
+    fn estimate_uses_advertised_output_when_request_omits_it() {
+        let body = br#"{"model":"m","messages":[]}"#;
+        let est = reservation_estimate(body, Some(8192));
+        assert!(est >= 8192, "est was {est}");
+    }
+
+    #[test]
+    fn estimate_falls_back_when_nothing_advertised() {
+        let body = br#"{"model":"m"}"#;
+        let est = reservation_estimate(body, None);
+        assert!(est >= FALLBACK_MAX_OUTPUT, "est was {est}");
+    }
+}

crates/cortex-gateway/src/metrics.rs

@@ -63,4 +63,16 @@ fn describe_metrics() {
         "cortex_cold_starts_total",
         "Total number of cold-start model loads"
     );
+    metrics::describe_counter!(
+        "cortex_spend_tokens_total",
+        "Total metered tokens (prompt + completion) per principal, labelled by account/key (#51)"
+    );
+    metrics::describe_counter!(
+        "cortex_spend_prompt_tokens_total",
+        "Metered prompt tokens per principal, labelled by account/key (#51)"
+    );
+    metrics::describe_counter!(
+        "cortex_spend_completion_tokens_total",
+        "Metered completion tokens per principal, labelled by account/key (#51)"
+    );
 }

crates/cortex-gateway/src/poller.rs

@@ -5,12 +5,29 @@ use crate::state::CortexState;
 use chrono::Utc;
 use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
 use cortex_core::harness::ModelInfo;
-use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_core::node::{ModelEntry, ModelStatus, NodeState};
 use std::sync::Arc;
 use std::time::Duration;
 
 const POLL_INTERVAL: Duration = Duration::from_secs(10);
 
+/// Consecutive failed `/models` polls before a node is marked unhealthy.
+/// Debounces transient misses (a busy neuron briefly slow to answer) so a
+/// single blip can't yank a node — and its models — out of routing. At the
+/// 10s poll interval this tolerates ~20s of flapping before evicting.
+const POLL_FAILURE_THRESHOLD: u32 = 3;
+
+/// Record a failed poll for `node`, marking it unhealthy only once failures
+/// reach [`POLL_FAILURE_THRESHOLD`]. Below the threshold the node keeps its
+/// last-known health, riding over transient misses. A successful poll resets
+/// the counter (see the success arm in `poll_once`).
+fn record_poll_failure(node: &mut NodeState) {
+    node.consecutive_poll_failures = node.consecutive_poll_failures.saturating_add(1);
+    if node.consecutive_poll_failures >= POLL_FAILURE_THRESHOLD {
+        node.healthy = false;
+    }
+}
+
 /// Runs forever, polling all neurons on a fixed interval.
 pub async fn poll_loop(fleet: Arc<CortexState>) {
     loop {
@@ -138,13 +155,14 @@ async fn poll_neuron(fleet: &CortexState, name: &str, endpoint: &str) {
                     // Remove models no longer reported by the neuron.
                     node.models.retain(|id, _| seen.contains(id));
 
+                    node.consecutive_poll_failures = 0;
                     node.healthy = true;
                     node.last_poll = Some(Utc::now());
                     tracing::debug!(node = name, models = models.len(), "poll ok");
                 }
                 Err(e) => {
                     tracing::warn!(node = name, error = %e, "failed to parse /models response");
-                    node.healthy = false;
+                    record_poll_failure(node);
                 }
             }
         }
@@ -154,11 +172,11 @@ async fn poll_neuron(fleet: &CortexState, name: &str, endpoint: &str) {
                 status = %resp.status(),
                 "neuron returned non-success status"
             );
-            node.healthy = false;
+            record_poll_failure(node);
         }
         Err(e) => {
             tracing::warn!(node = name, error = %e, "failed to reach neuron");
-            node.healthy = false;
+            record_poll_failure(node);
         }
     }
 
@@ -200,6 +218,9 @@ async fn poll_health(fleet: &CortexState, name: &str, endpoint: &str) {
             let mut nodes = fleet.nodes.write().await;
             if let Some(node) = nodes.get_mut(name) {
                 node.activation = Some(h.activation);
+                // Per-model admission load (#53) → keyed by id for the
+                // load-aware router (#55).
+                node.model_load = h.models.into_iter().map(|m| (m.id.clone(), m)).collect();
             }
         }
         Err(e) => {

crates/cortex-gateway/src/proxy.rs

@@ -1,21 +1,27 @@
 //! Streaming HTTP reverse proxy to neuron backends.
 //!
-//! For streaming requests, SSE chunks are forwarded as they arrive.
+//! The streaming *mechanism* — forward an SSE body chunk-for-chunk without
-//! The proxy captures timing information for metrics but does not
+//! buffering, observing the bytes for metrics — lives in the shared
-//! buffer the full response.
+//! [`helexa_stream`] crate (#71), so cortex and helexa-router use one
+//! implementation. This module supplies cortex's *policy*: the
+//! [`CortexMetrics`] observer (per-request token metrics + per-principal
+//! reservation settle), cortex's logging contract, and the cortex error
+//! envelope. The usage-extraction helper is re-exported from the shared
+//! crate so existing call sites keep working.
 
 use crate::router::RouteDecision;
-use anyhow::Result;
+use axum::http::HeaderMap;
-use axum::body::Body;
+use axum::http::StatusCode;
-use axum::http::{HeaderMap, StatusCode};
 use axum::response::{IntoResponse, Response};
-use futures::Stream;
+use helexa_stream::{BodyTail, ChunkObserver, StreamError};
-use futures::stream::BoxStream;
 use reqwest::Client;
-use std::pin::Pin;
-use std::task::{Context, Poll};
 use std::time::Instant;
 
+/// Re-export the shared usage-extraction helper. Several cortex modules
+/// (`handlers`, `anthropic_sse`) pull token counts out of a buffered body
+/// tail via this function; it lives in `helexa-stream` now.
+pub use helexa_stream::last_count_for;
+
 /// Proxy a request body to the resolved backend node and stream the response.
 ///
 /// Logging contract: every call emits exactly one structured event at
@@ -31,6 +37,7 @@ pub async fn forward_request(
     headers: HeaderMap,
     body: bytes::Bytes,
     model_id: &str,
+    usage_sink: Option<crate::metering::UsageSink>,
 ) -> Result<Response, ProxyError> {
     let request_start = Instant::now();
     let url = format!("{}{}", route.endpoint, path);
@@ -41,66 +48,41 @@ pub async fn forward_request(
         "proxying request"
     );
 
-    let mut req_builder = client.post(&url).body(body);
+    let observer = CortexMetrics::new(model_id, &route.node_name, request_start, usage_sink);
 
-    // Forward relevant headers.
+    let response = helexa_stream::forward_streaming(client, &url, headers, body, observer)
-    for (key, value) in headers.iter() {
+        .await
-        if key == "host" || key == "content-length" {
+        .map_err(|e| {
-            continue; // reqwest sets these
+            match &e {
-        }
+                StreamError::Upstream(err) => tracing::warn!(
-        req_builder = req_builder.header(key, value);
+                    node = %route.node_name,
-    }
+                    url = %url,
+                    error = %err,
+                    "proxy: upstream request failed (network)"
+                ),
+                StreamError::ResponseBuild(err) => tracing::warn!(
+                    node = %route.node_name,
+                    url = %url,
+                    error = %err,
+                    "proxy: failed to build response"
+                ),
+            }
+            ProxyError::from(e)
+        })?;
 
-    let upstream_resp = match req_builder.send().await {
+    if !response.status().is_success() {
-        Ok(r) => r,
-        Err(e) => {
-            tracing::warn!(
-                node = %route.node_name,
-                url = %url,
-                error = %e,
-                "proxy: upstream request failed (network)"
-            );
-            return Err(ProxyError::Upstream(e));
-        }
-    };
-
-    let upstream_status = upstream_resp.status();
-    if !upstream_status.is_success() {
         // Streaming body — can't snippet without breaking the stream
         // pass-through. Log status + URL; the client still gets the
         // upstream status, just without the leaked body.
         tracing::warn!(
             node = %route.node_name,
             url = %url,
-            status = upstream_status.as_u16(),
+            status = response.status().as_u16(),
             "proxy: upstream returned non-2xx"
         );
     }
 
-    let status = StatusCode::from_u16(upstream_status.as_u16()).unwrap_or(StatusCode::BAD_GATEWAY);
+    Ok(response)
-
-    let resp_headers = upstream_resp.headers().clone();
-    let stream = TokenMetricsStream::new(
-        Box::pin(upstream_resp.bytes_stream()),
-        TokenMetrics::new(model_id, &route.node_name, request_start),
-    );
-
-    let body = Body::from_stream(stream);
-
-    let mut response = Response::builder().status(status);
-    for (key, value) in resp_headers.iter() {
-        response = response.header(key, value);
-    }
-
-    response.body(body).map_err(|e| {
-        tracing::warn!(
-            node = %route.node_name,
-            url = %url,
-            error = %e,
-            "proxy: failed to build response"
-        );
-        ProxyError::ResponseBuild(e.to_string())
-    })
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -111,6 +93,15 @@ pub enum ProxyError {
     ResponseBuild(String),
 }
 
+impl From<StreamError> for ProxyError {
+    fn from(e: StreamError) -> Self {
+        match e {
+            StreamError::Upstream(err) => ProxyError::Upstream(err),
+            StreamError::ResponseBuild(msg) => ProxyError::ResponseBuild(msg),
+        }
+    }
+}
+
 impl IntoResponse for ProxyError {
     fn into_response(self) -> Response {
         let (status, code, message) = match &self {
@@ -138,9 +129,10 @@ impl IntoResponse for ProxyError {
 //
 // The proxy never buffers or re-serialises the upstream body — chunks
 // are forwarded verbatim. For metrics it observes each chunk's arrival
-// time and keeps a bounded tail of the body text, from which the final
+// time and keeps a bounded tail of the body text (via the shared
-// OpenAI `usage` object (present on the last SSE chunk and on
+// `helexa_stream::BodyTail`), from which the final OpenAI `usage` object
-// non-streaming JSON bodies alike) yields engine-truth token counts.
+// (present on the last SSE chunk and on non-streaming JSON bodies alike)
+// yields engine-truth token counts.
 //
 // Emitted per request, labelled {model, node}:
 //   cortex_time_to_first_token_seconds  (histogram) — first body chunk
@@ -154,42 +146,29 @@ impl IntoResponse for ProxyError {
 /// non-streaming bodies.
 const TAIL_CAP_BYTES: usize = 64 * 1024;
 
-/// Find the value of the LAST `"key": <integer>` occurrence in `tail`.
+/// cortex's [`ChunkObserver`]: per-request token metrics plus the
-/// Pure and chunk-boundary-safe (the tail is contiguous appended text).
+/// per-principal reservation settle. Drives cortex policy over the shared
-/// The quoted-needle form means `completion_tokens` never matches
+/// streaming mechanism.
-/// `completion_tokens_details`.
+struct CortexMetrics {
-pub(crate) fn last_count_for(tail: &str, key: &str) -> Option<u64> {
-    let needle = format!("\"{key}\"");
-    let mut result = None;
-    for (idx, _) in tail.match_indices(&needle) {
-        let rest = tail[idx + needle.len()..].trim_start();
-        let Some(rest) = rest.strip_prefix(':') else {
-            continue;
-        };
-        let rest = rest.trim_start();
-        let digits: &str = &rest[..rest
-            .char_indices()
-            .find(|(_, c)| !c.is_ascii_digit())
-            .map(|(i, _)| i)
-            .unwrap_or(rest.len())];
-        if let Ok(v) = digits.parse::<u64>() {
-            result = Some(v);
-        }
-    }
-    result
-}
-
-struct TokenMetrics {
     labels: [(&'static str, String); 2],
     request_start: Instant,
     first_chunk: Option<Instant>,
     last_chunk: Option<Instant>,
-    tail: String,
+    tail: BodyTail,
     finished: bool,
+    /// Per-principal metering hook (#51). Invoked exactly once in `finish`
+    /// with the observed `(prompt, completion)` so the reservation can be
+    /// settled and spend recorded. `None` for anonymous requests.
+    usage_sink: Option<crate::metering::UsageSink>,
 }
 
-impl TokenMetrics {
+impl CortexMetrics {
-    fn new(model_id: &str, node_name: &str, request_start: Instant) -> Self {
+    fn new(
+        model_id: &str,
+        node_name: &str,
+        request_start: Instant,
+        usage_sink: Option<crate::metering::UsageSink>,
+    ) -> Self {
         Self {
             labels: [
                 ("model", model_id.to_string()),
@@ -198,25 +177,19 @@ impl TokenMetrics {
             request_start,
             first_chunk: None,
             last_chunk: None,
-            tail: String::new(),
+            tail: BodyTail::new(TAIL_CAP_BYTES),
             finished: false,
+            usage_sink,
         }
     }
+}
 
+impl ChunkObserver for CortexMetrics {
     fn observe(&mut self, chunk: &[u8]) {
         let now = Instant::now();
         self.first_chunk.get_or_insert(now);
         self.last_chunk = Some(now);
-        self.tail.push_str(&String::from_utf8_lossy(chunk));
+        self.tail.push(chunk);
-        if self.tail.len() > TAIL_CAP_BYTES {
-            // Keep the newest half; the usage object is always at the
-            // very end of the body. Split at a char boundary.
-            let mut cut = self.tail.len() - TAIL_CAP_BYTES / 2;
-            while !self.tail.is_char_boundary(cut) {
-                cut += 1;
-            }
-            self.tail.drain(..cut);
-        }
     }
 
     /// Emit the metrics exactly once — called on clean stream end and
@@ -227,130 +200,45 @@ impl TokenMetrics {
             return;
         }
         self.finished = true;
-        let Some(first) = self.first_chunk else {
-            return; // no body ever arrived — nothing to record
-        };
-        let ttft = first.duration_since(self.request_start).as_secs_f64();
-        metrics::histogram!("cortex_time_to_first_token_seconds", &self.labels).record(ttft);
 
-        if let Some(prompt) = last_count_for(&self.tail, "prompt_tokens") {
+        let prompt = last_count_for(self.tail.as_str(), "prompt_tokens");
-            metrics::counter!("cortex_prompt_tokens_total", &self.labels).increment(prompt);
+        let completion = last_count_for(self.tail.as_str(), "completion_tokens");
-        }
-        let Some(completion) = last_count_for(&self.tail, "completion_tokens") else {
-            return;
-        };
-        if completion == 0 {
-            return;
-        }
-        metrics::counter!("cortex_completion_tokens_total", &self.labels).increment(completion);
 
-        let last = self.last_chunk.unwrap_or(first);
+        // Per-model metrics — only when body chunks actually arrived.
-        let decode_window = last.duration_since(first).as_secs_f64();
+        if let Some(first) = self.first_chunk {
-        // Streaming: rate over the decode window (first→last chunk).
+            let ttft = first.duration_since(self.request_start).as_secs_f64();
-        // Non-streaming bodies arrive as ~one chunk (window ≈ 0), where
+            metrics::histogram!("cortex_time_to_first_token_seconds", &self.labels).record(ttft);
-        // the only honest denominator is the full request duration.
-        let secs = if decode_window >= 0.1 {
-            decode_window
-        } else {
-            last.duration_since(self.request_start).as_secs_f64()
-        };
-        if secs > 0.0 {
-            metrics::histogram!("cortex_tokens_per_second", &self.labels)
-                .record(completion as f64 / secs);
-        }
-    }
-}
 
-/// Pass-through stream wrapper that feeds [`TokenMetrics`]. Emits on
+            if let Some(prompt) = prompt {
-/// clean end-of-stream; the Drop impl covers client disconnects.
+                metrics::counter!("cortex_prompt_tokens_total", &self.labels).increment(prompt);
-struct TokenMetricsStream {
-    inner: BoxStream<'static, Result<bytes::Bytes, reqwest::Error>>,
-    metrics: TokenMetrics,
-}
-
-impl TokenMetricsStream {
-    fn new(
-        inner: BoxStream<'static, Result<bytes::Bytes, reqwest::Error>>,
-        metrics: TokenMetrics,
-    ) -> Self {
-        Self { inner, metrics }
-    }
-}
-
-impl Stream for TokenMetricsStream {
-    type Item = Result<bytes::Bytes, reqwest::Error>;
-
-    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
-        let this = self.get_mut();
-        match this.inner.as_mut().poll_next(cx) {
-            Poll::Ready(Some(Ok(chunk))) => {
-                this.metrics.observe(&chunk);
-                Poll::Ready(Some(Ok(chunk)))
             }
-            Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
+            if let Some(completion) = completion.filter(|c| *c > 0) {
-            Poll::Ready(None) => {
+                metrics::counter!("cortex_completion_tokens_total", &self.labels)
-                this.metrics.finish();
+                    .increment(completion);
-                Poll::Ready(None)
+
+                let last = self.last_chunk.unwrap_or(first);
+                let decode_window = last.duration_since(first).as_secs_f64();
+                // Streaming: rate over the decode window (first→last chunk).
+                // Non-streaming bodies arrive as ~one chunk (window ≈ 0),
+                // where the only honest denominator is the full request
+                // duration.
+                let secs = if decode_window >= 0.1 {
+                    decode_window
+                } else {
+                    last.duration_since(self.request_start).as_secs_f64()
+                };
+                if secs > 0.0 {
+                    metrics::histogram!("cortex_tokens_per_second", &self.labels)
+                        .record(completion as f64 / secs);
+                }
             }
-            Poll::Pending => Poll::Pending,
+        }
+
+        // Per-principal metering + reservation settle (#51). Always runs so
+        // the reservation is resolved even when no usage/body was observed
+        // (sink with (0, 0) → settle 0 → release).
+        if let Some(sink) = self.usage_sink.take() {
+            sink(prompt.unwrap_or(0), completion.unwrap_or(0));
         }
     }
 }
-
-impl Drop for TokenMetricsStream {
-    fn drop(&mut self) {
-        self.metrics.finish();
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::last_count_for;
-
-    #[test]
-    fn extracts_counts_from_final_sse_usage_chunk() {
-        let tail = concat!(
-            "data: {\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n",
-            "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":225,",
-            "\"completion_tokens\":42,\"total_tokens\":267}}\n\n",
-            "data: [DONE]\n\n"
-        );
-        assert_eq!(last_count_for(tail, "prompt_tokens"), Some(225));
-        assert_eq!(last_count_for(tail, "completion_tokens"), Some(42));
-    }
-
-    #[test]
-    fn extracts_counts_from_non_streaming_body() {
-        let tail = "{\"choices\":[{\"message\":{\"content\":\"hi\"}}],\
-                    \"usage\":{\"prompt_tokens\": 12, \"completion_tokens\": 7}}";
-        assert_eq!(last_count_for(tail, "prompt_tokens"), Some(12));
-        assert_eq!(last_count_for(tail, "completion_tokens"), Some(7));
-    }
-
-    #[test]
-    fn ignores_details_variants_and_takes_last_occurrence() {
-        // completion_tokens_details must not shadow completion_tokens,
-        // and the LAST usage object wins (matters when content echoes
-        // a usage-shaped string earlier in the stream).
-        let tail = concat!(
-            "data: {\"usage\":{\"completion_tokens\":1}}\n\n",
-            "data: {\"usage\":{\"completion_tokens\":99,",
-            "\"completion_tokens_details\":{\"reasoning_tokens\":3}}}\n\n"
-        );
-        assert_eq!(last_count_for(tail, "completion_tokens"), Some(99));
-    }
-
-    #[test]
-    fn absent_keys_yield_none() {
-        assert_eq!(
-            last_count_for("data: [DONE]\n\n", "completion_tokens"),
-            None
-        );
-        assert_eq!(last_count_for("", "prompt_tokens"), None);
-        // key present but non-numeric value
-        assert_eq!(
-            last_count_for("\"completion_tokens\": null", "completion_tokens"),
-            None
-        );
-    }
-}

crates/cortex-gateway/src/router.rs

@@ -50,6 +50,10 @@ pub enum RouteError {
         "model '{model_id}' is in the catalogue but no healthy neuron's topology satisfies its constraints"
     )]
     NoFeasibleNeuron { model_id: String },
+    #[error(
+        "model '{model_id}' is feasible on a neuron that is currently unhealthy — retry shortly"
+    )]
+    FeasibleNodeUnhealthy { model_id: String },
     #[error("cold-load of '{model_id}' on '{node}' failed: {message}")]
     ColdLoadFailed {
         model_id: String,
@@ -68,7 +72,9 @@ impl RouteError {
     /// safe to retry the same request); everything else is 404.
     pub fn http_status(&self) -> u16 {
         match self {
-            RouteError::NoHealthyNodes | RouteError::ModelRecovering { .. } => 503,
+            RouteError::NoHealthyNodes
+            | RouteError::ModelRecovering { .. }
+            | RouteError::FeasibleNodeUnhealthy { .. } => 503,
             _ => 404,
         }
     }
@@ -81,7 +87,8 @@ impl RouteError {
             | RouteError::EndpointResolveFailed(_, _)
             | RouteError::NoFeasibleNeuron { .. }
             | RouteError::ColdLoadFailed { .. }
-            | RouteError::ModelRecovering { .. } => "api_error",
+            | RouteError::ModelRecovering { .. }
+            | RouteError::FeasibleNodeUnhealthy { .. } => "api_error",
         }
     }
 
@@ -94,6 +101,7 @@ impl RouteError {
             RouteError::NoFeasibleNeuron { .. } => "service_unavailable",
             RouteError::ColdLoadFailed { .. } => "service_unavailable",
             RouteError::ModelRecovering { .. } => "service_unavailable",
+            RouteError::FeasibleNodeUnhealthy { .. } => "service_unavailable",
         }
     }
 
@@ -105,6 +113,7 @@ impl RouteError {
     pub fn retry_after_secs(&self) -> Option<u64> {
         match self {
             RouteError::ModelRecovering { .. } => Some(2),
+            RouteError::FeasibleNodeUnhealthy { .. } => Some(3),
             RouteError::NoHealthyNodes => Some(5),
             _ => None,
         }
@@ -132,7 +141,9 @@ pub async fn resolve(
     // Snapshot loaded / unloaded / recovering state from the poller cache.
     let (loaded_route, unloaded_route, recovering_node, any_healthy) = {
         let nodes = fleet.nodes.read().await;
-        let mut loaded_route = None;
+        // All healthy nodes with the model loaded, each with its current
+        // admission load (#53) so we can pick the least-busy replica (#55).
+        let mut loaded_candidates: Vec<(String, String, usize)> = Vec::new();
         let mut unloaded_route = None;
         let mut recovering_node = None;
         let mut any_healthy = false;
@@ -144,8 +155,15 @@ pub async fn resolve(
             if let Some(entry) = node.models.get(model_id) {
                 match entry.status {
                     ModelStatus::Loaded | ModelStatus::Reloading => {
-                        loaded_route = Some((node.name.clone(), node.endpoint.clone(), false));
+                        // Least-busy score: in-flight + queued from the
-                        break;
+                        // neuron's last /health (#53). Unknown load (no poll
+                        // yet) scores 0 so the replica stays eligible.
+                        let score = node
+                            .model_load
+                            .get(model_id)
+                            .map(|l| l.in_flight + l.queue_depth)
+                            .unwrap_or(0);
+                        loaded_candidates.push((node.name.clone(), node.endpoint.clone(), score));
                     }
                     ModelStatus::Unloaded => {
                         if unloaded_route.is_none() {
@@ -175,6 +193,12 @@ pub async fn resolve(
                 }
             }
         }
+        // Pick the least-busy loaded replica; ties break by node name for
+        // deterministic routing. `false` = not a cold start.
+        let loaded_route = loaded_candidates
+            .into_iter()
+            .min_by(|a, b| a.2.cmp(&b.2).then_with(|| a.0.cmp(&b.0)))
+            .map(|(name, endpoint, _score)| (name, endpoint, false));
         (loaded_route, unloaded_route, recovering_node, any_healthy)
     };
 
@@ -237,11 +261,32 @@ async fn pick_feasible_neuron(
         b.2.cmp(&a.2) // pinned first (true > false)
             .then(a.0.cmp(&b.0))
     });
-    let pick = candidates.into_iter().next();
+    if let Some((n, e, _)) = candidates.into_iter().next() {
-    pick.map(|(n, e, _)| (n, e))
+        return Ok((n, e));
-        .ok_or_else(|| RouteError::NoFeasibleNeuron {
+    }
+
+    // No *healthy* feasible neuron. Distinguish a transient outage from a
+    // permanent misconfiguration: if some neuron is topologically feasible
+    // but currently unhealthy (e.g. it briefly missed polls while busy),
+    // this is retryable — return 503 + Retry-After so the client backs off
+    // and retries instead of treating a 404 as a hard failure. Only when no
+    // neuron could *ever* satisfy the topology is it a permanent 404.
+    let feasible_but_unhealthy = nodes.values().any(|node| {
+        !node.healthy
+            && node
+                .discovery
+                .as_ref()
+                .is_some_and(|disc| profile.is_feasible_on(&node.name, &disc.devices))
+    });
+    if feasible_but_unhealthy {
+        Err(RouteError::FeasibleNodeUnhealthy {
             model_id: profile.id.clone(),
         })
+    } else {
+        Err(RouteError::NoFeasibleNeuron {
+            model_id: profile.id.clone(),
+        })
+    }
 }
 
 /// Issue `POST {endpoint}/models/load` for this profile on this neuron,

crates/cortex-gateway/src/served_usage.rs

@@ -0,0 +1,105 @@
+//! Served-usage ledger (#58): cortex meters, per principal and per UTC day,
+//! the tokens it has served on behalf of mesh accounts, and periodically
+//! reports **absolute** cumulative counters to helexa-upstream for
+//! reconciliation (operators are compensated for served tokens).
+//!
+//! Counters are cumulative-since-process-start for the current period;
+//! upstream upserts them monotonically (GREATEST), so re-sending the same
+//! value is idempotent and a flush that races another is harmless. (A
+//! process restart resets the in-memory counter; the monotonic upsert keeps
+//! upstream from regressing — at most it under-counts the restarted window,
+//! acceptable for beta. One cortex per operator token is assumed.)
+
+use serde::Serialize;
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+#[derive(Debug, Clone, Serialize, PartialEq, Eq)]
+pub struct ServedRow {
+    pub account_id: String,
+    pub key_id: String,
+    pub period: String, // YYYY-MM-DD (UTC)
+    pub served_tokens: u64,
+}
+
+#[derive(Default)]
+pub struct ServedUsage {
+    inner: Mutex<HashMap<(String, String, String), u64>>,
+}
+
+impl ServedUsage {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    /// Add served tokens for a principal in today's (UTC) period.
+    pub fn add(&self, account_id: &str, key_id: &str, tokens: u64) {
+        if tokens == 0 {
+            return;
+        }
+        let period = chrono::Utc::now().format("%Y-%m-%d").to_string();
+        let mut m = self.inner.lock().expect("served-usage lock");
+        *m.entry((account_id.to_string(), key_id.to_string(), period))
+            .or_insert(0) += tokens;
+    }
+
+    /// Absolute cumulative counters, for a flush to upstream.
+    pub fn snapshot(&self) -> Vec<ServedRow> {
+        let m = self.inner.lock().expect("served-usage lock");
+        m.iter()
+            .map(|((account_id, key_id, period), &served_tokens)| ServedRow {
+                account_id: account_id.clone(),
+                key_id: key_id.clone(),
+                period: period.clone(),
+                served_tokens,
+            })
+            .collect()
+    }
+}
+
+/// POST the absolute counters to upstream's `/authz/v1/served-usage`.
+pub async fn report(
+    client: &reqwest::Client,
+    base_url: &str,
+    bearer: &str,
+    rows: &[ServedRow],
+) -> Result<(), reqwest::Error> {
+    if rows.is_empty() {
+        return Ok(());
+    }
+    let url = format!("{}/authz/v1/served-usage", base_url.trim_end_matches('/'));
+    client
+        .post(url)
+        .bearer_auth(bearer)
+        .json(&serde_json::json!({ "rows": rows }))
+        .send()
+        .await?
+        .error_for_status()?;
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn accumulates_per_principal_and_period() {
+        let su = ServedUsage::new();
+        su.add("acct", "key", 10);
+        su.add("acct", "key", 5);
+        su.add("acct", "other", 7);
+        su.add("acct", "key", 0); // no-op
+        let mut rows = su.snapshot();
+        rows.sort_by(|a, b| a.key_id.cmp(&b.key_id));
+        assert_eq!(rows.len(), 2);
+        let key_row = rows.iter().find(|r| r.key_id == "key").unwrap();
+        assert_eq!(key_row.served_tokens, 15);
+        assert_eq!(
+            rows.iter()
+                .find(|r| r.key_id == "other")
+                .unwrap()
+                .served_tokens,
+            7
+        );
+    }
+}

crates/cortex-gateway/src/state.rs

@@ -1,4 +1,6 @@
+use crate::entitlements_chain::ChainedEntitlementProvider;
 use crate::entitlements_local::LocalEntitlementProvider;
+use crate::entitlements_upstream::UpstreamEntitlementProvider;
 use cortex_core::catalogue::ModelCatalogue;
 use cortex_core::config::{EvictionSettings, GatewayConfig, NeuronEndpoint};
 use cortex_core::entitlements::EntitlementProvider;
@@ -20,6 +22,9 @@ pub struct CortexState {
     /// Whether to reject unauthenticated requests (#49). Read by the auth
     /// middleware once it lands.
     pub require_auth: bool,
+    /// Per-principal served-token tally (#58), reported to upstream for
+    /// operator reconciliation by the flush task when upstream is enabled.
+    pub served_usage: Arc<crate::served_usage::ServedUsage>,
 }
 
 impl CortexState {
@@ -37,14 +42,28 @@ impl CortexState {
                     last_poll: None,
                     discovery: None,
                     activation: None,
+                    model_load: HashMap::new(),
+                    consecutive_poll_failures: 0,
                 },
             );
         }
 
         let catalogue = ModelCatalogue::load(&config.models_config);
 
-        let entitlements: Arc<dyn EntitlementProvider> =
+        // Local provider always handles operator + infra keys. When the
-            Arc::new(LocalEntitlementProvider::from_config(&config.entitlements));
+        // upstream client is enabled (#57), wrap it in the chain so locally
+        // unknown keys fall through to the mesh authority; otherwise stay
+        // purely local.
+        let local = LocalEntitlementProvider::from_config(&config.entitlements);
+        let entitlements: Arc<dyn EntitlementProvider> = if config.upstream.enabled {
+            tracing::info!(url = %config.upstream.url, "upstream entitlement client enabled");
+            Arc::new(ChainedEntitlementProvider::new(
+                local,
+                UpstreamEntitlementProvider::new(&config.upstream),
+            ))
+        } else {
+            Arc::new(local)
+        };
 
         Self {
             nodes: RwLock::new(nodes),
@@ -57,6 +76,7 @@ impl CortexState {
                 .expect("failed to build HTTP client"),
             entitlements,
             require_auth: config.entitlements.require_auth,
+            served_usage: Arc::new(crate::served_usage::ServedUsage::new()),
         }
     }
 }

crates/cortex-gateway/tests/aliases.rs

@@ -57,6 +57,7 @@ async fn test_alias_resolves_in_chat_completions() {
         }],
         models_config: models_path.to_string_lossy().to_string(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -143,6 +144,7 @@ async fn test_aliases_surface_in_v1_models() {
         }],
         models_config: models_path.to_string_lossy().to_string(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -232,6 +234,7 @@ async fn test_alias_falls_through_for_unmapped_model() {
         }],
         models_config: models_path.to_string_lossy().to_string(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));

crates/cortex-gateway/tests/auth.rs

@@ -0,0 +1,273 @@
+//! Integration tests for API-key auth + principal resolution (#49).
+//!
+//! Verifies the #63 rejection contract (401 invalid_api_key via the #60
+//! envelope) and that an authenticated request reaches neuron carrying the
+//! internal principal headers — while a client-supplied principal header is
+//! stripped (anti-spoofing).
+
+use axum::Json;
+use axum::extract::Path;
+use axum::http::HeaderMap;
+use axum::routing::{get, post};
+use cortex_core::config::{
+    ApiKeyConfig, EntitlementsConfig, EvictionSettings, EvictionStrategy, GatewayConfig,
+    GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::entitlements::{CapWindow, HEADER_ACCOUNT_ID, HEADER_KEY_ID};
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::{Value, json};
+use std::sync::{Arc, Mutex};
+use tokio::net::TcpListener;
+
+/// What the mock neuron observed on the inbound `/v1/chat/completions`
+/// request: the principal headers cortex stamped (or didn't).
+#[derive(Default)]
+struct Seen {
+    account_id: Option<String>,
+    key_id: Option<String>,
+}
+
+/// Spawn a mock neuron that records the principal headers it receives and
+/// returns a trivial chat completion. Returns (base_url, observed).
+async fn spawn_capturing_neuron() -> (String, Arc<Mutex<Seen>>) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    let base_url = format!("http://{addr}");
+    let inference_url = base_url.clone();
+    let seen: Arc<Mutex<Seen>> = Arc::new(Mutex::new(Seen::default()));
+    let sink = Arc::clone(&seen);
+
+    let app = axum::Router::new()
+        .route(
+            "/models/{model_id}/endpoint",
+            get(move |Path(_): Path<String>| {
+                let url = inference_url.clone();
+                async move { Json(json!({ "url": url })) }
+            }),
+        )
+        .route(
+            "/v1/chat/completions",
+            post(move |headers: HeaderMap, Json(body): Json<Value>| {
+                let sink = Arc::clone(&sink);
+                async move {
+                    {
+                        let mut s = sink.lock().unwrap();
+                        s.account_id = headers
+                            .get(HEADER_ACCOUNT_ID)
+                            .and_then(|v| v.to_str().ok())
+                            .map(str::to_string);
+                        s.key_id = headers
+                            .get(HEADER_KEY_ID)
+                            .and_then(|v| v.to_str().ok())
+                            .map(str::to_string);
+                    }
+                    let model = body.get("model").and_then(Value::as_str).unwrap_or("m");
+                    Json(json!({
+                        "id": "chatcmpl-auth-001",
+                        "object": "chat.completion",
+                        "created": 1700000000_u64,
+                        "model": model,
+                        "choices": [{
+                            "index": 0,
+                            "message": {"role": "assistant", "content": "ok"},
+                            "finish_reason": "stop"
+                        }],
+                        "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}
+                    }))
+                }
+            }),
+        )
+        .with_state(());
+
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+
+    (base_url, seen)
+}
+
+/// Spawn a gateway with the given entitlements config, a single neuron, and
+/// `test-model` seeded as loaded (build_app spawns no poller).
+async fn spawn_gateway(neuron_url: &str, entitlements: EntitlementsConfig) -> String {
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron_url.to_string(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements,
+        upstream: Default::default(),
+    };
+
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let node = nodes.get_mut("mock-node").unwrap();
+        node.healthy = true;
+        node.models.insert(
+            "test-model".into(),
+            ModelEntry {
+                id: "test-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: Some(8000),
+                capabilities: Vec::new(),
+                tool_call: false,
+                reasoning: false,
+                limit: None,
+            },
+        );
+    }
+
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+fn one_key_config(require_auth: bool) -> EntitlementsConfig {
+    EntitlementsConfig {
+        require_auth,
+        keys: vec![ApiKeyConfig {
+            key: "sk-good".into(),
+            account_id: "acct-1".into(),
+            key_id: Some("key-1".into()),
+            hard_cap: None,
+            window: CapWindow::Balance,
+        }],
+    }
+}
+
+fn chat_body() -> Value {
+    json!({
+        "model": "test-model",
+        "messages": [{"role": "user", "content": "hi"}]
+    })
+}
+
+#[tokio::test]
+async fn missing_key_when_required_is_401_invalid_api_key() {
+    let (neuron, _seen) = spawn_capturing_neuron().await;
+    let gateway = spawn_gateway(&neuron, one_key_config(true)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .json(&chat_body())
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::UNAUTHORIZED);
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "invalid_api_key");
+    assert_eq!(body["error"]["type"], "invalid_request_error");
+}
+
+#[tokio::test]
+async fn unrecognized_key_is_ignored_when_auth_not_required() {
+    let (neuron, seen) = spawn_capturing_neuron().await;
+    // allow-anonymous mode: a placeholder/unknown bearer (as opencode,
+    // Open WebUI, Agent Zero, litellm all send by default) must NOT be
+    // rejected — it's ignored and the request is served anonymously.
+    let gateway = spawn_gateway(&neuron, one_key_config(false)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-dummy-placeholder")
+        .json(&chat_body())
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    let _ = resp.bytes().await.unwrap();
+    // Served, but anonymous — no principal stamped from the bogus key.
+    assert!(seen.lock().unwrap().account_id.is_none());
+}
+
+#[tokio::test]
+async fn invalid_key_is_401_when_auth_required() {
+    let (neuron, seen) = spawn_capturing_neuron().await;
+    // With auth required, a present-but-wrong credential is rejected.
+    let gateway = spawn_gateway(&neuron, one_key_config(true)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-wrong")
+        .json(&chat_body())
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::UNAUTHORIZED);
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "invalid_api_key");
+    // Rejected before dispatch — neuron never saw the request.
+    assert!(seen.lock().unwrap().account_id.is_none());
+}
+
+#[tokio::test]
+async fn valid_key_reaches_neuron_with_principal_headers() {
+    let (neuron, seen) = spawn_capturing_neuron().await;
+    let gateway = spawn_gateway(&neuron, one_key_config(true)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-good")
+        // A spoofed principal header must be stripped, not forwarded.
+        .header(HEADER_ACCOUNT_ID, "attacker")
+        .json(&chat_body())
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    let s = seen.lock().unwrap();
+    assert_eq!(s.account_id.as_deref(), Some("acct-1"));
+    assert_eq!(s.key_id.as_deref(), Some("key-1"));
+}
+
+#[tokio::test]
+async fn anonymous_allowed_when_auth_not_required() {
+    let (neuron, seen) = spawn_capturing_neuron().await;
+    let gateway = spawn_gateway(&neuron, EntitlementsConfig::default()).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .json(&chat_body())
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    // No principal resolved → no principal headers stamped.
+    let s = seen.lock().unwrap();
+    assert!(s.account_id.is_none());
+    assert!(s.key_id.is_none());
+}
+
+#[tokio::test]
+async fn health_is_public_even_when_auth_required() {
+    let (neuron, _seen) = spawn_capturing_neuron().await;
+    let gateway = spawn_gateway(&neuron, one_key_config(true)).await;
+
+    let resp = reqwest::Client::new()
+        .get(format!("{gateway}/health"))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+}

crates/cortex-gateway/tests/budget_enforcement.rs

@@ -0,0 +1,254 @@
+//! Integration tests for budget enforcement (#52) — the A0 seatbelt.
+//!
+//! A reservation over the key's hard cap is refused *before* neuron is hit,
+//! with the #63 code matching the cap-window semantics (rate_limit_exceeded
+//! + Retry-After for a resetting window, insufficient_quota for a hard
+//! balance). Spend never exceeds the cap. No 402, ever.
+
+use axum::Json;
+use axum::extract::Path;
+use axum::routing::{get, post};
+use cortex_core::config::{
+    ApiKeyConfig, EntitlementsConfig, EvictionSettings, EvictionStrategy, GatewayConfig,
+    GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::entitlements::{CapWindow, Principal};
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::{Value, json};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+use tokio::net::TcpListener;
+
+/// Mock neuron with a hit counter on the inference path, so a test can prove
+/// a request was (or wasn't) dispatched.
+async fn spawn_counting_neuron() -> (String, Arc<AtomicU64>) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    let base_url = format!("http://{addr}");
+    let inference_url = base_url.clone();
+    let hits = Arc::new(AtomicU64::new(0));
+    let sink = Arc::clone(&hits);
+
+    let app = axum::Router::new()
+        .route(
+            "/models/{model_id}/endpoint",
+            get(move |Path(_): Path<String>| {
+                let url = inference_url.clone();
+                async move { Json(json!({ "url": url })) }
+            }),
+        )
+        .route(
+            "/v1/chat/completions",
+            post(move |Json(body): Json<Value>| {
+                let sink = Arc::clone(&sink);
+                async move {
+                    sink.fetch_add(1, Ordering::SeqCst);
+                    let model = body.get("model").and_then(Value::as_str).unwrap_or("m");
+                    Json(json!({
+                        "id": "chatcmpl-budget",
+                        "object": "chat.completion",
+                        "created": 1700000000_u64,
+                        "model": model,
+                        "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}],
+                        "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}
+                    }))
+                }
+            }),
+        );
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    (base_url, hits)
+}
+
+async fn spawn_gateway(neuron_url: &str, key: ApiKeyConfig) -> (Arc<CortexState>, String) {
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron_url.to_string(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements: EntitlementsConfig {
+            require_auth: true,
+            keys: vec![key],
+        },
+        upstream: Default::default(),
+    };
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let node = nodes.get_mut("mock-node").unwrap();
+        node.healthy = true;
+        node.models.insert(
+            "test-model".into(),
+            ModelEntry {
+                id: "test-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: Some(8000),
+                capabilities: Vec::new(),
+                tool_call: false,
+                reasoning: false,
+                limit: None,
+            },
+        );
+    }
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    (fleet, format!("http://{addr}"))
+}
+
+fn key(window: CapWindow, hard_cap: u64) -> ApiKeyConfig {
+    ApiKeyConfig {
+        key: "sk-cap".into(),
+        account_id: "acct-cap".into(),
+        key_id: Some("key-cap".into()),
+        hard_cap: Some(hard_cap),
+        window,
+    }
+}
+
+fn chat(max_tokens: u64) -> Value {
+    json!({
+        "model": "test-model",
+        "max_tokens": max_tokens,
+        "messages": [{"role": "user", "content": "hi"}]
+    })
+}
+
+#[tokio::test]
+async fn balance_over_cap_is_429_insufficient_quota_before_dispatch() {
+    let (neuron, hits) = spawn_counting_neuron().await;
+    // Cap far below a single request's reservation (max_tokens 1000).
+    let (_fleet, gateway) = spawn_gateway(&neuron, key(CapWindow::Balance, 10)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-cap")
+        .json(&chat(1000))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::TOO_MANY_REQUESTS);
+    // Hard balance → no Retry-After.
+    assert!(resp.headers().get(reqwest::header::RETRY_AFTER).is_none());
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "insufficient_quota");
+    // Refused before dispatch — neuron never saw it.
+    assert_eq!(hits.load(Ordering::SeqCst), 0);
+}
+
+#[tokio::test]
+async fn rolling_over_cap_is_429_rate_limited_with_retry_after() {
+    let (neuron, hits) = spawn_counting_neuron().await;
+    let (_fleet, gateway) =
+        spawn_gateway(&neuron, key(CapWindow::Rolling { seconds: 3600 }, 10)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-cap")
+        .json(&chat(1000))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::TOO_MANY_REQUESTS);
+    let retry = resp
+        .headers()
+        .get(reqwest::header::RETRY_AFTER)
+        .expect("rolling-window rejection must carry Retry-After");
+    assert!(retry.to_str().unwrap().parse::<u64>().unwrap() >= 1);
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "rate_limit_exceeded");
+    assert_eq!(hits.load(Ordering::SeqCst), 0);
+}
+
+#[tokio::test]
+async fn within_cap_is_served() {
+    let (neuron, hits) = spawn_counting_neuron().await;
+    let (_fleet, gateway) = spawn_gateway(&neuron, key(CapWindow::Balance, 1_000_000)).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .bearer_auth("sk-cap")
+        .json(&chat(50))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    let _ = resp.bytes().await.unwrap();
+    assert_eq!(hits.load(Ordering::SeqCst), 1);
+}
+
+#[tokio::test]
+async fn a0_seatbelt_caps_a_runaway_fan_out() {
+    // An Agent-Zero-style key with a modest cap: a burst of requests drains
+    // it, then further requests are refused — the account stops draining and
+    // spend never exceeds the cap.
+    let (neuron, hits) = spawn_counting_neuron().await;
+    let (fleet, gateway) = spawn_gateway(&neuron, key(CapWindow::Balance, 100)).await;
+    let client = reqwest::Client::new();
+
+    let mut ok = 0;
+    let mut refused = 0;
+    for _ in 0..20 {
+        let resp = client
+            .post(format!("{gateway}/v1/chat/completions"))
+            .bearer_auth("sk-cap")
+            .json(&chat(20))
+            .send()
+            .await
+            .unwrap();
+        match resp.status() {
+            reqwest::StatusCode::OK => {
+                ok += 1;
+                let _ = resp.bytes().await.unwrap();
+            }
+            reqwest::StatusCode::TOO_MANY_REQUESTS => {
+                refused += 1;
+                let body: Value = resp.json().await.unwrap();
+                assert_eq!(body["error"]["code"], "insufficient_quota");
+            }
+            other => panic!("unexpected status {other}"),
+        }
+    }
+
+    assert!(ok >= 1, "some requests should be served");
+    assert!(refused >= 1, "the cap must eventually refuse the fan-out");
+    assert_eq!(
+        hits.load(Ordering::SeqCst),
+        ok,
+        "refused requests never dispatched"
+    );
+
+    // Spend never exceeded the hard cap (reservation prevents overshoot).
+    // Poll briefly for in-flight settles to land.
+    let principal = Principal {
+        account_id: "acct-cap".into(),
+        key_id: "key-cap".into(),
+    };
+    for _ in 0..50 {
+        let snap = fleet.entitlements.snapshot(&principal).await.unwrap();
+        if snap.reserved == 0 {
+            break;
+        }
+        tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+    }
+    let snap = fleet.entitlements.snapshot(&principal).await.unwrap();
+    assert!(snap.spent <= 100, "spent {} exceeded cap", snap.spent);
+}

crates/cortex-gateway/tests/common/mod.rs

@@ -430,6 +430,7 @@ pub async fn spawn_gateway_with_state(mock_url: &str) -> (Arc<CortexState>, Stri
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));

crates/cortex-gateway/tests/error_envelope.rs

@@ -89,6 +89,7 @@ async fn error_response_no_healthy_nodes() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(cortex_gateway::state::CortexState::from_config(&config));

crates/cortex-gateway/tests/eviction.rs

@@ -72,6 +72,7 @@ fn make_fleet(endpoint: &str, defrag_after: u32) -> Arc<CortexState> {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
     Arc::new(CortexState::from_config(&config))
 }

crates/cortex-gateway/tests/feasibility_routing.rs

@@ -0,0 +1,125 @@
+//! Router: a catalogued model whose only topologically-feasible neuron is
+//! currently unhealthy is a *transient* condition (retryable 503), not a
+//! permanent 404. This is the exact shape of the beast incident: benjy/
+//! quadbrat (1 GPU, healthy) can't host the 27B, and beast (2 GPU) — the
+//! sole feasible node — briefly drops out → clients must back off and retry,
+//! not hard-fail.
+
+use cortex_core::config::{
+    EvictionSettings, EvictionStrategy, GatewayConfig, GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::discovery::{DeviceInfo, DiscoveryResponse};
+use cortex_gateway::router::{self, RouteError};
+use cortex_gateway::state::CortexState;
+use std::sync::Arc;
+
+fn devices(n: usize) -> Vec<DeviceInfo> {
+    (0..n)
+        .map(|i| DeviceInfo {
+            index: i as u32,
+            name: "RTX 5090".into(),
+            vram_total_mb: 32_768,
+            compute_capability: "9.0".into(),
+        })
+        .collect()
+}
+
+fn discovery(host: &str, n_devices: usize) -> DiscoveryResponse {
+    DiscoveryResponse {
+        hostname: host.into(),
+        os: "Linux".into(),
+        kernel: "7.0".into(),
+        cuda_version: Some("13.0".into()),
+        driver_version: Some("999".into()),
+        devices: devices(n_devices),
+        harnesses: vec!["candle".into()],
+        cuda_unavailable_reason: None,
+        max_prompt_tokens: 49_152,
+    }
+}
+
+/// Catalogue with one model needing 2 devices. Returns a temp path.
+fn write_catalogue() -> std::path::PathBuf {
+    let toml = r#"
+[[models]]
+id = "big-model"
+harness = "candle"
+min_devices = 2
+"#;
+    let path = std::env::temp_dir().join("cortex_test_feasibility_models.toml");
+    std::fs::write(&path, toml).unwrap();
+    path
+}
+
+async fn fleet_with(big_healthy: bool, big_devices: usize) -> Arc<CortexState> {
+    let cat = write_catalogue();
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![
+            NeuronEndpoint {
+                name: "small".into(),
+                endpoint: "http://127.0.0.1:1".into(),
+            },
+            NeuronEndpoint {
+                name: "big".into(),
+                endpoint: "http://127.0.0.1:2".into(),
+            },
+        ],
+        models_config: cat.to_string_lossy().into_owned(),
+        entitlements: Default::default(),
+        upstream: Default::default(),
+    };
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        // "small" is healthy but only has 1 GPU → not feasible for the model.
+        let small = nodes.get_mut("small").unwrap();
+        small.healthy = true;
+        small.discovery = Some(discovery("small", 1));
+        // "big" has enough GPUs but its health is the variable under test.
+        let big = nodes.get_mut("big").unwrap();
+        big.healthy = big_healthy;
+        big.discovery = Some(discovery("big", big_devices));
+    }
+    fleet
+}
+
+#[tokio::test]
+async fn feasible_node_unhealthy_is_transient_503() {
+    // big (2 GPU, the only feasible node) is unhealthy; small (1 GPU) is
+    // healthy but can't host the model → retryable, not a permanent 404.
+    let fleet = fleet_with(false, 2).await;
+    let err = router::resolve(&fleet, "big-model")
+        .await
+        .expect_err("model can't be served right now");
+    assert!(
+        matches!(err, RouteError::FeasibleNodeUnhealthy { .. }),
+        "expected FeasibleNodeUnhealthy, got {err:?}"
+    );
+    assert_eq!(err.http_status(), 503);
+    assert_eq!(err.retry_after_secs(), Some(3));
+    assert_eq!(err.code(), "service_unavailable");
+}
+
+#[tokio::test]
+async fn no_node_can_ever_satisfy_is_permanent_404() {
+    // big is healthy but only has 1 GPU now (e.g. topology genuinely can't
+    // satisfy min_devices=2 anywhere) → permanent, non-retryable 404.
+    let fleet = fleet_with(true, 1).await;
+    let err = router::resolve(&fleet, "big-model")
+        .await
+        .expect_err("no feasible topology");
+    assert!(
+        matches!(err, RouteError::NoFeasibleNeuron { .. }),
+        "expected NoFeasibleNeuron, got {err:?}"
+    );
+    assert_eq!(err.http_status(), 404);
+    assert_eq!(err.retry_after_secs(), None);
+}

crates/cortex-gateway/tests/load_routing.rs

@@ -0,0 +1,190 @@
+//! Load-aware routing across replicas (#55).
+//!
+//! When a model is loaded on more than one healthy neuron, the router picks
+//! the least-busy replica using the per-model admission load each neuron
+//! reports on `GET /health` (#53), rather than always taking the first.
+
+mod common;
+
+use axum::Json;
+use axum::extract::Path;
+use axum::http::{StatusCode, header};
+use axum::response::IntoResponse;
+use axum::routing::{get, post};
+use cortex_core::config::{
+    EvictionSettings, EvictionStrategy, GatewayConfig, GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::discovery::ModelLoad;
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::{Value, json};
+use std::sync::Arc;
+use tokio::net::TcpListener;
+
+/// Seed a node as healthy with `test-model` loaded and a given admission load.
+async fn seed_loaded(fleet: &CortexState, node: &str, in_flight: usize, queue_depth: usize) {
+    let mut nodes = fleet.nodes.write().await;
+    let n = nodes.get_mut(node).expect("node exists");
+    n.healthy = true;
+    n.models.insert(
+        "test-model".into(),
+        ModelEntry {
+            id: "test-model".into(),
+            status: ModelStatus::Loaded,
+            last_accessed: None,
+            vram_estimate_mb: Some(8000),
+            capabilities: Vec::new(),
+            tool_call: false,
+            reasoning: false,
+            limit: None,
+        },
+    );
+    n.model_load.insert(
+        "test-model".into(),
+        ModelLoad {
+            id: "test-model".into(),
+            in_flight,
+            queue_depth,
+        },
+    );
+}
+
+/// Build a gateway state over two mock neurons (no poller; we seed state).
+async fn two_neuron_fleet(endpoint_a: &str, endpoint_b: &str) -> Arc<CortexState> {
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![
+            NeuronEndpoint {
+                name: "node-a".into(),
+                endpoint: endpoint_a.to_string(),
+            },
+            NeuronEndpoint {
+                name: "node-b".into(),
+                endpoint: endpoint_b.to_string(),
+            },
+        ],
+        models_config: "/dev/null".into(),
+        entitlements: Default::default(),
+        upstream: Default::default(),
+    };
+    Arc::new(CortexState::from_config(&config))
+}
+
+#[tokio::test]
+async fn routes_to_least_busy_replica() {
+    let neuron_a = common::spawn_mock_neuron().await;
+    let neuron_b = common::spawn_mock_neuron().await;
+    let fleet = two_neuron_fleet(&neuron_a, &neuron_b).await;
+
+    // A is busy (1 running + 3 queued), B is idle.
+    seed_loaded(&fleet, "node-a", 1, 3).await;
+    seed_loaded(&fleet, "node-b", 0, 0).await;
+
+    let route = cortex_gateway::router::resolve(&fleet, "test-model")
+        .await
+        .expect("model is loaded on both nodes");
+    assert_eq!(route.node_name, "node-b", "should pick the idle replica");
+
+    // Flip the load: now B is the busy one.
+    seed_loaded(&fleet, "node-a", 0, 0).await;
+    seed_loaded(&fleet, "node-b", 1, 5).await;
+    let route = cortex_gateway::router::resolve(&fleet, "test-model")
+        .await
+        .expect("still loaded");
+    assert_eq!(route.node_name, "node-a", "should follow the lighter load");
+}
+
+/// Mock neuron whose inference endpoint always returns a #63 backpressure
+/// envelope (503 + Retry-After) — simulating a saturated neuron.
+async fn spawn_busy_neuron() -> String {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    let base_url = format!("http://{addr}");
+    let inference_url = base_url.clone();
+    let app = axum::Router::new()
+        .route(
+            "/models/{model_id}/endpoint",
+            get(move |Path(_): Path<String>| {
+                let url = inference_url.clone();
+                async move { Json(json!({ "url": url })) }
+            }),
+        )
+        .route(
+            "/v1/chat/completions",
+            post(|| async {
+                let body = json!({"error": {
+                    "message": "model is busy (admission queue full); retry shortly",
+                    "type": "rate_limit_error",
+                    "code": "rate_limit_exceeded",
+                    "param": null
+                }});
+                (
+                    StatusCode::SERVICE_UNAVAILABLE,
+                    [(header::RETRY_AFTER, "6")],
+                    Json(body),
+                )
+                    .into_response()
+            }),
+        );
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    base_url
+}
+
+#[tokio::test]
+async fn neuron_backpressure_is_propagated_intact() {
+    // A saturated neuron's 503 + Retry-After + envelope must reach the client
+    // verbatim — not unwrapped, remapped, or stripped (#55 / #63).
+    let neuron = spawn_busy_neuron().await;
+    let fleet = two_neuron_fleet(&neuron, &neuron).await;
+    seed_loaded(&fleet, "node-a", 1, 8).await;
+
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+
+    let resp = reqwest::Client::new()
+        .post(format!("http://{addr}/v1/chat/completions"))
+        .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": "hi"}]}))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::SERVICE_UNAVAILABLE);
+    assert_eq!(
+        resp.headers()
+            .get(reqwest::header::RETRY_AFTER)
+            .and_then(|v| v.to_str().ok()),
+        Some("6"),
+        "Retry-After must survive the proxy"
+    );
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "rate_limit_exceeded");
+}
+
+#[tokio::test]
+async fn ties_break_deterministically_by_name() {
+    let neuron_a = common::spawn_mock_neuron().await;
+    let neuron_b = common::spawn_mock_neuron().await;
+    let fleet = two_neuron_fleet(&neuron_a, &neuron_b).await;
+
+    // Equal load on both → stable pick (lowest node name).
+    seed_loaded(&fleet, "node-a", 0, 0).await;
+    seed_loaded(&fleet, "node-b", 0, 0).await;
+
+    let route = cortex_gateway::router::resolve(&fleet, "test-model")
+        .await
+        .expect("loaded");
+    assert_eq!(route.node_name, "node-a", "ties break by name");
+}

crates/cortex-gateway/tests/metering.rs

@@ -0,0 +1,209 @@
+//! Integration tests for per-request token metering (#51).
+//!
+//! Drives authenticated requests through the gateway to a mock neuron that
+//! reports a fixed `usage` object, then asserts the EntitlementProvider's
+//! spend ledger reflects cumulative per-key spend and that reservations
+//! settle to actual (no outstanding reserved tokens once requests complete).
+
+mod common;
+
+use cortex_core::config::{
+    ApiKeyConfig, EntitlementsConfig, EvictionSettings, EvictionStrategy, GatewayConfig,
+    GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::entitlements::{CapWindow, Principal};
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::json;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::net::TcpListener;
+
+const ACCOUNT: &str = "acct-meter";
+const KEY_ID: &str = "key-meter";
+const BEARER: &str = "sk-meter";
+
+/// The mock neuron (common::spawn_mock_neuron) reports this fixed usage on
+/// every chat completion.
+const PROMPT_PER_REQ: u64 = 10;
+const COMPLETION_PER_REQ: u64 = 5;
+
+async fn spawn_metered_gateway(neuron_url: &str) -> (Arc<CortexState>, String) {
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron_url.to_string(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements: EntitlementsConfig {
+            require_auth: true,
+            keys: vec![ApiKeyConfig {
+                key: BEARER.into(),
+                account_id: ACCOUNT.into(),
+                key_id: Some(KEY_ID.into()),
+                hard_cap: Some(1_000_000),
+                window: CapWindow::Balance,
+            }],
+        },
+        upstream: Default::default(),
+    };
+
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let node = nodes.get_mut("mock-node").unwrap();
+        node.healthy = true;
+        node.models.insert(
+            "test-model".into(),
+            ModelEntry {
+                id: "test-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: Some(8000),
+                capabilities: Vec::new(),
+                tool_call: false,
+                reasoning: false,
+                limit: None,
+            },
+        );
+    }
+
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    (fleet, format!("http://{addr}"))
+}
+
+fn principal() -> Principal {
+    Principal {
+        account_id: ACCOUNT.into(),
+        key_id: KEY_ID.into(),
+    }
+}
+
+/// Poll the provider ledger until settled spend reaches `expected` (settle
+/// runs in a spawned task after the response stream finishes) or time out.
+async fn await_spent(fleet: &CortexState, expected: u64) -> u64 {
+    let principal = principal();
+    for _ in 0..100 {
+        let snap = fleet.entitlements.snapshot(&principal).await.unwrap();
+        if snap.spent >= expected {
+            return snap.spent;
+        }
+        tokio::time::sleep(Duration::from_millis(20)).await;
+    }
+    fleet.entitlements.snapshot(&principal).await.unwrap().spent
+}
+
+#[tokio::test]
+async fn cumulative_spend_is_metered_per_key() {
+    let neuron = common::spawn_mock_neuron().await;
+    let (fleet, gateway) = spawn_metered_gateway(&neuron).await;
+    let client = reqwest::Client::new();
+
+    const N: u64 = 3;
+    for _ in 0..N {
+        let resp = client
+            .post(format!("{gateway}/v1/chat/completions"))
+            .bearer_auth(BEARER)
+            .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": "hi"}]}))
+            .send()
+            .await
+            .unwrap();
+        assert_eq!(resp.status(), reqwest::StatusCode::OK);
+        // Drain the body so the response stream finishes and metering settles.
+        let _ = resp.bytes().await.unwrap();
+    }
+
+    let expected = N * (PROMPT_PER_REQ + COMPLETION_PER_REQ);
+    let spent = await_spent(&fleet, expected).await;
+    assert_eq!(
+        spent, expected,
+        "ledger must reflect cumulative per-key spend"
+    );
+
+    // Reservations settled to actual — nothing left outstanding.
+    let snap = fleet.entitlements.snapshot(&principal()).await.unwrap();
+    assert_eq!(snap.reserved, 0, "all reservations must settle/release");
+    assert_eq!(snap.hard_cap, Some(1_000_000));
+}
+
+#[tokio::test]
+async fn anonymous_request_records_no_spend() {
+    // require_auth=false so the unauthenticated request is served, but with
+    // no principal it must not touch any ledger.
+    let neuron = common::spawn_mock_neuron().await;
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron.clone(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements: EntitlementsConfig::default(),
+        upstream: Default::default(),
+    };
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let node = nodes.get_mut("mock-node").unwrap();
+        node.healthy = true;
+        node.models.insert(
+            "test-model".into(),
+            ModelEntry {
+                id: "test-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: Some(8000),
+                capabilities: Vec::new(),
+                tool_call: false,
+                reasoning: false,
+                limit: None,
+            },
+        );
+    }
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+
+    let resp = reqwest::Client::new()
+        .post(format!("http://{addr}/v1/chat/completions"))
+        .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": "hi"}]}))
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    let _ = resp.bytes().await.unwrap();
+
+    // An unconfigured principal has a zeroed snapshot — nothing was metered.
+    let snap = fleet
+        .entitlements
+        .snapshot(&Principal {
+            account_id: "nobody".into(),
+            key_id: "nobody".into(),
+        })
+        .await
+        .unwrap();
+    assert_eq!(snap.spent, 0);
+}

crates/cortex-gateway/tests/model_cost.rs

@@ -0,0 +1,132 @@
+//! Issue #68: the `cost` wire contract on `GET /v1/models`.
+//!
+//! `cost` is operator-set pricing sourced from the `models.toml` catalogue
+//! profile (the source of truth today; the marketplace clearing house #59
+//! later — both must read the same value metering/#51 bills against). The
+//! shape is the models.dev/opencode convention: **USD per 1,000,000 tokens,
+//! as JSON numbers**, with optional `cache_read`/`cache_write` tiers. This
+//! test pins:
+//!   - the units/shape (per-million floats, not per-token, not strings);
+//!   - that cache fields flow through when present and are omitted otherwise;
+//!   - the load-bearing **absent vs `0.0`** distinction (#68): a model with
+//!     no catalogue `cost` omits the key entirely (price unknown), distinct
+//!     from an explicit `0.0` (intentionally free).
+//!
+//! Catalogue-only models surface via Pass 1 of `list_models` even with no
+//! feasible neuron, so this is hermetic — no nodes or poller needed.
+
+use cortex_core::config::{
+    EvictionSettings, EvictionStrategy, GatewayConfig, GatewaySettings, NeuronEndpoint,
+};
+use cortex_gateway::state::CortexState;
+use std::sync::Arc;
+use tokio::net::TcpListener;
+
+#[tokio::test]
+async fn v1_models_cost_units_shape_and_absent_vs_zero() {
+    // Three catalogue models exercise the whole contract: a priced model
+    // with cache tiers, an intentionally-free model (explicit 0.0), and an
+    // unpriced model (no `cost` block at all).
+    let models_toml = r#"
+[[models]]
+id = "priced-model"
+harness = "candle"
+cost.input = 0.5
+cost.output = 1.5
+cost.cache_read = 0.05
+cost.cache_write = 0.6
+
+[[models]]
+id = "free-model"
+harness = "candle"
+cost.input = 0.0
+cost.output = 0.0
+
+[[models]]
+id = "unpriced-model"
+harness = "candle"
+"#;
+    let cat_path = std::env::temp_dir().join("cortex_test_issue68_models.toml");
+    std::fs::write(&cat_path, models_toml).unwrap();
+
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        // Never contacted: build_app does not spawn the poller, so the
+        // catalogue alone drives /v1/models.
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: "http://127.0.0.1:1".into(),
+        }],
+        models_config: cat_path.to_string_lossy().into_owned(),
+        entitlements: Default::default(),
+        upstream: Default::default(),
+    };
+
+    let fleet = Arc::new(CortexState::from_config(&config));
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+
+    let body: serde_json::Value = reqwest::Client::new()
+        .get(format!("http://{addr}/v1/models"))
+        .send()
+        .await
+        .unwrap()
+        .json()
+        .await
+        .unwrap();
+
+    let data = body["data"].as_array().expect("data is an array");
+    let entry = |id: &str| {
+        data.iter()
+            .find(|m| m["id"] == id)
+            .unwrap_or_else(|| panic!("{id} present in /v1/models"))
+            .clone()
+    };
+
+    // Priced model: exact values flow through as JSON numbers (USD per 1M
+    // tokens). If anything rescaled by 10⁶ or stringified, these fail.
+    let priced = entry("priced-model");
+    assert_eq!(priced["cost"]["input"], 0.5);
+    assert_eq!(priced["cost"]["output"], 1.5);
+    assert_eq!(priced["cost"]["cache_read"], 0.05);
+    assert_eq!(priced["cost"]["cache_write"], 0.6);
+    assert!(
+        priced["cost"]["input"].is_number(),
+        "cost.input must be a JSON number, not a string"
+    );
+
+    // Intentionally free: cost present, rates explicitly 0.0. Unset cache
+    // tiers are omitted (skip_serializing_if), not emitted as null/0.
+    let free = entry("free-model");
+    assert_eq!(free["cost"]["input"], 0.0);
+    assert_eq!(free["cost"]["output"], 0.0);
+    assert!(
+        free["cost"].get("cache_read").is_none(),
+        "absent cache tiers must be omitted, not null"
+    );
+    assert!(free["cost"].get("cache_write").is_none());
+
+    // Unpriced: the whole `cost` object is omitted — "price unknown",
+    // distinct from the free model's explicit 0.0. This is the #68
+    // distinction opencode needs to avoid showing $0 for a model whose
+    // price simply hasn't been declared.
+    let unpriced = entry("unpriced-model");
+    assert!(
+        unpriced.get("cost").is_none(),
+        "a model with no catalogue cost must omit `cost` entirely, got {:?}",
+        unpriced.get("cost")
+    );
+
+    let _ = std::fs::remove_file(&cat_path);
+}

crates/cortex-gateway/tests/model_limits.rs

@@ -8,8 +8,11 @@
 //!   - `cost` from the catalogue profile (operator-set pricing).
 //!   - `tool_call` / `reasoning` from the neuron's runtime detection (OR-ed in)
 //!
-//! Also a regression guard for the removal of `max_model_len` — the misnamed,
+//! Also asserts the flat, vLLM-convention duplicates (`max_model_len`,
-//! unconsumed vLLM-ism that this contract replaces.
+//! `max_input_tokens`, `max_output_tokens`) mirror `limit` (#78): the
+//! earlier removal of `max_model_len` as "unconsumed" was wrong — Hermes
+//! Agent (and the wider OpenAI client ecosystem) probes those flat keys
+//! and cannot see `limit.context`.
 
 use cortex_core::config::{
     EvictionSettings, EvictionStrategy, GatewayConfig, GatewaySettings, NeuronEndpoint,
@@ -55,6 +58,7 @@ capabilities = ["text"]
         }],
         models_config: cat_path.to_string_lossy().into_owned(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -84,6 +88,21 @@ capabilities = ["text"]
                 }),
             },
         );
+        // A model with no derivable limit: the flat #78 fields must be
+        // OMITTED (absent-vs-zero is load-bearing), never 0 or a guess.
+        node.models.insert(
+            "no-limit-model".into(),
+            ModelEntry {
+                id: "no-limit-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: None,
+                capabilities: vec!["text".into()],
+                tool_call: false,
+                reasoning: false,
+                limit: None,
+            },
+        );
     }
 
     let app = cortex_gateway::build_app(Arc::clone(&fleet));
@@ -122,11 +141,26 @@ capabilities = ["text"]
     assert_eq!(entry["tool_call"], true);
     assert_eq!(entry["reasoning"], true);
 
-    // Regression guard: the removed, unconsumed vLLM-ism must not reappear.
+    // Flat ecosystem duplicates (#78) mirror the advertised `limit` so
-    assert!(
+    // vLLM-convention probes (Hermes Agent) auto-detect the window.
-        entry.get("max_model_len").is_none(),
+    assert_eq!(entry["max_model_len"], 49152);
-        "max_model_len was removed; /v1/models must not advertise it"
+    assert_eq!(entry["max_input_tokens"], 40960);
-    );
+    assert_eq!(entry["max_output_tokens"], 8192);
+
+    // No limit → flat fields omitted entirely, never 0 or a guess.
+    let unknown = body["data"]
+        .as_array()
+        .unwrap()
+        .iter()
+        .find(|m| m["id"] == "no-limit-model")
+        .expect("no-limit-model present in /v1/models");
+    assert!(unknown.get("limit").is_none());
+    for key in ["max_model_len", "max_input_tokens", "max_output_tokens"] {
+        assert!(
+            unknown.get(key).is_none(),
+            "{key} must be omitted when the window is unknown"
+        );
+    }
 
     let _ = std::fs::remove_file(&cat_path);
 }

crates/cortex-gateway/tests/poller.rs

@@ -32,6 +32,7 @@ async fn test_poller_discovers_models() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -84,6 +85,7 @@ async fn test_poller_updates_gateway_models_endpoint() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -156,6 +158,7 @@ async fn test_models_endpoint_unions_capabilities_across_nodes() {
         ],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -219,6 +222,7 @@ async fn test_poller_marks_unreachable_node_unhealthy() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -228,10 +232,26 @@ async fn test_poller_marks_unreachable_node_unhealthy() {
         nodes.get_mut("dead-node").unwrap().healthy = true;
     }
 
+    // Debounce (#53 follow-up): a single missed poll must NOT evict a
+    // previously-healthy node — a busy neuron briefly slow to answer
+    // shouldn't yank its models out of routing.
     cortex_gateway::poller::poll_once(&fleet).await;
+    assert!(
+        fleet.nodes.read().await.get("dead-node").unwrap().healthy,
+        "one failed poll should not mark a healthy node unhealthy"
+    );
 
-    let nodes = fleet.nodes.read().await;
+    // It flips unhealthy only after POLL_FAILURE_THRESHOLD (3) consecutive
-    assert!(!nodes.get("dead-node").unwrap().healthy);
+    // failures.
+    cortex_gateway::poller::poll_once(&fleet).await;
+    cortex_gateway::poller::poll_once(&fleet).await;
+    assert!(
+        !fleet.nodes.read().await.get("dead-node").unwrap().healthy,
+        "three consecutive failed polls should mark the node unhealthy"
+    );
+
+    // A subsequent successful poll would reset the counter and restore
+    // health; covered implicitly by the discovery tests above.
 }
 
 #[tokio::test]
@@ -257,6 +277,7 @@ async fn test_poller_removes_stale_models() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -288,6 +309,7 @@ async fn test_poller_removes_stale_models() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet2 = Arc::new(CortexState::from_config(&config2));
@@ -370,6 +392,7 @@ async fn test_poller_captures_activation_from_health() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));
@@ -415,6 +438,7 @@ async fn test_poller_parses_recovering_status() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
 
     let fleet = Arc::new(CortexState::from_config(&config));

crates/cortex-gateway/tests/prompt_prevalidation.rs

@@ -0,0 +1,175 @@
+//! Fail-fast prompt pre-validation + advisory client hints (#56).
+//!
+//! cortex refuses a prompt that already exceeds the model's advertised
+//! context window before dispatching to neuron — the same #60
+//! `context_length_exceeded` envelope neuron would emit, just earlier — and
+//! attaches an advisory `X-Helexa-Advice` header for fingerprinted clients.
+
+use axum::Json;
+use axum::extract::Path;
+use axum::routing::{get, post};
+use cortex_core::config::{
+    EvictionSettings, EvictionStrategy, GatewayConfig, GatewaySettings, NeuronEndpoint,
+};
+use cortex_core::harness::ModelLimit;
+use cortex_core::node::{ModelEntry, ModelStatus};
+use cortex_gateway::state::CortexState;
+use serde_json::{Value, json};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicU64, Ordering};
+use tokio::net::TcpListener;
+
+/// Mock neuron with a hit counter, so a test can prove a request was (or
+/// wasn't) dispatched past the gateway's pre-validation.
+async fn spawn_counting_neuron() -> (String, Arc<AtomicU64>) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    let base_url = format!("http://{addr}");
+    let inference_url = base_url.clone();
+    let hits = Arc::new(AtomicU64::new(0));
+    let sink = Arc::clone(&hits);
+    let app = axum::Router::new()
+        .route(
+            "/models/{model_id}/endpoint",
+            get(move |Path(_): Path<String>| {
+                let url = inference_url.clone();
+                async move { Json(json!({ "url": url })) }
+            }),
+        )
+        .route(
+            "/v1/chat/completions",
+            post(move || {
+                let sink = Arc::clone(&sink);
+                async move {
+                    sink.fetch_add(1, Ordering::SeqCst);
+                    Json(json!({
+                        "id": "c", "object": "chat.completion", "created": 1_700_000_000_u64,
+                        "model": "test-model",
+                        "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}],
+                        "usage": {"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}
+                    }))
+                }
+            }),
+        );
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    (base_url, hits)
+}
+
+/// Gateway over one neuron with `test-model` loaded and a tiny advertised
+/// context window (so a modest prompt overflows it).
+async fn spawn_gateway(neuron: &str, context: usize) -> String {
+    let config = GatewayConfig {
+        gateway: GatewaySettings {
+            listen: "127.0.0.1:0".into(),
+            metrics_listen: "127.0.0.1:0".into(),
+        },
+        eviction: EvictionSettings {
+            strategy: EvictionStrategy::Lru,
+            defrag_after_cycles: 0,
+        },
+        neurons: vec![NeuronEndpoint {
+            name: "mock-node".into(),
+            endpoint: neuron.to_string(),
+        }],
+        models_config: "/dev/null".into(),
+        entitlements: Default::default(),
+        upstream: Default::default(),
+    };
+    let fleet = Arc::new(CortexState::from_config(&config));
+    {
+        let mut nodes = fleet.nodes.write().await;
+        let n = nodes.get_mut("mock-node").unwrap();
+        n.healthy = true;
+        n.models.insert(
+            "test-model".into(),
+            ModelEntry {
+                id: "test-model".into(),
+                status: ModelStatus::Loaded,
+                last_accessed: None,
+                vram_estimate_mb: Some(8000),
+                capabilities: Vec::new(),
+                tool_call: false,
+                reasoning: false,
+                limit: Some(ModelLimit {
+                    context,
+                    input: None,
+                    output: 16,
+                }),
+            },
+        );
+    }
+    let app = cortex_gateway::build_app(Arc::clone(&fleet));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+#[tokio::test]
+async fn over_long_prompt_is_rejected_before_dispatch() {
+    let (neuron, hits) = spawn_counting_neuron().await;
+    let gateway = spawn_gateway(&neuron, 50).await; // tiny 50-token window
+
+    // ~1200 chars → ~300 est tokens, well over 50.
+    let big = "word ".repeat(240);
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .header("user-agent", "litellm/1.0")
+        .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": big}]}))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::BAD_REQUEST);
+    // Advisory hint for the fingerprinted client (header only, never body).
+    assert!(
+        resp.headers().get("x-helexa-advice").is_some(),
+        "litellm should get advice"
+    );
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "context_length_exceeded");
+    assert_eq!(body["error"]["max"], 50);
+    // Refused at the edge — neuron never saw it.
+    assert_eq!(hits.load(Ordering::SeqCst), 0);
+}
+
+#[tokio::test]
+async fn within_context_passes_through() {
+    let (neuron, hits) = spawn_counting_neuron().await;
+    let gateway = spawn_gateway(&neuron, 4096).await;
+
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": "hi"}]}))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::OK);
+    let _ = resp.bytes().await.unwrap();
+    assert_eq!(hits.load(Ordering::SeqCst), 1, "served by neuron");
+}
+
+#[tokio::test]
+async fn unknown_client_gets_no_advice_header() {
+    let (neuron, _hits) = spawn_counting_neuron().await;
+    let gateway = spawn_gateway(&neuron, 50).await;
+
+    let big = "word ".repeat(240);
+    let resp = reqwest::Client::new()
+        .post(format!("{gateway}/v1/chat/completions"))
+        // no/unknown User-Agent → no advice, but still a clean 400
+        .json(&json!({"model": "test-model", "messages": [{"role": "user", "content": big}]}))
+        .send()
+        .await
+        .unwrap();
+
+    assert_eq!(resp.status(), reqwest::StatusCode::BAD_REQUEST);
+    assert!(resp.headers().get("x-helexa-advice").is_none());
+    let body: Value = resp.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "context_length_exceeded");
+}

crates/cortex-gateway/tests/proxy_basic.rs

@@ -118,6 +118,7 @@ async fn test_no_healthy_nodes() {
         }],
         models_config: "/dev/null".into(),
         entitlements: Default::default(),
+        upstream: Default::default(),
     };
     let fleet = std::sync::Arc::new(cortex_gateway::state::CortexState::from_config(&config));
 

crates/cortex-gateway/tests/upstream_chain.rs

@@ -0,0 +1,106 @@
+//! B3: the chained entitlement provider (local → upstream) and fail-closed
+//! semantics, exercised against a mock helexa-upstream `/authz/v1`.
+
+use axum::{Json, Router, routing::post};
+use cortex_core::config::{ApiKeyConfig, EntitlementsConfig, UpstreamClientConfig};
+use cortex_core::entitlements::{AuthError, EntitlementProvider};
+use cortex_gateway::entitlements_chain::ChainedEntitlementProvider;
+use cortex_gateway::entitlements_local::LocalEntitlementProvider;
+use cortex_gateway::entitlements_upstream::UpstreamEntitlementProvider;
+use serde_json::{Value, json};
+use tokio::net::TcpListener;
+
+/// Mock upstream: `mesh-key` resolves to a mesh account; anything else 401.
+/// reserve always grants reservation 1.
+async fn spawn_mock_upstream() -> String {
+    async fn resolve(Json(body): Json<Value>) -> axum::response::Response {
+        use axum::response::IntoResponse;
+        if body["api_key"] == "mesh-key" {
+            Json(json!({"principal": {"account_id": "mesh-acct", "key_id": "mesh-key-1"}}))
+                .into_response()
+        } else {
+            (
+                axum::http::StatusCode::UNAUTHORIZED,
+                Json(json!({"error": {"code": "invalid_api_key"}})),
+            )
+                .into_response()
+        }
+    }
+    async fn reserve() -> Json<Value> {
+        Json(json!({ "reservation_id": 1 }))
+    }
+    let app = Router::new()
+        .route("/authz/v1/resolve", post(resolve))
+        .route("/authz/v1/reserve", post(reserve));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+fn local_with_key() -> LocalEntitlementProvider {
+    let cfg = EntitlementsConfig {
+        require_auth: false,
+        keys: vec![ApiKeyConfig {
+            key: "local-key".into(),
+            account_id: "op".into(),
+            key_id: None,
+            hard_cap: None,
+            window: Default::default(),
+        }],
+    };
+    LocalEntitlementProvider::from_config(&cfg)
+}
+
+fn chain(local: LocalEntitlementProvider, url: &str) -> ChainedEntitlementProvider {
+    let upstream = UpstreamEntitlementProvider::new(&UpstreamClientConfig {
+        enabled: true,
+        url: url.to_string(),
+        bearer: "client-secret".into(),
+        timeout_secs: 5,
+        served_usage_report_interval_secs: 60,
+    });
+    ChainedEntitlementProvider::new(local, upstream)
+}
+
+#[tokio::test]
+async fn local_key_resolves_locally() {
+    let url = spawn_mock_upstream().await;
+    let c = chain(local_with_key(), &url);
+    let p = c.resolve("local-key").await.expect("local resolves");
+    assert_eq!(p.account_id, "op");
+}
+
+#[tokio::test]
+async fn unknown_key_falls_through_to_upstream() {
+    let url = spawn_mock_upstream().await;
+    let c = chain(local_with_key(), &url);
+    let p = c.resolve("mesh-key").await.expect("upstream resolves");
+    assert_eq!(p.account_id, "mesh-acct");
+    assert_eq!(p.key_id, "mesh-key-1");
+}
+
+#[tokio::test]
+async fn unknown_everywhere_is_invalid_key() {
+    let url = spawn_mock_upstream().await;
+    let c = chain(local_with_key(), &url);
+    match c.resolve("nope").await {
+        Err(AuthError::InvalidKey) => {}
+        other => panic!("expected InvalidKey, got {other:?}"),
+    }
+}
+
+#[tokio::test]
+async fn upstream_unreachable_fails_closed_as_unavailable() {
+    // No mock — point at a dead port. A locally-unknown key must surface
+    // Unavailable (→ 503), never InvalidKey (→ 401).
+    let c = chain(local_with_key(), "http://127.0.0.1:1");
+    match c.resolve("some-mesh-key").await {
+        Err(AuthError::Unavailable { retry_after_secs }) => assert!(retry_after_secs > 0),
+        other => panic!("expected Unavailable, got {other:?}"),
+    }
+    // A local key still resolves without touching upstream.
+    assert_eq!(c.resolve("local-key").await.unwrap().account_id, "op");
+}

crates/helexa-bench/src/api.rs

@@ -35,6 +35,9 @@ pub fn api_routes(state: ApiState) -> Router {
         .route("/api/health", get(health))
         .route("/api/dimensions", get(dimensions))
         .route("/api/summary", get(summary))
+        .route("/api/scaling", get(scaling))
+        .route("/api/swap", get(swap))
+        .route("/api/capability", get(capability))
         .route("/api/series", get(series))
         .route("/api/runs", get(runs))
         .layer(CorsLayer::permissive())
@@ -73,6 +76,30 @@ async fn summary(
     store.summary().map(Json).map_err(err500)
 }
 
+/// Context-length scaling curves per (target, model) — prefill & decode
+/// tok/s vs context, with decode-flatness (#88).
+async fn scaling(
+    State(s): State<ApiState>,
+) -> Result<Json<Vec<crate::store::ScalingCurve>>, ApiError> {
+    let store = s.lock().await;
+    store.scaling().map(Json).map_err(err500)
+}
+
+/// Cold-load / model-swap costs per (target, model) — reload latency + cold
+/// first-request (#90).
+async fn swap(State(s): State<ApiState>) -> Result<Json<Vec<crate::store::SwapCost>>, ApiError> {
+    let store = s.lock().await;
+    store.swap_costs().map(Json).map_err(err500)
+}
+
+/// Capability-probe runs — stored artifacts + quality scores (#91).
+async fn capability(
+    State(s): State<ApiState>,
+) -> Result<Json<Vec<crate::store::CapabilityRun>>, ApiError> {
+    let store = s.lock().await;
+    store.capability_runs(false).map(Json).map_err(err500)
+}
+
 #[derive(Debug, Deserialize)]
 struct SeriesQuery {
     /// Optional — when omitted the store resolves the host serving this model.

crates/helexa-bench/src/client.rs

@@ -3,10 +3,10 @@
 //! `openai` targets use the OpenAI-compatible surface (preliminary).
 
 use crate::config::{TargetConfig, TargetKind};
-use anyhow::{Context, Result};
+use anyhow::{Context, Result, anyhow};
 use cortex_core::build_info::BuildInfo;
-use cortex_core::discovery::DiscoveryResponse;
+use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
-use cortex_core::harness::ModelInfo;
+use cortex_core::harness::{ModelInfo, ModelSpec};
 use cortex_core::openai::ModelsResponse;
 use std::time::Duration;
 
@@ -94,6 +94,84 @@ impl TargetClient {
         Ok(Some(disco))
     }
 
+    /// Runtime device health (neuron only): per-GPU VRAM used/free,
+    /// utilization, and temperature from `GET /health`. Bench samples this
+    /// around each measured run to record VRAM high-water + GPU telemetry
+    /// (#87). Returns `Ok(None)` for non-neuron targets; a soft `Ok(None)`
+    /// (not an error) on transport failure so a flaky `/health` never fails
+    /// a measurement.
+    pub async fn fetch_health(&self, target: &TargetConfig) -> Result<Option<HealthResponse>> {
+        if target.kind != TargetKind::Neuron {
+            return Ok(None);
+        }
+        let base = target.endpoint.trim_end_matches('/');
+        let health = self
+            .http
+            .get(format!("{base}/health"))
+            .timeout(META_TIMEOUT)
+            .send()
+            .await
+            .context("GET /health")?
+            .error_for_status()
+            .context("GET /health status")?
+            .json::<HealthResponse>()
+            .await
+            .context("decoding /health")?;
+        Ok(Some(health))
+    }
+
+    /// Unload a model (neuron only): `POST /models/unload {model_id}`.
+    /// Used by the deliberate swap-cost measurement (#90), never the sweep.
+    pub async fn unload_model(&self, target: &TargetConfig, model_id: &str) -> Result<()> {
+        let base = target.endpoint.trim_end_matches('/');
+        self.http
+            .post(format!("{base}/models/unload"))
+            .json(&serde_json::json!({ "model_id": model_id }))
+            .send()
+            .await
+            .context("POST /models/unload")?
+            .error_for_status()
+            .context("POST /models/unload status")?;
+        Ok(())
+    }
+
+    /// Load a model from a spec (neuron only): `POST /models/load`. neuron
+    /// returns synchronously once loaded, so the call duration is the reload
+    /// cost the swap-cost measurement records (#90).
+    pub async fn load_model(&self, target: &TargetConfig, spec: &ModelSpec) -> Result<()> {
+        let base = target.endpoint.trim_end_matches('/');
+        self.http
+            .post(format!("{base}/models/load"))
+            .json(spec)
+            // A cold load can take tens of seconds; use the full request
+            // timeout rather than the short metadata one.
+            .send()
+            .await
+            .context("POST /models/load")?
+            .error_for_status()
+            .context("POST /models/load status")?;
+        Ok(())
+    }
+
+    /// Reconstruct a reload [`ModelSpec`] from a model's `/models` entry.
+    /// Tensor-parallel is inferred from the device count; `quant` is left
+    /// `None` for neuron to resolve from the catalogue / its prior load.
+    pub fn spec_from_info(info: &ModelInfo) -> Result<ModelSpec> {
+        if info.devices.is_empty() {
+            return Err(anyhow!(
+                "model '{}' reports no devices; cannot reconstruct a load spec",
+                info.id
+            ));
+        }
+        Ok(ModelSpec {
+            model_id: info.id.clone(),
+            harness: info.harness.clone(),
+            quant: None,
+            tensor_parallel: (info.devices.len() > 1).then_some(info.devices.len() as u32),
+            devices: Some(info.devices.clone()),
+        })
+    }
+
     /// Warm models — those ready to serve without a cold load.
     ///
     /// Neuron: `GET /models` filtered to `status == "loaded"` (skips

crates/helexa-bench/src/config.rs

@@ -104,6 +104,37 @@ pub struct ScenarioConfig {
     /// Max generated tokens per request.
     #[serde(default = "default_max_tokens")]
     pub max_tokens: u64,
+    /// Concurrency levels (#89): one `concurrency:<n>` scenario per entry,
+    /// each firing N simultaneous streams. Defaults to empty (opt-in) — a
+    /// burst puts real load on a serving fleet, so operators enable it
+    /// deliberately, e.g. `concurrency_levels = [2, 4, 8]`.
+    #[serde(default)]
+    pub concurrency_levels: Vec<u32>,
+    /// Approximate prompt size (tokens) used by the concurrency scenarios.
+    #[serde(default = "default_concurrency_prompt_tokens")]
+    pub concurrency_prompt_tokens: u32,
+    /// Capability probes (#91): one `capability:<name>` scenario per entry,
+    /// each running a fixed prompt and storing the full output artifact for
+    /// quality scoring (manual now, LLM-judge later). Defaults to empty
+    /// (opt-in) — these generate long outputs and exist to compare reasoning
+    /// quality across models, not to run on every sweep by default.
+    #[serde(default)]
+    pub capability_probes: Vec<CapabilityProbe>,
+}
+
+/// One capability probe: a named prompt whose output is stored and scored
+/// for quality (#91). The probe is deterministic (temperature 0) so the
+/// same model+build produces a stable artifact to score.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct CapabilityProbe {
+    /// Stable id fragment — the scenario becomes `capability:<name>`.
+    pub name: String,
+    /// The prompt sent verbatim as the user message.
+    pub prompt: String,
+    /// Generation budget for the probe (planning answers are long; the
+    /// default 256 is too small). Defaults to 2048.
+    #[serde(default = "default_capability_max_tokens")]
+    pub max_tokens: u64,
 }
 
 impl Default for ScenarioConfig {
@@ -111,6 +142,9 @@ impl Default for ScenarioConfig {
         ScenarioConfig {
             prompt_sizes: default_prompt_sizes(),
             max_tokens: default_max_tokens(),
+            concurrency_levels: Vec::new(),
+            concurrency_prompt_tokens: default_concurrency_prompt_tokens(),
+            capability_probes: Vec::new(),
         }
     }
 }
@@ -187,6 +221,12 @@ fn default_prompt_sizes() -> Vec<u32> {
 fn default_max_tokens() -> u64 {
     256
 }
+fn default_concurrency_prompt_tokens() -> u32 {
+    512
+}
+fn default_capability_max_tokens() -> u64 {
+    2048
+}
 
 #[cfg(test)]
 // Jail's closure must return figment::Result; the large-Err type is

crates/helexa-bench/src/main.rs

@@ -43,6 +43,33 @@ enum Command {
         #[arg(short, long, default_value = "helexa-bench.toml")]
         config: String,
     },
+    /// Measure cold-load / model-swap cost (#90): for each neuron target's
+    /// warm models, unload → time reload → time a cold first request, recorded
+    /// under scenario "swap". DELIBERATE — takes each model offline for its
+    /// reload, so run it in a maintenance window, not against live traffic.
+    SwapCost {
+        #[arg(short, long, default_value = "helexa-bench.toml")]
+        config: String,
+    },
+    /// Attach a quality score to a capability-probe run (#91). Find run ids
+    /// with `report --capability`. `--scorer` records who scored it
+    /// (defaults to "manual"); a future LLM-judge would set e.g. "llm:…".
+    Score {
+        #[arg(short, long, default_value = "helexa-bench.toml")]
+        config: String,
+        /// Override the SQLite path (skips reading the config file).
+        #[arg(long)]
+        db: Option<String>,
+        /// The run id to score.
+        #[arg(long)]
+        id: i64,
+        /// The quality score to attach (scale is the operator's rubric).
+        #[arg(long)]
+        score: f64,
+        /// Who/what produced the score.
+        #[arg(long, default_value = "manual")]
+        scorer: String,
+    },
     /// Render recorded results. Uses `--db` if given, else the db_path
     /// from `--config`.
     Report {
@@ -54,6 +81,19 @@ enum Command {
         /// Output format.
         #[arg(long, default_value = "md")]
         format: Format,
+        /// Render the context-length scaling view (prefill & decode tok/s
+        /// vs context per model, with decode-flatness) instead of the flat
+        /// results table (#88).
+        #[arg(long)]
+        scaling: bool,
+        /// Render the cold-load / model-swap cost view (#90) instead of the
+        /// flat results table.
+        #[arg(long)]
+        swap: bool,
+        /// Render the capability-probe view (#91): stored artifacts + quality
+        /// scores, with per-model median.
+        #[arg(long)]
+        capability: bool,
     },
 }
 
@@ -122,16 +162,79 @@ async fn run(cli: Cli) -> Result<()> {
             );
             Ok(())
         }
-        Command::Report { config, db, format } => {
+        Command::SwapCost { config } => {
+            let cfg = load_config(&config)?;
+            require_targets(&cfg)?;
+            let sweeper = Sweeper::new(cfg)?;
+            tracing::warn!(
+                "swap-cost: cycling each warm model (unload → reload → cold request); models go offline during reload"
+            );
+            let summary = sweeper.swap_cost_once().await?;
+            tracing::info!(
+                measured = summary.measured,
+                failed = summary.failed,
+                unreachable = summary.targets_unreachable,
+                "swap-cost measurement complete"
+            );
+            Ok(())
+        }
+        Command::Score {
+            config,
+            db,
+            id,
+            score,
+            scorer,
+        } => {
             let db_path = match db {
                 Some(p) => p,
                 None => load_config(&config)?.bench.db_path,
             };
             let store = Store::open(&db_path)?;
-            let rows = store.report_rows()?;
+            match store.set_score(id, score, &scorer)? {
-            let rendered = match format {
+                0 => anyhow::bail!("no run with id {id}"),
-                Format::Md => report::render_markdown(&rows),
+                _ => {
-                Format::Json => report::render_json(&rows)?,
+                    println!("scored run {id}: {score} ({scorer})");
+                    Ok(())
+                }
+            }
+        }
+        Command::Report {
+            config,
+            db,
+            format,
+            scaling,
+            swap,
+            capability,
+        } => {
+            let db_path = match db {
+                Some(p) => p,
+                None => load_config(&config)?.bench.db_path,
+            };
+            let store = Store::open(&db_path)?;
+            let rendered = if capability {
+                let runs = store.capability_runs(false)?;
+                match format {
+                    Format::Md => report::render_capability_markdown(&runs),
+                    Format::Json => report::render_capability_json(&runs)?,
+                }
+            } else if swap {
+                let costs = store.swap_costs()?;
+                match format {
+                    Format::Md => report::render_swap_markdown(&costs),
+                    Format::Json => report::render_swap_json(&costs)?,
+                }
+            } else if scaling {
+                let curves = store.scaling()?;
+                match format {
+                    Format::Md => report::render_scaling_markdown(&curves),
+                    Format::Json => report::render_scaling_json(&curves)?,
+                }
+            } else {
+                let rows = store.report_rows()?;
+                match format {
+                    Format::Md => report::render_markdown(&rows),
+                    Format::Json => report::render_json(&rows)?,
+                }
             };
             println!("{rendered}");
             Ok(())

crates/helexa-bench/src/report.rs

@@ -3,28 +3,36 @@
 //! doc: engine, model, prompt tok, TTFT (s), decode tok/s, total (s),
 //! plus the build SHA each cell was measured against.
 
-use crate::store::ReportRow;
+use crate::store::{CapabilityRun, ReportRow, ScalingCurve, SwapCost};
 use anyhow::Result;
 
 pub fn render_markdown(rows: &[ReportRow]) -> String {
     let mut out = String::new();
     out.push_str(
-        "| engine | model | prompt tok | TTFT (s) | decode tok/s | total (s) | build | n |\n",
+        "| engine | model | prompt tok | prefill tok/s | TTFT (s) | TTFT p95 | \
+         decode tok/s | total (s) | total p95 | VRAM (GB) | conc | queue ms | rej | build | n |\n",
     );
-    out.push_str("|---|---|---:|---:|---:|---:|---|---:|\n");
+    out.push_str("|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|---:|\n");
     for r in rows {
         let ptok = r
             .prompt_tokens
             .map(|t| t.to_string())
             .unwrap_or_else(|| format!("~{}", r.prompt_size_approx));
         out.push_str(&format!(
-            "| {} | {} | {} | {} | {} | {} | `{}` | {} |\n",
+            "| {} | {} | {} | {} | {} | {} | {} | {} | {} | {} | {} | {} | {} | `{}` | {} |\n",
             r.target_name,
             r.model_id,
             ptok,
+            fmt_opt(r.prefill_tps_median, 1),
             fmt_opt(r.ttft_s_median, 3),
+            fmt_opt(r.ttft_s_p95, 3),
             fmt_opt(r.decode_tps_median, 1),
             fmt_opt(r.total_s_median, 3),
+            fmt_opt(r.total_s_p95, 3),
+            fmt_vram(r.vram_used_mb_median, r.vram_total_mb),
+            fmt_u64(r.concurrency),
+            fmt_opt(r.queue_wait_ms_median, 0),
+            fmt_opt(r.rejected_median, 0),
             r.git_sha,
             r.samples,
         ));
@@ -43,8 +51,23 @@ pub fn render_json(rows: &[ReportRow]) -> Result<String> {
                 "prompt_size_approx": r.prompt_size_approx,
                 "prompt_tokens": r.prompt_tokens,
                 "ttft_s_median": r.ttft_s_median,
+                "ttft_s_p95": r.ttft_s_p95,
+                "ttft_s_p99": r.ttft_s_p99,
                 "decode_tps_median": r.decode_tps_median,
                 "total_s_median": r.total_s_median,
+                "total_s_p95": r.total_s_p95,
+                "total_s_p99": r.total_s_p99,
+                "prefill_ms_median": r.prefill_ms_median,
+                "decode_ms_median": r.decode_ms_median,
+                "prefill_tps_median": r.prefill_tps_median,
+                "vram_used_mb_median": r.vram_used_mb_median,
+                "vram_total_mb": r.vram_total_mb,
+                "gpu_util_pct_median": r.gpu_util_pct_median,
+                "gpu_temp_c_median": r.gpu_temp_c_median,
+                "concurrency": r.concurrency,
+                "ttft_p95_load_s": r.ttft_p95_load_s,
+                "queue_wait_ms_median": r.queue_wait_ms_median,
+                "rejected_median": r.rejected_median,
                 "git_sha": r.git_sha,
                 "samples": r.samples,
                 "gpu": r.gpu,
@@ -54,6 +77,179 @@ pub fn render_json(rows: &[ReportRow]) -> Result<String> {
     Ok(serde_json::to_string_pretty(&arr)?)
 }
 
+/// Context-length scaling view (#88): one block per (target, model) with
+/// prefill & decode tok/s vs context, then the decode-flatness verdict.
+pub fn render_scaling_markdown(curves: &[ScalingCurve]) -> String {
+    let mut out = String::new();
+    for c in curves {
+        let gpu = c.gpu.as_deref().unwrap_or("");
+        out.push_str(&format!(
+            "### {} · {}  (`{}`{})\n\n",
+            c.target_name,
+            c.model_id,
+            c.git_sha,
+            if gpu.is_empty() {
+                String::new()
+            } else {
+                format!(", {gpu}")
+            },
+        ));
+        out.push_str("| ctx tok | prefill tok/s | decode tok/s | n |\n");
+        out.push_str("|---:|---:|---:|---:|\n");
+        for p in &c.points {
+            let ctx = p
+                .prompt_tokens
+                .map(|t| t.to_string())
+                .unwrap_or_else(|| format!("~{}", p.prompt_size));
+            out.push_str(&format!(
+                "| {} | {} | {} | {} |\n",
+                ctx,
+                fmt_opt(p.prefill_tps, 1),
+                fmt_opt(p.decode_tps, 1),
+                p.samples,
+            ));
+        }
+        match c.decode_flatness {
+            Some(f) => out.push_str(&format!(
+                "\ndecode flatness: {f:.2} — decode tok/s {} across the context range \
+                 ({})\n\n",
+                if f >= 0.9 {
+                    "holds"
+                } else if f >= 0.7 {
+                    "softens"
+                } else {
+                    "drops sharply"
+                },
+                if f >= 0.9 {
+                    "Gated-DeltaNet O(1) decode confirmed"
+                } else {
+                    "investigate where it breaks"
+                },
+            )),
+            None => out.push_str("\ndecode flatness: — (need ≥2 context points)\n\n"),
+        }
+    }
+    out
+}
+
+pub fn render_scaling_json(curves: &[ScalingCurve]) -> Result<String> {
+    Ok(serde_json::to_string_pretty(curves)?)
+}
+
+/// Cold-load / model-swap cost view (#90): reload latency + cold
+/// first-request per model.
+pub fn render_swap_markdown(costs: &[SwapCost]) -> String {
+    let mut out = String::new();
+    out.push_str(
+        "| engine | model | unload (s) | reload (s) | cold TTFT (s) | cold total (s) | build | n |\n",
+    );
+    out.push_str("|---|---|---:|---:|---:|---:|---|---:|\n");
+    for c in costs {
+        out.push_str(&format!(
+            "| {} | {} | {} | {} | {} | {} | `{}` | {} |\n",
+            c.target_name,
+            c.model_id,
+            fmt_ms_as_s(c.unload_ms_median),
+            fmt_ms_as_s(c.load_ms_median),
+            fmt_opt(c.cold_ttft_s_median, 3),
+            fmt_opt(c.cold_total_s_median, 3),
+            c.git_sha,
+            c.samples,
+        ));
+    }
+    out
+}
+
+pub fn render_swap_json(costs: &[SwapCost]) -> Result<String> {
+    Ok(serde_json::to_string_pretty(costs)?)
+}
+
+/// Capability-probe view (#91): per (model, probe) the median quality score
+/// (the A/B number), then each run's id, score, and an artifact snippet so
+/// unscored runs can be located and scored (`helexa-bench score --id …`).
+pub fn render_capability_markdown(runs: &[CapabilityRun]) -> String {
+    use std::collections::BTreeMap;
+    let mut groups: BTreeMap<(String, String, String), Vec<&CapabilityRun>> = BTreeMap::new();
+    for r in runs {
+        groups
+            .entry((
+                r.target_name.clone(),
+                r.model_id.clone(),
+                r.scenario_id.clone(),
+            ))
+            .or_default()
+            .push(r);
+    }
+    let mut out = String::new();
+    for ((target, model, scenario), rs) in groups {
+        let scores: Vec<f64> = rs.iter().filter_map(|r| r.quality_score).collect();
+        let median = median_slice(&scores);
+        out.push_str(&format!(
+            "### {target} · {model} · {scenario} — median score {} ({}/{} scored)\n\n",
+            median
+                .map(|m| format!("{m:.1}"))
+                .unwrap_or_else(|| "—".into()),
+            scores.len(),
+            rs.len(),
+        ));
+        out.push_str("| run | score | scorer | build | artifact (snippet) |\n");
+        out.push_str("|---:|---:|---|---|---|\n");
+        for r in rs {
+            out.push_str(&format!(
+                "| {} | {} | {} | `{}` | {} |\n",
+                r.id,
+                r.quality_score
+                    .map(|s| format!("{s:.1}"))
+                    .unwrap_or_else(|| "—".into()),
+                r.scorer.as_deref().unwrap_or("—"),
+                r.git_sha,
+                snippet(r.artifact.as_deref()),
+            ));
+        }
+        out.push('\n');
+    }
+    out
+}
+
+pub fn render_capability_json(runs: &[CapabilityRun]) -> Result<String> {
+    Ok(serde_json::to_string_pretty(runs)?)
+}
+
+/// First ~80 chars of an artifact on one line, for the table cell.
+fn snippet(artifact: Option<&str>) -> String {
+    match artifact {
+        Some(a) => {
+            let one_line: String = a.split_whitespace().collect::<Vec<_>>().join(" ");
+            let trimmed: String = one_line.chars().take(80).collect();
+            if one_line.chars().count() > 80 {
+                format!("{trimmed}…")
+            } else {
+                trimmed
+            }
+        }
+        None => "—".to_string(),
+    }
+}
+
+fn median_slice(v: &[f64]) -> Option<f64> {
+    if v.is_empty() {
+        return None;
+    }
+    let mut s = v.to_vec();
+    s.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    let lo = (s.len() - 1) / 2;
+    let hi = s.len() / 2;
+    Some((s[lo] + s[hi]) / 2.0)
+}
+
+/// Milliseconds rendered as seconds (reload costs read naturally in s).
+fn fmt_ms_as_s(ms: Option<f64>) -> String {
+    match ms {
+        Some(x) => format!("{:.2}", x / 1000.0),
+        None => "—".to_string(),
+    }
+}
+
 fn fmt_opt(v: Option<f64>, places: usize) -> String {
     match v {
         Some(x) => format!("{x:.places$}"),
@@ -61,9 +257,113 @@ fn fmt_opt(v: Option<f64>, places: usize) -> String {
     }
 }
 
+/// Integer cell (concurrency width); `—` when unset (non-concurrency rows).
+fn fmt_u64(v: Option<u64>) -> String {
+    match v {
+        Some(x) => x.to_string(),
+        None => "—".to_string(),
+    }
+}
+
+/// `used/total` in GB (e.g. `42.0/64.0`) — the headroom-at-a-glance cell.
+/// `used` alone if the node total is unknown; `—` if no telemetry.
+fn fmt_vram(used_mb: Option<f64>, total_mb: Option<u64>) -> String {
+    match (used_mb, total_mb) {
+        (Some(u), Some(t)) => format!("{:.1}/{:.1}", u / 1024.0, t as f64 / 1024.0),
+        (Some(u), None) => format!("{:.1}", u / 1024.0),
+        _ => "—".to_string(),
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::store::{ScalingCurve, ScalingPoint};
+
+    #[test]
+    fn capability_markdown_groups_with_median_and_snippet() {
+        let runs = vec![
+            CapabilityRun {
+                id: 7,
+                ts: "t".into(),
+                target_name: "beast".into(),
+                model_id: "m".into(),
+                scenario_id: "capability:plan".into(),
+                git_sha: "abc".into(),
+                quality_score: Some(8.0),
+                scorer: Some("manual".into()),
+                artifact: Some("A detailed plan with trade-offs and sequencing.".into()),
+            },
+            CapabilityRun {
+                id: 8,
+                ts: "t".into(),
+                target_name: "beast".into(),
+                model_id: "m".into(),
+                scenario_id: "capability:plan".into(),
+                git_sha: "abc".into(),
+                quality_score: Some(6.0),
+                scorer: Some("manual".into()),
+                artifact: Some("Shorter plan.".into()),
+            },
+        ];
+        let md = render_capability_markdown(&runs);
+        assert!(md.contains("capability:plan"));
+        assert!(md.contains("median score 7.0")); // median(8,6)
+        assert!(md.contains("trade-offs"));
+        assert!(md.contains("| 7 |"));
+    }
+
+    #[test]
+    fn swap_markdown_renders_reload_and_cold_costs() {
+        let costs = vec![SwapCost {
+            target_name: "beast".into(),
+            model_id: "Qwen/Qwen3.6-27B".into(),
+            git_sha: "abc1234".into(),
+            gpu: Some("2× RTX 5090".into()),
+            unload_ms_median: Some(320.0),
+            load_ms_median: Some(25000.0),
+            cold_ttft_s_median: Some(2.5),
+            cold_total_s_median: Some(5.0),
+            samples: 3,
+        }];
+        let md = render_swap_markdown(&costs);
+        assert!(md.contains("reload (s)"));
+        assert!(md.contains("beast"));
+        assert!(md.contains("25.00")); // 25000 ms → 25.00 s
+        assert!(md.contains("2.500"));
+    }
+
+    #[test]
+    fn scaling_markdown_renders_curve_and_flatness() {
+        let curves = vec![ScalingCurve {
+            target_name: "beast".into(),
+            model_id: "Qwen/Qwen3.6-27B".into(),
+            git_sha: "abc1234".into(),
+            gpu: Some("2× RTX 5090".into()),
+            points: vec![
+                ScalingPoint {
+                    prompt_size: 128,
+                    prompt_tokens: Some(130),
+                    prefill_tps: Some(900.0),
+                    decode_tps: Some(50.0),
+                    samples: 5,
+                },
+                ScalingPoint {
+                    prompt_size: 4096,
+                    prompt_tokens: Some(4100),
+                    prefill_tps: Some(2800.0),
+                    decode_tps: Some(48.0),
+                    samples: 5,
+                },
+            ],
+            decode_flatness: Some(0.96),
+        }];
+        let md = render_scaling_markdown(&curves);
+        assert!(md.contains("### beast · Qwen/Qwen3.6-27B"));
+        assert!(md.contains("ctx tok"));
+        assert!(md.contains("decode flatness: 0.96"));
+        assert!(md.contains("holds"));
+    }
 
     #[test]
     fn markdown_has_header_and_row() {
@@ -77,14 +377,36 @@ mod tests {
             ttft_s_median: Some(0.123),
             decode_tps_median: Some(45.6),
             total_s_median: Some(1.234),
+            ttft_s_p95: Some(0.222),
+            ttft_s_p99: Some(0.250),
+            total_s_p95: Some(1.5),
+            total_s_p99: Some(1.6),
+            prefill_ms_median: Some(120.0),
+            decode_ms_median: Some(1100.0),
+            prefill_tps_median: Some(1066.7),
+            vram_used_mb_median: Some(43008.0),
+            vram_total_mb: Some(65536),
+            gpu_util_pct_median: Some(89.0),
+            gpu_temp_c_median: Some(64.0),
+            concurrency: None,
+            ttft_p95_load_s: None,
+            queue_wait_ms_median: None,
+            rejected_median: None,
             samples: 5,
             gpu: Some("2× RTX 5090".into()),
         }];
         let md = render_markdown(&rows);
         assert!(md.contains("| engine |"));
+        assert!(md.contains("prefill tok/s"));
+        assert!(md.contains("VRAM (GB)"));
+        assert!(md.contains("conc"));
         assert!(md.contains("beast"));
         assert!(md.contains("`30d50d6`"));
         assert!(md.contains("0.123"));
+        // p95 column rendered.
+        assert!(md.contains("0.222"));
+        // VRAM used/total in GB (43008/65536 MiB → 42.0/64.0).
+        assert!(md.contains("42.0/64.0"));
     }
 
     #[test]
@@ -99,6 +421,21 @@ mod tests {
             ttft_s_median: Some(0.1),
             decode_tps_median: None,
             total_s_median: Some(0.5),
+            ttft_s_p95: Some(0.1),
+            ttft_s_p99: Some(0.1),
+            total_s_p95: Some(0.5),
+            total_s_p99: Some(0.5),
+            prefill_ms_median: None,
+            decode_ms_median: None,
+            prefill_tps_median: None,
+            vram_used_mb_median: None,
+            vram_total_mb: None,
+            gpu_util_pct_median: None,
+            gpu_temp_c_median: None,
+            concurrency: None,
+            ttft_p95_load_s: None,
+            queue_wait_ms_median: None,
+            rejected_median: None,
             samples: 1,
             gpu: None,
         }];

crates/helexa-bench/src/scenario.rs

@@ -62,6 +62,41 @@ pub struct ScenarioMetrics {
     pub prompt_tokens: Option<u64>,
     /// Completion tokens: from `usage` when present, else content-chunk count.
     pub completion_tokens: u64,
+    /// Server-measured prefill duration (ms), from the `usage.helexa_timing`
+    /// extension (#85). `None` when the server didn't emit it (external
+    /// engines, non-instrumented paths). The honest prefill-phase number,
+    /// distinct from client-observed `ttft_s` which also includes request
+    /// setup + first-byte network latency.
+    pub prefill_ms: Option<u64>,
+    /// Server-measured decode duration (ms), from `usage.helexa_timing`.
+    pub decode_ms: Option<u64>,
+    /// Tokens submitted to prefill — the denominator for prefill tok/s.
+    pub prefill_tokens: Option<u64>,
+    // ── Concurrency / agentic-load fields (#89) ──────────────────────────
+    // Set only by the concurrency scenario, which fans out N simultaneous
+    // streams to characterize the real a0/hermes/opencode workload that
+    // batch-1 single-request measurement can't see. `None` for single
+    // requests. For a concurrency burst, the inherited fields carry the
+    // aggregate: `ttft_s` = median TTFT across streams, `decode_tps` = node
+    // throughput (total tokens / burst window), `total_s` = burst wall-clock,
+    // `completion_tokens` = total across streams.
+    /// Number of simultaneous streams in the burst (the cell dimension).
+    pub concurrency: Option<u32>,
+    /// p95 of per-stream TTFT within the burst — the tail under simultaneous
+    /// load, where batch-1 serialization actually hurts.
+    pub ttft_p95_s: Option<f64>,
+    /// Median per-stream admission queue-wait (ms), approximated as
+    /// `ttft − prefill_ms` (#85): on a batch-1 server, later streams wait for
+    /// earlier ones, so TTFT inflates while server prefill stays constant —
+    /// the gap is the wait. `None` if streams didn't report `helexa_timing`.
+    pub queue_wait_ms_median: Option<f64>,
+    /// Streams shed by admission control (HTTP 429/503) during the burst —
+    /// honest backpressure, not silent failures.
+    pub rejected: Option<u32>,
+    /// Full generated text, captured only by the capability probe (#91) so
+    /// the output can be quality-scored later (manual or LLM-judge). `None`
+    /// for latency/throughput scenarios, which discard the text.
+    pub artifact: Option<String>,
 }
 
 #[async_trait]
@@ -85,10 +120,13 @@ pub trait Scenario: Send + Sync {
     async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics>;
 }
 
-/// Build the active scenario set from config. One chat-latency scenario
+/// Build the active scenario set from config: one chat-latency scenario per
-/// per configured prompt size.
+/// prompt size, plus one concurrency scenario per configured level (#89).
+/// Concurrency levels default to empty (opt-in), since a burst puts real
+/// simultaneous load on a serving fleet — operators enable it deliberately.
 pub fn build_scenarios(cfg: &ScenarioConfig) -> Vec<Box<dyn Scenario>> {
-    cfg.prompt_sizes
+    let mut scenarios: Vec<Box<dyn Scenario>> = cfg
+        .prompt_sizes
         .iter()
         .map(|&size| {
             Box::new(ChatLatencyScenario {
@@ -96,7 +134,46 @@ pub fn build_scenarios(cfg: &ScenarioConfig) -> Vec<Box<dyn Scenario>> {
                 approx_prompt_tokens: size,
             }) as Box<dyn Scenario>
         })
-        .collect()
+        .collect();
+    for &n in &cfg.concurrency_levels {
+        scenarios.push(Box::new(ConcurrencyScenario {
+            id: format!("concurrency:{n}"),
+            concurrency: n,
+            approx_prompt_tokens: cfg.concurrency_prompt_tokens,
+        }) as Box<dyn Scenario>);
+    }
+    for probe in &cfg.capability_probes {
+        scenarios.push(Box::new(CapabilityScenario {
+            id: format!("capability:{}", probe.name),
+            prompt: probe.prompt.clone(),
+            max_tokens: probe.max_tokens,
+        }) as Box<dyn Scenario>);
+    }
+    scenarios
+}
+
+/// A single small streamed request, timed like a chat-latency run. Used by
+/// the swap-cost measurement (#90) to capture the cold first-request latency
+/// straight after a reload. Reuses the shared SSE-timing core.
+pub async fn cold_probe(ctx: &RunCtx<'_>) -> Result<ScenarioMetrics> {
+    let prompt = build_prompt(128);
+    let payload = chat_payload(ctx, &prompt);
+    tokio::time::timeout(ctx.timeout, stream_and_measure(ctx, &payload))
+        .await
+        .map_err(|_| anyhow!("cold probe timed out after {:?}", ctx.timeout))?
+}
+
+/// The chat-completions request body shared by the latency and concurrency
+/// scenarios — streamed, deterministic (temperature 0), usage included.
+fn chat_payload(ctx: &RunCtx, prompt: &str) -> serde_json::Value {
+    json!({
+        "model": ctx.model_id,
+        "messages": [{"role": "user", "content": prompt}],
+        "max_tokens": ctx.max_tokens,
+        "temperature": 0,
+        "stream": true,
+        "stream_options": {"include_usage": true},
+    })
 }
 
 /// Streamed single-request chat-completions latency probe — the batch-1
@@ -118,15 +195,7 @@ impl Scenario for ChatLatencyScenario {
 
     async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics> {
         let prompt = build_prompt(self.approx_prompt_tokens);
-        let payload = json!({
+        let payload = chat_payload(ctx, &prompt);
-            "model": ctx.model_id,
-            "messages": [{"role": "user", "content": prompt}],
-            "max_tokens": ctx.max_tokens,
-            "temperature": 0,
-            "stream": true,
-            "stream_options": {"include_usage": true},
-        });
-
         let fut = stream_and_measure(ctx, &payload);
         tokio::time::timeout(ctx.timeout, fut)
             .await
@@ -134,11 +203,182 @@ impl Scenario for ChatLatencyScenario {
     }
 }
 
+/// Fan-out load probe: fire `concurrency` identical streams at once and
+/// measure how the fleet behaves under simultaneous pressure (#89). This is
+/// the only scenario that exercises the real a0/hermes/opencode pattern —
+/// many agentic requests per user turn — which batch-1 single-request
+/// timing cannot characterize. On a batch-1 serialized server, aggregate
+/// throughput stays ~flat while TTFT/queue-wait inflate with `concurrency`;
+/// that gap is the evidence for/against continuous batching.
+pub struct ConcurrencyScenario {
+    id: String,
+    concurrency: u32,
+    approx_prompt_tokens: u32,
+}
+
+#[async_trait]
+impl Scenario for ConcurrencyScenario {
+    fn id(&self) -> &str {
+        &self.id
+    }
+
+    fn prompt_size(&self) -> u32 {
+        self.approx_prompt_tokens
+    }
+
+    async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics> {
+        let prompt = build_prompt(self.approx_prompt_tokens);
+        let payload = chat_payload(ctx, &prompt);
+
+        // Fire all streams at once; each is independently timed and capped by
+        // the per-request timeout so one hung stream can't stall the burst.
+        let burst_start = Instant::now();
+        let futs = (0..self.concurrency).map(|_| async {
+            tokio::time::timeout(ctx.timeout, stream_and_measure(ctx, &payload)).await
+        });
+        let results = futures::future::join_all(futs).await;
+        let burst_window = burst_start.elapsed().as_secs_f64();
+
+        let mut streams: Vec<ScenarioMetrics> = Vec::new();
+        let mut rejected: u32 = 0;
+        for r in results {
+            match r {
+                Ok(Ok(m)) => streams.push(m),
+                // Admission backpressure (429/503) is shed load, counted
+                // separately from genuine failures/timeouts.
+                Ok(Err(e)) if is_admission_reject(&e) => rejected += 1,
+                Ok(Err(_)) | Err(_) => {}
+            }
+        }
+        if streams.is_empty() {
+            return Err(anyhow!(
+                "all {} concurrent streams failed ({rejected} shed by admission)",
+                self.concurrency
+            ));
+        }
+
+        let total_tokens: u64 = streams.iter().map(|m| m.completion_tokens).sum();
+        let ttfts: Vec<f64> = streams.iter().map(|m| m.ttft_s).collect();
+        // queue-wait ≈ TTFT − server prefill (#85); only for streams that
+        // reported helexa_timing.
+        let queue_waits: Vec<f64> = streams
+            .iter()
+            .filter_map(|m| {
+                m.prefill_ms
+                    .map(|p| (m.ttft_s * 1000.0 - p as f64).max(0.0))
+            })
+            .collect();
+        // Aggregate decode throughput across the whole node for the burst.
+        let aggregate_tps = if burst_window > 0.0 {
+            Some(total_tokens as f64 / burst_window)
+        } else {
+            None
+        };
+
+        Ok(ScenarioMetrics {
+            ttft_s: median(&ttfts).unwrap_or(0.0),
+            decode_tps: aggregate_tps,
+            total_s: burst_window,
+            prompt_tokens: streams.iter().find_map(|m| m.prompt_tokens),
+            completion_tokens: total_tokens,
+            prefill_ms: None,
+            decode_ms: None,
+            prefill_tokens: None,
+            concurrency: Some(self.concurrency),
+            ttft_p95_s: percentile(&ttfts, 95.0),
+            queue_wait_ms_median: median(&queue_waits),
+            rejected: Some(rejected),
+            artifact: None,
+        })
+    }
+}
+
+/// Quality probe (#91): runs a fixed prompt and stores the full generated
+/// text as an artifact for later scoring (manual now, LLM-judge later). The
+/// point is to compare reasoning/planning quality across models — the axis
+/// speed-only scenarios miss — so the frontier A/B (F3) picks on capability,
+/// not just throughput.
+pub struct CapabilityScenario {
+    id: String,
+    prompt: String,
+    max_tokens: u64,
+}
+
+#[async_trait]
+impl Scenario for CapabilityScenario {
+    fn id(&self) -> &str {
+        &self.id
+    }
+
+    /// Capability probes have no synthetic prompt-token target; the cell is
+    /// keyed by the scenario id alone.
+    fn prompt_size(&self) -> u32 {
+        0
+    }
+
+    async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics> {
+        let payload = json!({
+            "model": ctx.model_id,
+            "messages": [{"role": "user", "content": self.prompt}],
+            "max_tokens": self.max_tokens,
+            "temperature": 0,
+            "stream": true,
+            "stream_options": {"include_usage": true},
+        });
+        let fut = stream_and_measure_inner(ctx, &payload, true);
+        tokio::time::timeout(ctx.timeout, fut)
+            .await
+            .map_err(|_| anyhow!("capability probe timed out after {:?}", ctx.timeout))?
+    }
+}
+
+/// Whether a stream error was admission backpressure (HTTP 429/503) rather
+/// than a genuine failure. `stream_and_measure` renders the upstream status
+/// into the error string, so a substring check is sufficient.
+fn is_admission_reject(e: &anyhow::Error) -> bool {
+    let s = e.to_string();
+    s.contains("429") || s.contains("503")
+}
+
+/// Median of a slice (sorted copy). `None` if empty.
+fn median(values: &[f64]) -> Option<f64> {
+    if values.is_empty() {
+        return None;
+    }
+    let mut v = values.to_vec();
+    v.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    let lo = (v.len() - 1) / 2;
+    let hi = v.len() / 2;
+    Some((v[lo] + v[hi]) / 2.0)
+}
+
+/// Nearest-rank percentile of a slice (`p` in 0..=100). `None` if empty.
+fn percentile(values: &[f64], p: f64) -> Option<f64> {
+    if values.is_empty() {
+        return None;
+    }
+    let mut v = values.to_vec();
+    v.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    let rank = (p / 100.0 * v.len() as f64).ceil() as usize;
+    Some(v[rank.clamp(1, v.len()) - 1])
+}
+
 /// The SSE-timing core, ported from `bench.py::one_run`. Kept free of the
 /// `Scenario` trait so it's unit-testable against a mock byte stream.
 async fn stream_and_measure(
     ctx: &RunCtx<'_>,
     payload: &serde_json::Value,
+) -> Result<ScenarioMetrics> {
+    stream_and_measure_inner(ctx, payload, false).await
+}
+
+/// As [`stream_and_measure`] but accumulates the full visible text when
+/// `capture_text` is set — used by the capability probe (#91) to store the
+/// generated artifact for later quality scoring.
+async fn stream_and_measure_inner(
+    ctx: &RunCtx<'_>,
+    payload: &serde_json::Value,
+    capture_text: bool,
 ) -> Result<ScenarioMetrics> {
     let start = Instant::now();
     let resp = ctx
@@ -160,6 +400,10 @@ async fn stream_and_measure(
     let mut chunk_count: u64 = 0;
     let mut prompt_tokens: Option<u64> = None;
     let mut completion_tokens: Option<u64> = None;
+    let mut prefill_ms: Option<u64> = None;
+    let mut decode_ms: Option<u64> = None;
+    let mut prefill_tokens: Option<u64> = None;
+    let mut captured = String::new();
 
     while let Some(event) = stream.next().await {
         let event = event.context("reading SSE stream")?;
@@ -172,46 +416,86 @@ async fn stream_and_measure(
             Ok(c) => c,
             Err(_) => continue, // tolerate non-JSON keepalive frames
         };
-        if let Some(choice) = chunk.choices.first()
+        if let Some(choice) = chunk.choices.first() {
-            && choice
+            // Liveness counts ANY generated delta (#117). Thinking
+            // models (Qwen3-Next-Thinking, Qwen3 with thinking on)
+            // stream `reasoning_content` first — sometimes for their
+            // entire budget — and a content-only view misread that as
+            // a dead stream ("no content chunks received") while also
+            // producing impossible client-side rates (reasoning-
+            // inclusive token counts over a visible-content-only
+            // window; observed: "244 tok/s" on a 3060). For
+            // non-thinking models the first delta IS content, so
+            // `ttft_s` semantics are unchanged for them.
+            let content = choice
                 .delta
                 .get("content")
                 .and_then(|c| c.as_str())
-                .is_some_and(|s| !s.is_empty())
+                .filter(|c| !c.is_empty());
-        {
+            let reasoning = choice
-            if first.is_none() {
+                .delta
-                first = Some(now);
+                .get("reasoning_content")
+                .and_then(|c| c.as_str())
+                .filter(|c| !c.is_empty());
+            if content.is_some() || reasoning.is_some() {
+                if first.is_none() {
+                    first = Some(now);
+                }
+                last = Some(now);
+                chunk_count += 1;
+            }
+            if capture_text && let Some(text) = content {
+                captured.push_str(text);
             }
-            last = Some(now);
-            chunk_count += 1;
         }
         if let Some(usage) = chunk.usage {
             prompt_tokens = Some(usage.prompt_tokens);
             completion_tokens = Some(usage.completion_tokens);
+            if let Some(t) = usage.helexa_timing {
+                prefill_ms = Some(t.prefill_ms);
+                decode_ms = Some(t.decode_ms);
+                prefill_tokens = Some(t.prefill_tokens);
+            }
         }
     }
     let end = Instant::now();
 
-    let first = first.ok_or_else(|| anyhow!("no content chunks received"))?;
+    let first = first.ok_or_else(|| anyhow!("no generated chunks received"))?;
 
-    // neuron emits one SSE chunk per visible token, so chunk_count is an
+    // neuron emits one SSE chunk per generated token, so chunk_count is
-    // engine-truth count when no usage frame is sent.
+    // an engine-truth count when no usage frame is sent.
     let tokens = completion_tokens.filter(|&t| t > 0).unwrap_or(chunk_count);
-    // decode rate is only meaningful over a real inter-chunk window.
+    // Decode rate: prefer the server-measured split (#85) — it counts
+    // every generated token over the actual decode window, immune to
+    // reasoning-suppression frame mismatches. Fall back to the client
+    // inter-chunk window with the CHUNK count (same frame) — never
+    // usage.completion_tokens over the chunk window, which mixes a
+    // reasoning-inclusive numerator with a visible-only denominator.
     let window = last
         .filter(|&l| l > first)
         .map(|l| (l - first).as_secs_f64())
         .unwrap_or(0.0);
+    let decode_tps = match decode_ms {
+        Some(ms) if ms > 200 && tokens > 0 => Some(tokens as f64 / (ms as f64 / 1000.0)),
+        _ if window > 0.2 => Some(chunk_count as f64 / window),
+        _ => None,
+    };
     Ok(ScenarioMetrics {
         ttft_s: (first - start).as_secs_f64(),
-        decode_tps: if window > 0.2 {
+        decode_tps,
-            Some(tokens as f64 / window)
-        } else {
-            None
-        },
         total_s: (end - start).as_secs_f64(),
         prompt_tokens,
         completion_tokens: tokens,
+        prefill_ms,
+        decode_ms,
+        prefill_tokens,
+        // Concurrency fields unset on the single-request path; the
+        // concurrency scenario builds its own aggregate (#89).
+        concurrency: None,
+        ttft_p95_s: None,
+        queue_wait_ms_median: None,
+        rejected: None,
+        artifact: if capture_text { Some(captured) } else { None },
     })
 }
 
@@ -229,6 +513,54 @@ mod tests {
         assert!(small.ends_with("/no_think"));
     }
 
+    #[test]
+    fn median_and_percentile_basics() {
+        assert_eq!(median(&[3.0, 1.0, 2.0]), Some(2.0));
+        assert_eq!(median(&[]), None);
+        let v = [1.0, 2.0, 3.0, 4.0, 5.0];
+        assert_eq!(percentile(&v, 50.0), Some(3.0));
+        assert_eq!(percentile(&v, 95.0), Some(5.0)); // nearest-rank → max with n=5
+        assert_eq!(percentile(&[], 95.0), None);
+    }
+
+    #[test]
+    fn admission_rejects_detected_by_status() {
+        assert!(is_admission_reject(&anyhow!(
+            "upstream returned 429 Too Many Requests"
+        )));
+        assert!(is_admission_reject(&anyhow!(
+            "upstream returned 503 Service Unavailable"
+        )));
+        assert!(!is_admission_reject(&anyhow!(
+            "upstream returned 500 Internal"
+        )));
+        assert!(!is_admission_reject(&anyhow!("connection refused")));
+    }
+
+    #[test]
+    fn concurrency_scenarios_built_from_config() {
+        use crate::config::{CapabilityProbe, ScenarioConfig};
+        let cfg = ScenarioConfig {
+            prompt_sizes: vec![128],
+            max_tokens: 64,
+            concurrency_levels: vec![2, 8],
+            concurrency_prompt_tokens: 512,
+            capability_probes: vec![CapabilityProbe {
+                name: "plan".into(),
+                prompt: "Write a plan.".into(),
+                max_tokens: 2048,
+            }],
+        };
+        let ids: Vec<String> = build_scenarios(&cfg)
+            .iter()
+            .map(|s| s.id().to_string())
+            .collect();
+        assert!(ids.contains(&"chat:128".to_string()));
+        assert!(ids.contains(&"concurrency:2".to_string()));
+        assert!(ids.contains(&"concurrency:8".to_string()));
+        assert!(ids.contains(&"capability:plan".to_string()));
+    }
+
     #[test]
     fn prompt_floor_for_tiny_targets() {
         // max(approx,16) floor means even 0 yields a non-trivial prompt.

crates/helexa-bench/src/store.rs

@@ -51,6 +51,35 @@ pub struct RunRecord {
     pub decode_tps: Option<f64>,
     pub total_s: Option<f64>,
     pub completion_tokens: Option<u64>,
+    // server-measured prefill/decode split (#85), null on engines/paths
+    // that don't emit `usage.helexa_timing`.
+    pub prefill_ms: Option<u64>,
+    pub decode_ms: Option<u64>,
+    pub prefill_tokens: Option<u64>,
+    // GPU telemetry sampled from /health around the run (#87), null for
+    // non-neuron targets or when /health was unreachable. vram_used_mb is
+    // the node sum; util/temp are the hottest single device.
+    pub vram_used_mb: Option<u64>,
+    pub gpu_util_pct: Option<u32>,
+    pub gpu_temp_c: Option<u32>,
+    // concurrency / agentic-load burst metrics (#89), null for single-request
+    // scenarios. For a burst, ttft_s/decode_tps/total_s carry the aggregate.
+    pub concurrency: Option<u32>,
+    pub ttft_p95_s: Option<f64>,
+    pub queue_wait_ms: Option<f64>,
+    pub rejected: Option<u32>,
+    // cold-load / model-swap cost (#90), set only by the deliberate
+    // `swap-cost` measurement (scenario_id = "swap"). The other metric
+    // fields carry the cold first-request after reload.
+    pub swap_unload_ms: Option<u64>,
+    pub swap_load_ms: Option<u64>,
+    // capability probe (#91): the full generated text (scenario_id =
+    // "capability:<name>"), scored for quality later. `quality_score` /
+    // `scorer` are null at insert — set by `score` (manual) or a future
+    // LLM-judge.
+    pub artifact: Option<String>,
+    pub quality_score: Option<f64>,
+    pub scorer: Option<String>,
     // outcome
     pub ok: bool,
     pub error: Option<String>,
@@ -123,6 +152,21 @@ impl Store {
                 decode_tps           REAL,
                 total_s              REAL,
                 completion_tokens    INTEGER,
+                prefill_ms           INTEGER,
+                decode_ms            INTEGER,
+                prefill_tokens       INTEGER,
+                vram_used_mb         INTEGER,
+                gpu_util_pct         INTEGER,
+                gpu_temp_c           INTEGER,
+                concurrency          INTEGER,
+                ttft_p95_s           REAL,
+                queue_wait_ms        REAL,
+                rejected             INTEGER,
+                swap_unload_ms       INTEGER,
+                swap_load_ms         INTEGER,
+                artifact             TEXT,
+                quality_score        REAL,
+                scorer               TEXT,
                 ok                   INTEGER NOT NULL,
                 error                TEXT
             );
@@ -133,6 +177,51 @@ impl Store {
             "#,
         )
         .context("initialising sqlite schema")?;
+        // Additive migrations for DBs created before a column existed.
+        // `CREATE TABLE IF NOT EXISTS` above only seeds fresh DBs; existing
+        // ones need the columns backfilled (as NULL) so older rows coexist
+        // with new metrics. There is no migration framework — each entry is
+        // an idempotent "add if missing".
+        Self::ensure_columns(
+            conn,
+            "runs",
+            &[
+                ("prefill_ms", "INTEGER"),
+                ("decode_ms", "INTEGER"),
+                ("prefill_tokens", "INTEGER"),
+                ("vram_used_mb", "INTEGER"),
+                ("gpu_util_pct", "INTEGER"),
+                ("gpu_temp_c", "INTEGER"),
+                ("concurrency", "INTEGER"),
+                ("ttft_p95_s", "REAL"),
+                ("queue_wait_ms", "REAL"),
+                ("rejected", "INTEGER"),
+                ("swap_unload_ms", "INTEGER"),
+                ("swap_load_ms", "INTEGER"),
+                ("artifact", "TEXT"),
+                ("quality_score", "REAL"),
+                ("scorer", "TEXT"),
+            ],
+        )?;
+        Ok(())
+    }
+
+    /// Add any of `columns` that the table is missing (`ALTER TABLE ADD
+    /// COLUMN`). Idempotent: existing columns are read from
+    /// `PRAGMA table_info` and skipped, so this is safe to run on every open.
+    fn ensure_columns(conn: &Connection, table: &str, columns: &[(&str, &str)]) -> Result<()> {
+        let mut existing = std::collections::HashSet::new();
+        let mut stmt = conn.prepare(&format!("PRAGMA table_info({table})"))?;
+        let names = stmt.query_map([], |row| row.get::<_, String>(1))?;
+        for name in names {
+            existing.insert(name?);
+        }
+        for (name, ty) in columns {
+            if !existing.contains(*name) {
+                conn.execute_batch(&format!("ALTER TABLE {table} ADD COLUMN {name} {ty};"))
+                    .with_context(|| format!("adding column {table}.{name}"))?;
+            }
+        }
         Ok(())
     }
 
@@ -166,6 +255,11 @@ impl Store {
                 model_id, harness, capabilities_json, devices_json,
                 scenario_id, prompt_size_approx, prompt_tokens_actual, max_tokens,
                 ttft_s, decode_tps, total_s, completion_tokens,
+                prefill_ms, decode_ms, prefill_tokens,
+                vram_used_mb, gpu_util_pct, gpu_temp_c,
+                concurrency, ttft_p95_s, queue_wait_ms, rejected,
+                swap_unload_ms, swap_load_ms,
+                artifact, quality_score, scorer,
                 ok, error
             ) VALUES (
                 ?1, ?2, ?3, ?4,
@@ -176,7 +270,12 @@ impl Store {
                 ?20, ?21, ?22, ?23,
                 ?24, ?25, ?26, ?27,
                 ?28, ?29, ?30, ?31,
-                ?32, ?33
+                ?32, ?33, ?34,
+                ?35, ?36, ?37,
+                ?38, ?39, ?40, ?41,
+                ?42, ?43,
+                ?44, ?45, ?46,
+                ?47, ?48
             )",
             params![
                 r.ts,
@@ -210,6 +309,21 @@ impl Store {
                 r.decode_tps,
                 r.total_s,
                 r.completion_tokens,
+                r.prefill_ms,
+                r.decode_ms,
+                r.prefill_tokens,
+                r.vram_used_mb,
+                r.gpu_util_pct,
+                r.gpu_temp_c,
+                r.concurrency,
+                r.ttft_p95_s,
+                r.queue_wait_ms,
+                r.rejected,
+                r.swap_unload_ms,
+                r.swap_load_ms,
+                r.artifact,
+                r.quality_score,
+                r.scorer,
                 r.ok as i64,
                 r.error,
             ],
@@ -224,7 +338,10 @@ impl Store {
         // successful run, then median that SHA's samples.
         let mut stmt = self.conn.prepare(
             "SELECT target_name, model_id, scenario_id, prompt_size_approx, git_sha,
-                    ttft_s, decode_tps, total_s, prompt_tokens_actual, gpus_json
+                    ttft_s, decode_tps, total_s, prompt_tokens_actual, gpus_json,
+                    prefill_ms, decode_ms, prefill_tokens,
+                    vram_used_mb, gpu_util_pct, gpu_temp_c,
+                    concurrency, ttft_p95_s, queue_wait_ms, rejected
              FROM runs
              WHERE ok=1
              ORDER BY target_name, model_id, scenario_id, id",
@@ -241,12 +358,188 @@ impl Store {
                 total_s: row.get(7)?,
                 prompt_tokens_actual: row.get(8)?,
                 gpus_json: row.get(9)?,
+                prefill_ms: row.get(10)?,
+                decode_ms: row.get(11)?,
+                prefill_tokens: row.get(12)?,
+                vram_used_mb: row.get(13)?,
+                gpu_util_pct: row.get(14)?,
+                gpu_temp_c: row.get(15)?,
+                concurrency: row.get(16)?,
+                ttft_p95_s: row.get(17)?,
+                queue_wait_ms: row.get(18)?,
+                rejected: row.get(19)?,
             })
         })?;
         let raws: Vec<RawRow> = rows.collect::<rusqlite::Result<_>>()?;
         Ok(aggregate(raws))
     }
 
+    /// Context-length scaling curves (#88): per (target, model), the latest
+    /// build's `chat:<n>` cells pivoted by prompt size into prefill & decode
+    /// tok/s vs context. The headline is `decode_flatness` — decode tok/s at
+    /// the largest context divided by the smallest. Near 1.0 confirms the
+    /// Gated-DeltaNet O(1)-in-sequence-length decode; a sharp drop locates
+    /// where the model stops scaling for free.
+    pub fn scaling(&self) -> Result<Vec<ScalingCurve>> {
+        use std::collections::BTreeMap;
+        // Reuse the already-aggregated report cells; the chat:<n> rows are
+        // exactly the per-context measurement points.
+        let mut by_model: BTreeMap<(String, String), Vec<ReportRow>> = BTreeMap::new();
+        for r in self.report_rows()? {
+            if r.scenario_id.starts_with("chat:") {
+                by_model
+                    .entry((r.target_name.clone(), r.model_id.clone()))
+                    .or_default()
+                    .push(r);
+            }
+        }
+        let mut out = Vec::new();
+        for ((target_name, model_id), mut rows) in by_model {
+            rows.sort_by_key(|r| r.prompt_size_approx);
+            let points: Vec<ScalingPoint> = rows
+                .iter()
+                .map(|r| ScalingPoint {
+                    prompt_size: r.prompt_size_approx,
+                    prompt_tokens: r.prompt_tokens,
+                    prefill_tps: r.prefill_tps_median,
+                    decode_tps: r.decode_tps_median,
+                    samples: r.samples,
+                })
+                .collect();
+            // Flatness across the smallest→largest points that both have a
+            // decode rate (skips cells where the decode window was too short).
+            let with_decode: Vec<&ScalingPoint> =
+                points.iter().filter(|p| p.decode_tps.is_some()).collect();
+            let decode_flatness = match (with_decode.first(), with_decode.last()) {
+                (Some(lo), Some(hi)) if with_decode.len() >= 2 => {
+                    match (lo.decode_tps, hi.decode_tps) {
+                        (Some(a), Some(b)) if a > 0.0 => Some(b / a),
+                        _ => None,
+                    }
+                }
+                _ => None,
+            };
+            out.push(ScalingCurve {
+                target_name,
+                model_id,
+                git_sha: rows.first().map(|r| r.git_sha.clone()).unwrap_or_default(),
+                gpu: rows.iter().find_map(|r| r.gpu.clone()),
+                points,
+                decode_flatness,
+            });
+        }
+        Ok(out)
+    }
+
+    /// Cold-load / model-swap costs (#90): per (target, model) at the latest
+    /// build, the median unload→reload latency and the cold first-request
+    /// latency after reload (the `scenario_id = "swap"` rows).
+    pub fn swap_costs(&self) -> Result<Vec<SwapCost>> {
+        use std::collections::BTreeMap;
+        let mut stmt = self.conn.prepare(
+            "SELECT target_name, model_id, git_sha, gpus_json,
+                    swap_unload_ms, swap_load_ms, ttft_s, total_s
+             FROM runs
+             WHERE ok=1 AND scenario_id='swap'
+             ORDER BY target_name, model_id, id",
+        )?;
+        struct Raw {
+            target: String,
+            model: String,
+            sha: String,
+            gpus_json: Option<String>,
+            unload_ms: Option<f64>,
+            load_ms: Option<f64>,
+            ttft_s: Option<f64>,
+            total_s: Option<f64>,
+        }
+        let raws: Vec<Raw> = stmt
+            .query_map([], |r| {
+                Ok(Raw {
+                    target: r.get(0)?,
+                    model: r.get(1)?,
+                    sha: r.get(2)?,
+                    gpus_json: r.get(3)?,
+                    unload_ms: r.get::<_, Option<i64>>(4)?.map(|v| v as f64),
+                    load_ms: r.get::<_, Option<i64>>(5)?.map(|v| v as f64),
+                    ttft_s: r.get(6)?,
+                    total_s: r.get(7)?,
+                })
+            })?
+            .collect::<rusqlite::Result<_>>()?;
+        let mut by: BTreeMap<(String, String), Vec<Raw>> = BTreeMap::new();
+        for r in raws {
+            by.entry((r.target.clone(), r.model.clone()))
+                .or_default()
+                .push(r);
+        }
+        let mut out = Vec::new();
+        for ((target, model), rows) in by {
+            let latest = rows.last().map(|r| r.sha.clone()).unwrap_or_default();
+            let cell: Vec<&Raw> = rows.iter().filter(|r| r.sha == latest).collect();
+            out.push(SwapCost {
+                target_name: target,
+                model_id: model,
+                git_sha: latest,
+                gpu: cell
+                    .iter()
+                    .find_map(|r| r.gpus_json.as_deref().and_then(gpu_label)),
+                unload_ms_median: median(cell.iter().filter_map(|r| r.unload_ms)),
+                load_ms_median: median(cell.iter().filter_map(|r| r.load_ms)),
+                cold_ttft_s_median: median(cell.iter().filter_map(|r| r.ttft_s)),
+                cold_total_s_median: median(cell.iter().filter_map(|r| r.total_s)),
+                samples: cell.len(),
+            });
+        }
+        Ok(out)
+    }
+
+    /// Capability-probe runs (#91): the stored output artifacts and their
+    /// quality scores. `unscored_only` filters to runs awaiting a score —
+    /// the worklist for manual scoring or a future LLM-judge.
+    pub fn capability_runs(&self, unscored_only: bool) -> Result<Vec<CapabilityRun>> {
+        let sql = format!(
+            "SELECT id, ts, target_name, model_id, scenario_id, git_sha,
+                    quality_score, scorer, artifact
+             FROM runs
+             WHERE ok=1 AND scenario_id LIKE 'capability:%'{}
+             ORDER BY id DESC",
+            if unscored_only {
+                " AND quality_score IS NULL"
+            } else {
+                ""
+            },
+        );
+        let mut stmt = self.conn.prepare(&sql)?;
+        let rows = stmt
+            .query_map([], |r| {
+                Ok(CapabilityRun {
+                    id: r.get(0)?,
+                    ts: r.get(1)?,
+                    target_name: r.get(2)?,
+                    model_id: r.get(3)?,
+                    scenario_id: r.get(4)?,
+                    git_sha: r.get(5)?,
+                    quality_score: r.get(6)?,
+                    scorer: r.get(7)?,
+                    artifact: r.get(8)?,
+                })
+            })?
+            .collect::<rusqlite::Result<_>>()?;
+        Ok(rows)
+    }
+
+    /// Attach a quality score to one capability run (#91). `scorer` records
+    /// who/what scored it ("manual", "llm:claude-…"). Returns the row count
+    /// updated (0 if the id doesn't exist).
+    pub fn set_score(&self, run_id: i64, score: f64, scorer: &str) -> Result<usize> {
+        let n = self.conn.execute(
+            "UPDATE runs SET quality_score=?1, scorer=?2 WHERE id=?3",
+            params![score, scorer, run_id],
+        )?;
+        Ok(n)
+    }
+
     // ── Read API surface (consumed by api.rs) ─────────────────────────
 
     /// Total recorded runs (for `/api/health`).
@@ -379,7 +672,10 @@ impl Store {
             "SELECT id, ts, target_name, hostname, git_sha, build_timestamp, package_version,
                     model_id, harness, scenario_id, prompt_size_approx, prompt_tokens_actual,
                     max_tokens, ttft_s, decode_tps, total_s, completion_tokens, ok, error,
-                    gpus_json
+                    gpus_json, prefill_ms, decode_ms, prefill_tokens,
+                    vram_used_mb, gpu_util_pct, gpu_temp_c,
+                    concurrency, ttft_p95_s, queue_wait_ms, rejected,
+                    swap_unload_ms, swap_load_ms
              FROM runs",
         );
         let mut conds: Vec<String> = Vec::new();
@@ -435,6 +731,18 @@ impl Store {
                     completion_tokens: r.get(16)?,
                     ok: r.get::<_, i64>(17)? != 0,
                     error: r.get(18)?,
+                    prefill_ms: r.get(20)?,
+                    decode_ms: r.get(21)?,
+                    prefill_tokens: r.get(22)?,
+                    vram_used_mb: r.get(23)?,
+                    gpu_util_pct: r.get(24)?,
+                    gpu_temp_c: r.get(25)?,
+                    concurrency: r.get(26)?,
+                    ttft_p95_s: r.get(27)?,
+                    queue_wait_ms: r.get(28)?,
+                    rejected: r.get(29)?,
+                    swap_unload_ms: r.get(30)?,
+                    swap_load_ms: r.get(31)?,
                 })
             })?
             .collect::<rusqlite::Result<_>>()?;
@@ -554,6 +862,18 @@ pub struct RunRow {
     pub decode_tps: Option<f64>,
     pub total_s: Option<f64>,
     pub completion_tokens: Option<u64>,
+    pub prefill_ms: Option<u64>,
+    pub decode_ms: Option<u64>,
+    pub prefill_tokens: Option<u64>,
+    pub vram_used_mb: Option<u64>,
+    pub gpu_util_pct: Option<u64>,
+    pub gpu_temp_c: Option<u64>,
+    pub concurrency: Option<u64>,
+    pub ttft_p95_s: Option<f64>,
+    pub queue_wait_ms: Option<f64>,
+    pub rejected: Option<u64>,
+    pub swap_unload_ms: Option<u64>,
+    pub swap_load_ms: Option<u64>,
     pub ok: bool,
     pub error: Option<String>,
 }
@@ -569,6 +889,16 @@ struct RawRow {
     total_s: Option<f64>,
     prompt_tokens_actual: Option<u64>,
     gpus_json: Option<String>,
+    prefill_ms: Option<u64>,
+    decode_ms: Option<u64>,
+    prefill_tokens: Option<u64>,
+    vram_used_mb: Option<u64>,
+    gpu_util_pct: Option<u64>,
+    gpu_temp_c: Option<u64>,
+    concurrency: Option<u64>,
+    ttft_p95_s: Option<f64>,
+    queue_wait_ms: Option<f64>,
+    rejected: Option<u64>,
 }
 
 /// An aggregated cell ready for the report table.
@@ -583,11 +913,100 @@ pub struct ReportRow {
     pub ttft_s_median: Option<f64>,
     pub decode_tps_median: Option<f64>,
     pub total_s_median: Option<f64>,
+    /// Latency tail percentiles — where batch-1 pain actually shows up, and
+    /// invisible behind a bare median. p95/p99 nearest-rank; with few
+    /// samples they collapse toward the max (honest, not interpolated).
+    pub ttft_s_p95: Option<f64>,
+    pub ttft_s_p99: Option<f64>,
+    pub total_s_p95: Option<f64>,
+    pub total_s_p99: Option<f64>,
+    /// Server-measured prefill/decode split (#85). `prefill_tps_median` is
+    /// the true prompt-encoding rate (prefill_tokens / prefill_ms),
+    /// complementing `decode_tps_median` (the generation rate).
+    pub prefill_ms_median: Option<f64>,
+    pub decode_ms_median: Option<f64>,
+    pub prefill_tps_median: Option<f64>,
+    /// GPU telemetry sampled from /health around the run (#87). `vram_used_mb`
+    /// is the node sum; `vram_total_mb` (from discovery) lets the report show
+    /// real headroom — the "2/3 used" hunch as a number. util/temp are the
+    /// hottest device. All `None` for non-neuron targets.
+    pub vram_used_mb_median: Option<f64>,
+    pub vram_total_mb: Option<u64>,
+    pub gpu_util_pct_median: Option<f64>,
+    pub gpu_temp_c_median: Option<f64>,
+    /// Concurrency / agentic-load burst metrics (#89). `concurrency` is the
+    /// burst width (constant per cell). `ttft_p95_load_s` is the within-burst
+    /// TTFT tail; `queue_wait_ms_median` the admission wait; `rejected_median`
+    /// the per-burst shed count. All `None` for non-concurrency scenarios.
+    pub concurrency: Option<u64>,
+    pub ttft_p95_load_s: Option<f64>,
+    pub queue_wait_ms_median: Option<f64>,
+    pub rejected_median: Option<f64>,
     pub samples: usize,
     /// Public-facing resource name (the host's GPU(s)), e.g. "2× RTX 5090".
     pub gpu: Option<String>,
 }
 
+/// One context-length scaling curve for a (target, model) at its latest
+/// build — the points ordered by prompt size, plus the decode-flatness
+/// summary (#88).
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct ScalingCurve {
+    pub target_name: String,
+    pub model_id: String,
+    pub git_sha: String,
+    pub gpu: Option<String>,
+    pub points: Vec<ScalingPoint>,
+    /// decode tok/s at the largest context ÷ at the smallest. ~1.0 = flat
+    /// (GDN O(1) decode); <1 quantifies the drop-off. `None` with <2 points.
+    pub decode_flatness: Option<f64>,
+}
+
+/// Cold-load / model-swap cost for a (target, model) at its latest build
+/// (#90): the reload latency and the cold first-request after it.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct SwapCost {
+    pub target_name: String,
+    pub model_id: String,
+    pub git_sha: String,
+    pub gpu: Option<String>,
+    /// Time to free the model (`POST /models/unload`), ms.
+    pub unload_ms_median: Option<f64>,
+    /// Time to reload it (`POST /models/load`, synchronous), ms — the
+    /// headline swap cost feeding the vision cold-swap policy (F4e).
+    pub load_ms_median: Option<f64>,
+    /// TTFT of the first request after reload (cold caches), seconds.
+    pub cold_ttft_s_median: Option<f64>,
+    /// Total wall-clock of that cold first request, seconds.
+    pub cold_total_s_median: Option<f64>,
+    pub samples: usize,
+}
+
+/// One capability-probe run (#91): the stored output artifact plus its
+/// quality score (null until scored manually or by a future LLM-judge).
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct CapabilityRun {
+    pub id: i64,
+    pub ts: String,
+    pub target_name: String,
+    pub model_id: String,
+    pub scenario_id: String,
+    pub git_sha: String,
+    pub quality_score: Option<f64>,
+    pub scorer: Option<String>,
+    pub artifact: Option<String>,
+}
+
+/// One point on a [`ScalingCurve`]: the throughput at a given context size.
+#[derive(Debug, Clone, serde::Serialize)]
+pub struct ScalingPoint {
+    pub prompt_size: u32,
+    pub prompt_tokens: Option<u64>,
+    pub prefill_tps: Option<f64>,
+    pub decode_tps: Option<f64>,
+    pub samples: usize,
+}
+
 /// Group by (target, model, scenario), keep only the latest SHA's rows
 /// (latest = the SHA of the last-inserted row, since input is id-ordered),
 /// and median each metric.
@@ -611,6 +1030,11 @@ fn aggregate(raws: Vec<RawRow>) -> Vec<ReportRow> {
         let latest_sha = rows.last().map(|r| r.git_sha.clone()).unwrap_or_default();
         let cell: Vec<&RawRow> = rows.iter().filter(|r| r.git_sha == latest_sha).collect();
         let prompt_size_approx = cell.first().map(|r| r.prompt_size_approx).unwrap_or(0);
+        // Per-row prefill tok/s, derived from the server-measured split.
+        let prefill_tps = |r: &&RawRow| match (r.prefill_tokens, r.prefill_ms) {
+            (Some(tok), Some(ms)) if ms > 0 => Some(tok as f64 * 1000.0 / ms as f64),
+            _ => None,
+        };
         out.push(ReportRow {
             target_name,
             model_id,
@@ -621,6 +1045,27 @@ fn aggregate(raws: Vec<RawRow>) -> Vec<ReportRow> {
             ttft_s_median: median(cell.iter().filter_map(|r| r.ttft_s)),
             decode_tps_median: median(cell.iter().filter_map(|r| r.decode_tps)),
             total_s_median: median(cell.iter().filter_map(|r| r.total_s)),
+            ttft_s_p95: percentile(cell.iter().filter_map(|r| r.ttft_s), 95.0),
+            ttft_s_p99: percentile(cell.iter().filter_map(|r| r.ttft_s), 99.0),
+            total_s_p95: percentile(cell.iter().filter_map(|r| r.total_s), 95.0),
+            total_s_p99: percentile(cell.iter().filter_map(|r| r.total_s), 99.0),
+            prefill_ms_median: median(cell.iter().filter_map(|r| r.prefill_ms.map(|m| m as f64))),
+            decode_ms_median: median(cell.iter().filter_map(|r| r.decode_ms.map(|m| m as f64))),
+            prefill_tps_median: median(cell.iter().filter_map(prefill_tps)),
+            vram_used_mb_median: median(
+                cell.iter().filter_map(|r| r.vram_used_mb.map(|v| v as f64)),
+            ),
+            vram_total_mb: cell
+                .iter()
+                .find_map(|r| r.gpus_json.as_deref().and_then(gpu_total_vram_mb)),
+            gpu_util_pct_median: median(
+                cell.iter().filter_map(|r| r.gpu_util_pct.map(|v| v as f64)),
+            ),
+            gpu_temp_c_median: median(cell.iter().filter_map(|r| r.gpu_temp_c.map(|v| v as f64))),
+            concurrency: cell.iter().find_map(|r| r.concurrency),
+            ttft_p95_load_s: median(cell.iter().filter_map(|r| r.ttft_p95_s)),
+            queue_wait_ms_median: median(cell.iter().filter_map(|r| r.queue_wait_ms)),
+            rejected_median: median(cell.iter().filter_map(|r| r.rejected.map(|v| v as f64))),
             samples: cell.len(),
             gpu: cell
                 .iter()
@@ -630,6 +1075,19 @@ fn aggregate(raws: Vec<RawRow>) -> Vec<ReportRow> {
     out
 }
 
+/// Node total VRAM in MB, summed across the devices in a run's stored
+/// `gpus_json` (the discovery `DeviceInfo` list, each with `vram_total_mb`).
+/// Pairs with the sampled `vram_used_mb` to report real headroom (#87).
+/// `None` when empty/absent or no device declares a total.
+fn gpu_total_vram_mb(gpus_json: &str) -> Option<u64> {
+    let devices: Vec<serde_json::Value> = serde_json::from_str(gpus_json).ok()?;
+    let total: u64 = devices
+        .iter()
+        .filter_map(|d| d.get("vram_total_mb").and_then(|v| v.as_u64()))
+        .sum();
+    (total > 0).then_some(total)
+}
+
 /// Compact GPU label from a run's stored `gpus_json` (the discovery device
 /// list) — e.g. "2× RTX 5090", "RTX 4090". `None` when empty/absent. Used
 /// as the public-facing resource name in place of internal hostnames.
@@ -680,6 +1138,22 @@ fn median(values: impl Iterator<Item = f64>) -> Option<f64> {
     Some((v[lo] + v[hi]) / 2.0)
 }
 
+/// Nearest-rank percentile (`p` in 0..=100). Chosen over interpolation
+/// because bench cells hold only a handful of samples: with n=5, p95/p99
+/// resolve to the max, which honestly says "this is the worst we saw"
+/// rather than inventing a value between samples we never observed.
+fn percentile(values: impl Iterator<Item = f64>, p: f64) -> Option<f64> {
+    let mut v: Vec<f64> = values.collect();
+    if v.is_empty() {
+        return None;
+    }
+    v.sort_by(|a, b| a.partial_cmp(b).unwrap_or(std::cmp::Ordering::Equal));
+    // rank = ceil(p/100 * n), clamped to [1, n]; index is rank-1.
+    let rank = (p / 100.0 * v.len() as f64).ceil() as usize;
+    let idx = rank.clamp(1, v.len()) - 1;
+    Some(v[idx])
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -717,6 +1191,21 @@ mod tests {
             decode_tps: Some(50.0),
             total_s: Some(1.0),
             completion_tokens: Some(50),
+            prefill_ms: Some(200),
+            decode_ms: Some(1000),
+            prefill_tokens: Some(130),
+            vram_used_mb: Some(42000),
+            gpu_util_pct: Some(88),
+            gpu_temp_c: Some(64),
+            concurrency: None,
+            ttft_p95_s: None,
+            queue_wait_ms: None,
+            rejected: None,
+            swap_unload_ms: None,
+            swap_load_ms: None,
+            artifact: None,
+            quality_score: None,
+            scorer: None,
             ok,
             error: if ok { None } else { Some("boom".into()) },
         }
@@ -755,6 +1244,197 @@ mod tests {
         assert!((rows[0].ttft_s_median.unwrap() - 0.3).abs() < 1e-9);
     }
 
+    #[test]
+    fn report_surfaces_percentiles_and_prefill_split() {
+        let s = Store::open_in_memory().unwrap();
+        // Five samples on one cell with spread TTFT so percentiles differ
+        // from the median, plus a server-measured prefill/decode split.
+        for (i, ttft) in [0.10, 0.12, 0.14, 0.16, 0.50].iter().enumerate() {
+            let mut r = rec("beast", "sha", "m", "chat:128", true);
+            r.ttft_s = Some(*ttft);
+            r.total_s = Some(ttft + 1.0);
+            r.prefill_ms = Some(200 + i as u64);
+            r.prefill_tokens = Some(400);
+            s.insert_run(&r).unwrap();
+        }
+        let rows = s.report_rows().unwrap();
+        assert_eq!(rows.len(), 1);
+        let row = &rows[0];
+        assert_eq!(row.samples, 5);
+        // p50 is the middle value; p95/p99 (nearest-rank, n=5) hit the max.
+        assert!((row.ttft_s_median.unwrap() - 0.14).abs() < 1e-9);
+        assert!((row.ttft_s_p95.unwrap() - 0.50).abs() < 1e-9);
+        assert!((row.ttft_s_p99.unwrap() - 0.50).abs() < 1e-9);
+        // prefill tok/s = 400 tok / ~0.2 s ≈ 2000 tok/s.
+        assert!(row.prefill_tps_median.unwrap() > 1900.0);
+        assert!(row.prefill_ms_median.is_some());
+    }
+
+    #[test]
+    fn report_surfaces_vram_and_gpu_telemetry() {
+        let s = Store::open_in_memory().unwrap();
+        let mut r = rec("beast", "sha", "m", "chat:128", true);
+        // Node total VRAM from discovery devices → headroom denominator.
+        r.gpus_json =
+            Some(r#"[{"name":"RTX 5090","vram_total_mb":32000},{"name":"RTX 5090","vram_total_mb":32000}]"#.into());
+        r.vram_used_mb = Some(42000);
+        r.gpu_util_pct = Some(90);
+        r.gpu_temp_c = Some(66);
+        s.insert_run(&r).unwrap();
+        let rows = s.report_rows().unwrap();
+        let row = &rows[0];
+        assert_eq!(row.vram_used_mb_median, Some(42000.0));
+        assert_eq!(row.vram_total_mb, Some(64000)); // 2× 32000
+        assert_eq!(row.gpu_util_pct_median, Some(90.0));
+        assert_eq!(row.gpu_temp_c_median, Some(66.0));
+    }
+
+    #[test]
+    fn report_surfaces_concurrency_burst_metrics() {
+        let s = Store::open_in_memory().unwrap();
+        // Two concurrency:8 burst-runs with shed load and a queue-wait tail.
+        for (qw, rej) in [(120.0, 1u32), (180.0, 3u32)] {
+            let mut r = rec("beast", "sha", "m", "concurrency:8", true);
+            r.concurrency = Some(8);
+            r.ttft_p95_s = Some(0.9);
+            r.queue_wait_ms = Some(qw);
+            r.rejected = Some(rej);
+            s.insert_run(&r).unwrap();
+        }
+        let row = s
+            .report_rows()
+            .unwrap()
+            .into_iter()
+            .find(|r| r.scenario_id == "concurrency:8")
+            .unwrap();
+        assert_eq!(row.concurrency, Some(8));
+        assert_eq!(row.queue_wait_ms_median, Some(150.0)); // median(120,180)
+        assert_eq!(row.rejected_median, Some(2.0)); // median(1,3)
+        assert_eq!(row.ttft_p95_load_s, Some(0.9));
+    }
+
+    #[test]
+    fn capability_runs_store_artifact_and_accept_scores() {
+        let s = Store::open_in_memory().unwrap();
+        let mut r = rec("beast", "sha", "m", "capability:plan", true);
+        r.artifact = Some("a thorough implementation plan...".into());
+        s.insert_run(&r).unwrap();
+        // A non-capability row must not appear in capability_runs.
+        s.insert_run(&rec("beast", "sha", "m", "chat:128", true))
+            .unwrap();
+
+        let unscored = s.capability_runs(true).unwrap();
+        assert_eq!(unscored.len(), 1);
+        let id = unscored[0].id;
+        assert!(unscored[0].artifact.as_deref().unwrap().contains("plan"));
+        assert!(unscored[0].quality_score.is_none());
+
+        // Score it; it then drops off the unscored worklist.
+        assert_eq!(s.set_score(id, 7.5, "manual").unwrap(), 1);
+        assert!(s.capability_runs(true).unwrap().is_empty());
+        let all = s.capability_runs(false).unwrap();
+        assert_eq!(all[0].quality_score, Some(7.5));
+        assert_eq!(all[0].scorer.as_deref(), Some("manual"));
+        // Unknown id updates nothing.
+        assert_eq!(s.set_score(999_999, 1.0, "manual").unwrap(), 0);
+    }
+
+    #[test]
+    fn swap_costs_pivots_swap_rows() {
+        let s = Store::open_in_memory().unwrap();
+        for (unload, load) in [(300u64, 24000u64), (340, 26000)] {
+            let mut r = rec("beast", "sha", "m", "swap", true);
+            r.swap_unload_ms = Some(unload);
+            r.swap_load_ms = Some(load);
+            r.ttft_s = Some(2.5); // cold first-request
+            r.total_s = Some(5.0);
+            s.insert_run(&r).unwrap();
+        }
+        // A non-swap row must be ignored by swap_costs.
+        s.insert_run(&rec("beast", "sha", "m", "chat:128", true))
+            .unwrap();
+        let costs = s.swap_costs().unwrap();
+        assert_eq!(costs.len(), 1);
+        let c = &costs[0];
+        assert_eq!(c.unload_ms_median, Some(320.0)); // median(300,340)
+        assert_eq!(c.load_ms_median, Some(25000.0)); // median(24000,26000)
+        assert_eq!(c.cold_ttft_s_median, Some(2.5));
+        assert_eq!(c.samples, 2);
+    }
+
+    #[test]
+    fn scaling_pivots_chat_cells_and_computes_flatness() {
+        let s = Store::open_in_memory().unwrap();
+        // Two context points for one model: decode tok/s 50 @128, 45 @4096.
+        let mut small = rec("beast", "sha", "m", "chat:128", true);
+        small.prompt_size_approx = 128;
+        small.decode_tps = Some(50.0);
+        small.prefill_ms = Some(100);
+        small.prefill_tokens = Some(128);
+        s.insert_run(&small).unwrap();
+        let mut big = rec("beast", "sha", "m", "chat:4096", true);
+        big.prompt_size_approx = 4096;
+        big.decode_tps = Some(45.0);
+        big.prefill_ms = Some(1000);
+        big.prefill_tokens = Some(4096);
+        s.insert_run(&big).unwrap();
+        // A concurrency cell must NOT leak into the scaling curve.
+        let mut conc = rec("beast", "sha", "m", "concurrency:8", true);
+        conc.concurrency = Some(8);
+        s.insert_run(&conc).unwrap();
+
+        let curves = s.scaling().unwrap();
+        assert_eq!(curves.len(), 1);
+        let c = &curves[0];
+        assert_eq!(c.points.len(), 2); // only the two chat:<n> points
+        assert_eq!(c.points[0].prompt_size, 128); // ordered ascending
+        assert_eq!(c.points[1].prompt_size, 4096);
+        // flatness = decode@largest / decode@smallest = 45/50 = 0.9
+        assert!((c.decode_flatness.unwrap() - 0.9).abs() < 1e-9);
+    }
+
+    #[test]
+    fn gpu_total_vram_sums_devices() {
+        let j = r#"[{"vram_total_mb":32000},{"vram_total_mb":32000}]"#;
+        assert_eq!(gpu_total_vram_mb(j), Some(64000));
+        assert_eq!(gpu_total_vram_mb("[]"), None);
+        assert_eq!(gpu_total_vram_mb("not json"), None);
+    }
+
+    #[test]
+    fn percentile_nearest_rank() {
+        let vals = || [1.0, 2.0, 3.0, 4.0, 5.0].into_iter();
+        assert_eq!(percentile(vals(), 50.0), Some(3.0));
+        assert_eq!(percentile(vals(), 95.0), Some(5.0));
+        assert_eq!(percentile(vals(), 99.0), Some(5.0));
+        assert_eq!(percentile(std::iter::empty(), 95.0), None);
+    }
+
+    #[test]
+    fn migration_is_idempotent_and_backfills() {
+        // A DB whose `runs` table predates the prefill columns: create the
+        // pre-#85 shape, insert a row, then run ensure_columns twice.
+        let conn = Connection::open_in_memory().unwrap();
+        conn.execute_batch(
+            "CREATE TABLE runs (id INTEGER PRIMARY KEY, ttft_s REAL);
+             INSERT INTO runs (ttft_s) VALUES (0.1);",
+        )
+        .unwrap();
+        for _ in 0..2 {
+            Store::ensure_columns(
+                &conn,
+                "runs",
+                &[("prefill_ms", "INTEGER"), ("decode_ms", "INTEGER")],
+            )
+            .unwrap();
+        }
+        // Columns now exist and the old row reads them back as NULL.
+        let got: Option<i64> = conn
+            .query_row("SELECT prefill_ms FROM runs", [], |r| r.get(0))
+            .unwrap();
+        assert_eq!(got, None);
+    }
+
     #[test]
     fn gpu_label_formats() {
         let two = r#"[{"name":"NVIDIA GeForce RTX 5090"},{"name":"NVIDIA GeForce RTX 5090"}]"#;

crates/helexa-bench/src/sweep.rs

@@ -9,11 +9,11 @@
 
 use crate::client::TargetClient;
 use crate::config::{BenchConfig, TargetConfig, TargetKind};
-use crate::scenario::{RunCtx, build_scenarios};
+use crate::scenario::{RunCtx, ScenarioMetrics, build_scenarios};
 use crate::store::{RunRecord, Store};
 use anyhow::Result;
 use cortex_core::build_info::BuildInfo;
-use cortex_core::discovery::DiscoveryResponse;
+use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
 use cortex_core::harness::ModelInfo;
 
 /// helexa-bench's own build version.
@@ -38,6 +38,38 @@ pub struct SweepSummary {
     pub targets_unreachable: usize,
 }
 
+/// Node-level GPU telemetry folded from one `/health` snapshot (#87):
+/// VRAM used summed across the node's devices, and the hottest/busiest
+/// single device for utilization and temperature.
+#[derive(Debug, Clone, Copy)]
+struct HealthAgg {
+    vram_used_mb: u64,
+    gpu_util_pct: u32,
+    gpu_temp_c: u32,
+}
+
+/// Cold-load / model-swap timing for one measure_swap cycle (#90).
+#[derive(Debug, Clone, Copy)]
+struct SwapTiming {
+    unload_ms: u64,
+    load_ms: u64,
+}
+
+impl HealthAgg {
+    fn from_health(h: &HealthResponse) -> Self {
+        HealthAgg {
+            vram_used_mb: h.devices.iter().map(|d| d.vram_used_mb).sum(),
+            gpu_util_pct: h
+                .devices
+                .iter()
+                .map(|d| d.utilization_pct)
+                .max()
+                .unwrap_or(0),
+            gpu_temp_c: h.devices.iter().map(|d| d.temp_c).max().unwrap_or(0),
+        }
+    }
+}
+
 pub struct Sweeper {
     cfg: BenchConfig,
     client: TargetClient,
@@ -72,6 +104,105 @@ impl Sweeper {
         }
     }
 
+    /// Deliberate cold-load / model-swap cost measurement (#90), invoked by
+    /// the `swap-cost` subcommand — **never** the continuous sweep. For each
+    /// neuron target and each currently-warm model: unload it, time the
+    /// reload, then time a cold first request. This takes the model offline
+    /// for the reload, so it is an explicit operator action (maintenance
+    /// window), recorded under `scenario_id = "swap"`.
+    pub async fn swap_cost_once(&self) -> Result<SweepSummary> {
+        let mut summary = SweepSummary::default();
+        for target in &self.cfg.targets {
+            if target.kind != TargetKind::Neuron {
+                continue; // load/unload is a neuron-native operation
+            }
+            let build = match self.client.fetch_version(target).await {
+                Ok(b) => b,
+                Err(e) => {
+                    summary.targets_unreachable += 1;
+                    tracing::warn!(target = %target.name, error = %format!("{e:#}"), "swap: target unreachable");
+                    continue;
+                }
+            };
+            let discovery = self.client.fetch_discovery(target).await.unwrap_or(None);
+            let models = self.client.warm_models(target).await.unwrap_or_default();
+            for model in &models {
+                match self
+                    .measure_swap(target, &build, discovery.as_ref(), model)
+                    .await
+                {
+                    Ok(()) => summary.measured += 1,
+                    Err(e) => {
+                        summary.failed += 1;
+                        tracing::warn!(target = %target.name, model = %model.id, error = %format!("{e:#}"), "swap: measurement failed");
+                    }
+                }
+            }
+        }
+        Ok(summary)
+    }
+
+    /// Unload → timed reload → timed cold first request for one model.
+    async fn measure_swap(
+        &self,
+        target: &TargetConfig,
+        build: &BuildInfo,
+        discovery: Option<&DiscoveryResponse>,
+        model: &ModelInfo,
+    ) -> Result<()> {
+        let spec = TargetClient::spec_from_info(model)?;
+        tracing::warn!(target = %target.name, model = %model.id, "swap: unloading (model goes offline until reload)");
+
+        let t0 = std::time::Instant::now();
+        self.client.unload_model(target, &model.id).await?;
+        let unload_ms = t0.elapsed().as_millis() as u64;
+
+        let t1 = std::time::Instant::now();
+        self.client.load_model(target, &spec).await?;
+        let load_ms = t1.elapsed().as_millis() as u64;
+        tracing::info!(target = %target.name, model = %model.id, unload_ms, load_ms, "swap: reloaded; measuring cold first request");
+
+        // Cold first request — caches empty straight after the load.
+        let ctx = RunCtx {
+            client: self.client.http(),
+            chat_url: self.client.chat_url(target),
+            model_id: model.id.clone(),
+            max_tokens: self.cfg.scenarios.max_tokens,
+            timeout: self.cfg.bench.request_timeout(),
+        };
+        let cold = crate::scenario::cold_probe(&ctx).await;
+        let swap = SwapTiming { unload_ms, load_ms };
+        let rec = match &cold {
+            Ok(m) => self.build_record(
+                target,
+                build,
+                discovery,
+                model,
+                "swap",
+                0,
+                Ok(m),
+                None,
+                Some(swap),
+            ),
+            Err(e) => {
+                let msg = format!("{e:#}");
+                self.build_record(
+                    target,
+                    build,
+                    discovery,
+                    model,
+                    "swap",
+                    0,
+                    Err(&msg),
+                    None,
+                    Some(swap),
+                )
+            }
+        };
+        self.store.insert_run(&rec)?;
+        Ok(())
+    }
+
     /// One full pass over all targets.
     pub async fn run_once(&self) -> Result<SweepSummary> {
         let mut summary = SweepSummary::default();
@@ -131,7 +262,21 @@ impl Sweeper {
                 }
 
                 for i in 0..need {
-                    match scenario.run(&ctx).await {
+                    let result = scenario.run(&ctx).await;
+                    // Sample GPU telemetry right after the run, while the
+                    // model is loaded and decode VRAM is at its recent peak
+                    // (#87). neuron's /health is ~5s-cached, so this is a
+                    // coarse high-water proxy, not an instantaneous peak — but
+                    // it's the headroom signal we can read over the wire. A
+                    // flaky /health degrades to None, never a failed run.
+                    let health = self
+                        .client
+                        .fetch_health(target)
+                        .await
+                        .ok()
+                        .flatten()
+                        .map(|h| HealthAgg::from_health(&h));
+                    match result {
                         Ok(m) => {
                             let rec = self.build_record(
                                 target,
@@ -141,6 +286,8 @@ impl Sweeper {
                                 scenario.id(),
                                 scenario.prompt_size(),
                                 Ok(&m),
+                                health,
+                                None,
                             );
                             self.store.insert_run(&rec)?;
                             summary.measured += 1;
@@ -160,6 +307,8 @@ impl Sweeper {
                                 scenario.id(),
                                 scenario.prompt_size(),
                                 Err(&msg),
+                                health,
+                                None,
                             );
                             self.store.insert_run(&rec)?;
                             summary.failed += 1;
@@ -186,19 +335,14 @@ impl Sweeper {
         scenario_id: &str,
         prompt_size: u32,
         result: Result<&crate::scenario::ScenarioMetrics, &str>,
+        health: Option<HealthAgg>,
+        swap: Option<SwapTiming>,
     ) -> RunRecord {
-        let (ok, error, ttft, decode, total, prompt_tokens, completion) = match result {
+        let (m, error): (Option<&ScenarioMetrics>, Option<String>) = match result {
-            Ok(m) => (
+            Ok(m) => (Some(m), None),
-                true,
+            Err(e) => (None, Some(e.to_string())),
-                None,
-                Some(m.ttft_s),
-                m.decode_tps,
-                Some(m.total_s),
-                m.prompt_tokens,
-                Some(m.completion_tokens),
-            ),
-            Err(e) => (false, Some(e.to_string()), None, None, None, None, None),
         };
+        let ok = m.is_some();
 
         RunRecord {
             ts: chrono::Utc::now().to_rfc3339(),
@@ -230,12 +374,29 @@ impl Sweeper {
                 .unwrap_or_else(|_| "[]".to_string()),
             scenario_id: scenario_id.to_string(),
             prompt_size_approx: prompt_size,
-            prompt_tokens_actual: prompt_tokens,
+            prompt_tokens_actual: m.and_then(|m| m.prompt_tokens),
             max_tokens: self.cfg.scenarios.max_tokens,
-            ttft_s: ttft,
+            ttft_s: m.map(|m| m.ttft_s),
-            decode_tps: decode,
+            decode_tps: m.and_then(|m| m.decode_tps),
-            total_s: total,
+            total_s: m.map(|m| m.total_s),
-            completion_tokens: completion,
+            completion_tokens: m.map(|m| m.completion_tokens),
+            prefill_ms: m.and_then(|m| m.prefill_ms),
+            decode_ms: m.and_then(|m| m.decode_ms),
+            prefill_tokens: m.and_then(|m| m.prefill_tokens),
+            vram_used_mb: health.map(|h| h.vram_used_mb),
+            gpu_util_pct: health.map(|h| h.gpu_util_pct),
+            gpu_temp_c: health.map(|h| h.gpu_temp_c),
+            concurrency: m.and_then(|m| m.concurrency),
+            ttft_p95_s: m.and_then(|m| m.ttft_p95_s),
+            queue_wait_ms: m.and_then(|m| m.queue_wait_ms_median),
+            rejected: m.and_then(|m| m.rejected),
+            swap_unload_ms: swap.map(|s| s.unload_ms),
+            swap_load_ms: swap.map(|s| s.load_ms),
+            // Capability artifact (#91); score/scorer are attached later by
+            // the `score` subcommand or a future LLM-judge.
+            artifact: m.and_then(|m| m.artifact.clone()),
+            quality_score: None,
+            scorer: None,
             ok,
             error,
         }

crates/helexa-bench/tests/api.rs

@@ -46,6 +46,21 @@ fn rec(
         decode_tps: if ok { Some(30.0) } else { None },
         total_s: if ok { Some(2.0) } else { None },
         completion_tokens: if ok { Some(60) } else { None },
+        prefill_ms: if ok { Some(150) } else { None },
+        decode_ms: if ok { Some(1800) } else { None },
+        prefill_tokens: if ok { Some(130) } else { None },
+        vram_used_mb: if ok { Some(42000) } else { None },
+        gpu_util_pct: if ok { Some(85) } else { None },
+        gpu_temp_c: if ok { Some(63) } else { None },
+        concurrency: None,
+        ttft_p95_s: None,
+        queue_wait_ms: None,
+        rejected: None,
+        swap_unload_ms: None,
+        swap_load_ms: None,
+        artifact: None,
+        quality_score: None,
+        scorer: None,
         ok,
         error: if ok { None } else { Some("boom".into()) },
     }

crates/helexa-bench/tests/sweep_integration.rs

@@ -89,6 +89,9 @@ fn config_for(endpoint: String, db_path: String) -> BenchConfig {
         scenarios: ScenarioConfig {
             prompt_sizes: vec![128], // single scenario keeps assertions simple
             max_tokens: 16,
+            concurrency_levels: Vec::new(),
+            concurrency_prompt_tokens: 512,
+            capability_probes: Vec::new(),
         },
         api: Default::default(),
         targets: vec![TargetConfig {

crates/helexa-router/Cargo.toml

@@ -0,0 +1,41 @@
+[package]
+name = "helexa-router"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[[bin]]
+name = "helexa-router"
+path = "src/main.rs"
+
+[lib]
+name = "helexa_router"
+path = "src/lib.rs"
+
+[dependencies]
+cortex-core = { workspace = true }
+helexa-stream = { path = "../helexa-stream" }
+
+tokio = { workspace = true }
+axum = { workspace = true }
+tower-http = { workspace = true }
+reqwest = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+figment = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+clap = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+chrono = { workspace = true }
+
+[dev-dependencies]
+# Jail (isolated cwd + env) for config tests.
+figment = { workspace = true, features = ["test"] }
+# Self-signed cert generation + a minimal HTTPS server for the outbound
+# TLS-pinning tests (#74).
+rcgen = "0.13"
+rustls = "0.23"
+tokio-rustls = "0.26"

crates/helexa-router/src/catalogue.rs

@@ -0,0 +1,260 @@
+//! Federation catalogue (#75) — the router's aggregate `/v1/models`.
+//!
+//! Presents the **deduped union** of every reachable cortex's `/v1/models`
+//! as the router's own catalogue, so an opencode client doing discovery
+//! against the router resolves the whole federation without knowing about
+//! operators or cortexes (resolves #61's "Router/discovery contract").
+//!
+//! Re-tiering: the fractal design is neuron ← cortex ← router. At the
+//! router tier the "nodes" are **cortexes**, so the merged entry's
+//! `feasible_on` / `locations` are rewritten to **operator names**, not the
+//! neuron names a cortex reports. That keeps the federation view honest
+//! ("served by these operators") without leaking each operator's internal
+//! topology (neuron names, per-device VRAM) to end users.
+//!
+//! Conflict resolution when operators advertise the same model with
+//! different enrichment:
+//! - **`limit`** → the *tightest* (smallest `context`), so a client never
+//!   overflows the most-constrained operator that might serve it (same rule
+//!   cortex uses across its neurons).
+//! - **`cost`** → the *cheapest* (lowest input, then output), the
+//!   federation "from" price. Richer policy (a range, region/price-aware
+//!   selection) couples to #68 and is left as a follow-up.
+
+use crate::state::{CortexTopology, entry_feasible};
+use cortex_core::harness::{ModelCost, ModelLimit};
+use cortex_core::node::{CortexModelEntry, ModelLocation, ModelStatus};
+use std::collections::HashMap;
+
+/// Build the federation catalogue: the deduped union of every reachable
+/// cortex's serveable models, merged across operators and sorted by id.
+pub fn aggregate_models(topology: &HashMap<String, CortexTopology>) -> Vec<CortexModelEntry> {
+    // Iterate cortexes in name order so `feasible_on` / `locations` and the
+    // limit/cost tie-breaks are deterministic regardless of map ordering.
+    let mut cortexes: Vec<(&String, &CortexTopology)> = topology.iter().collect();
+    cortexes.sort_by(|a, b| a.0.cmp(b.0));
+
+    let mut merged: HashMap<String, CortexModelEntry> = HashMap::new();
+    for (cortex_name, t) in cortexes {
+        if !t.reachable {
+            continue;
+        }
+        for entry in t.models.values() {
+            // Only surface models the cortex can actually serve — a
+            // catalogue-only entry no neuron can host shouldn't appear in
+            // the federation view.
+            if !entry_feasible(entry) {
+                continue;
+            }
+            merged
+                .entry(entry.id.clone())
+                .and_modify(|acc| merge_into(acc, cortex_name, entry))
+                .or_insert_with(|| router_entry(cortex_name, entry));
+        }
+    }
+
+    let mut out: Vec<CortexModelEntry> = merged.into_values().collect();
+    out.sort_by(|a, b| a.id.cmp(&b.id));
+    // Re-derive the flat ecosystem fields (#78) from the merged (tightest)
+    // limit — the values deserialized from each cortex are per-operator and
+    // may not match the federation-wide merge.
+    for e in &mut out {
+        e.sync_flat_limit();
+    }
+    out
+}
+
+/// Seed a federation entry from the first cortex that serves the model,
+/// re-tiering `feasible_on` / `locations` to the operator name.
+fn router_entry(cortex: &str, e: &CortexModelEntry) -> CortexModelEntry {
+    CortexModelEntry {
+        id: e.id.clone(),
+        object: "model".into(),
+        created: e.created,
+        owned_by: e.owned_by.clone(),
+        loaded: e.loaded,
+        feasible_on: vec![cortex.to_string()],
+        locations: loaded_location(cortex, e),
+        capabilities: e.capabilities.clone(),
+        limit: e.limit.clone(),
+        cost: e.cost.clone(),
+        tool_call: e.tool_call,
+        reasoning: e.reasoning,
+        // Derived from `limit` by the final sync pass in aggregate_models.
+        max_model_len: None,
+        max_input_tokens: None,
+        max_output_tokens: None,
+    }
+}
+
+/// Fold another cortex's view of the same model into the merged entry.
+fn merge_into(acc: &mut CortexModelEntry, cortex: &str, e: &CortexModelEntry) {
+    acc.loaded |= e.loaded;
+    acc.feasible_on.push(cortex.to_string());
+    acc.locations.extend(loaded_location(cortex, e));
+    for cap in &e.capabilities {
+        if !acc.capabilities.contains(cap) {
+            acc.capabilities.push(cap.clone());
+        }
+    }
+    acc.tool_call |= e.tool_call;
+    acc.reasoning |= e.reasoning;
+    acc.limit = tightest_limit(acc.limit.take(), e.limit.clone());
+    acc.cost = cheapest_cost(acc.cost.take(), e.cost.clone());
+}
+
+/// A single cortex-tier location when the model is loaded at that operator;
+/// empty when only cold-loadable. Neuron-level VRAM is deliberately dropped.
+fn loaded_location(cortex: &str, e: &CortexModelEntry) -> Vec<ModelLocation> {
+    if e.loaded {
+        vec![ModelLocation {
+            node: cortex.to_string(),
+            status: ModelStatus::Loaded,
+            vram_estimate_mb: None,
+        }]
+    } else {
+        Vec::new()
+    }
+}
+
+/// Smaller `context` wins — never advertise more headroom than the
+/// most-constrained operator can honour.
+fn tightest_limit(a: Option<ModelLimit>, b: Option<ModelLimit>) -> Option<ModelLimit> {
+    match (a, b) {
+        (None, x) | (x, None) => x,
+        (Some(a), Some(b)) => Some(if b.context < a.context { b } else { a }),
+    }
+}
+
+/// Cheapest by (input, output) price — the federation "from" price.
+fn cheapest_cost(a: Option<ModelCost>, b: Option<ModelCost>) -> Option<ModelCost> {
+    match (a, b) {
+        (None, x) | (x, None) => x,
+        (Some(a), Some(b)) => Some(if (b.input, b.output) < (a.input, a.output) {
+            b
+        } else {
+            a
+        }),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::state::CortexTopology;
+
+    fn entry(id: &str, loaded: bool, feasible: bool) -> CortexModelEntry {
+        CortexModelEntry {
+            id: id.into(),
+            object: "model".into(),
+            created: 0,
+            owned_by: "helexa".into(),
+            loaded,
+            feasible_on: if feasible || loaded {
+                vec!["some-neuron".into()]
+            } else {
+                vec![]
+            },
+            locations: vec![],
+            capabilities: vec![],
+            limit: None,
+            cost: None,
+            tool_call: false,
+            reasoning: false,
+            max_model_len: None,
+            max_input_tokens: None,
+            max_output_tokens: None,
+        }
+    }
+
+    fn cortex(reachable: bool, entries: Vec<CortexModelEntry>) -> CortexTopology {
+        CortexTopology {
+            reachable,
+            consecutive_failures: 0,
+            last_poll: None,
+            healthy_nodes: 1,
+            total_nodes: 1,
+            models: entries.into_iter().map(|e| (e.id.clone(), e)).collect(),
+        }
+    }
+
+    #[test]
+    fn dedupes_and_merges_availability_across_cortexes() {
+        let mut topo = HashMap::new();
+        // c-a: model loaded. c-b: same model only cold-loadable.
+        topo.insert("c-a".into(), cortex(true, vec![entry("m", true, true)]));
+        topo.insert("c-b".into(), cortex(true, vec![entry("m", false, true)]));
+
+        let out = aggregate_models(&topo);
+        assert_eq!(out.len(), 1, "duplicate model id collapses to one");
+        let m = &out[0];
+        assert!(m.loaded, "loaded somewhere → loaded");
+        // feasible_on re-tiered to operator names, both present, sorted.
+        assert_eq!(m.feasible_on, vec!["c-a".to_string(), "c-b".to_string()]);
+        // Only the loaded operator contributes a location, named by operator.
+        assert_eq!(m.locations.len(), 1);
+        assert_eq!(m.locations[0].node, "c-a");
+        assert_eq!(m.locations[0].vram_estimate_mb, None);
+    }
+
+    #[test]
+    fn unreachable_cortex_is_excluded() {
+        let mut topo = HashMap::new();
+        topo.insert("up".into(), cortex(true, vec![entry("m", true, true)]));
+        topo.insert(
+            "down".into(),
+            cortex(false, vec![entry("other", true, true)]),
+        );
+        let out = aggregate_models(&topo);
+        assert_eq!(out.len(), 1);
+        assert_eq!(out[0].id, "m");
+    }
+
+    #[test]
+    fn catalogue_only_infeasible_entries_are_hidden() {
+        let mut topo = HashMap::new();
+        topo.insert("c".into(), cortex(true, vec![entry("ghost", false, false)]));
+        assert!(aggregate_models(&topo).is_empty());
+    }
+
+    #[test]
+    fn preserves_tightest_limit_and_cheapest_cost() {
+        let mut a = entry("m", true, true);
+        a.limit = Some(ModelLimit {
+            context: 32_768,
+            input: None,
+            output: 4096,
+        });
+        a.cost = Some(ModelCost {
+            input: 0.50,
+            output: 1.50,
+            cache_read: None,
+            cache_write: None,
+        });
+        let mut b = entry("m", true, true);
+        b.limit = Some(ModelLimit {
+            context: 16_384, // tighter
+            input: None,
+            output: 4096,
+        });
+        b.cost = Some(ModelCost {
+            input: 0.20, // cheaper
+            output: 0.80,
+            cache_read: None,
+            cache_write: None,
+        });
+
+        let mut topo = HashMap::new();
+        topo.insert("c-a".into(), cortex(true, vec![a]));
+        topo.insert("c-b".into(), cortex(true, vec![b]));
+
+        let out = aggregate_models(&topo);
+        assert_eq!(out.len(), 1);
+        assert_eq!(out[0].limit.as_ref().unwrap().context, 16_384);
+        assert_eq!(out[0].cost.as_ref().unwrap().input, 0.20);
+        // Flat #78 fields re-derived from the merged (tightest) limit.
+        assert_eq!(out[0].max_model_len, Some(16_384));
+        assert_eq!(out[0].max_input_tokens, None);
+        assert_eq!(out[0].max_output_tokens, Some(4096));
+    }
+}

crates/helexa-router/src/config.rs

@@ -0,0 +1,100 @@
+use figment::{
+    Figment,
+    providers::{Env, Format, Toml},
+};
+use serde::{Deserialize, Serialize};
+use std::path::Path;
+
+/// Top-level `helexa-router` configuration.
+///
+/// Loaded from TOML with `HELEXA_ROUTER_`-prefixed env overrides (using
+/// `__` as the nesting separator), matching the cortex/neuron convention.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct RouterConfig {
+    pub router: RouterSettings,
+    /// Downstream cortex endpoints the router can dispatch to. The skeleton
+    /// (#70) only loads these; capacity/catalogue polling (#72) and
+    /// capacity-aware dispatch (#73) consume them later.
+    #[serde(default)]
+    pub cortexes: Vec<CortexEndpoint>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct RouterSettings {
+    /// Address to listen on for the inbound API (e.g. "0.0.0.0:8088").
+    ///
+    /// Plaintext only — operator/edge nginx terminates client TLS in front
+    /// of the router (see #69's TLS posture). The router never owns an
+    /// inbound TLS listener.
+    pub listen: String,
+    /// How often (seconds) the background poller refreshes each cortex's
+    /// health + `/v1/models` topology (#72). Defaults to 10s, matching the
+    /// cortex↔neuron poll cadence one tier down.
+    #[serde(default = "default_poll_interval_secs")]
+    pub poll_interval_secs: u64,
+    /// This router instance's region (e.g. "eu-west"). When set, dispatch
+    /// (#73) prefers cortexes whose `region` matches, before falling back to
+    /// any feasible cortex. `None` → no geo affinity.
+    #[serde(default)]
+    pub region: Option<String>,
+}
+
+fn default_poll_interval_secs() -> u64 {
+    10
+}
+
+/// One downstream cortex the router may proxy to. The router verifies the
+/// cortex's outbound TLS cert (#74) and routes on capacity (#73); it holds
+/// no entitlement logic of its own and forwards the client bearer verbatim.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct CortexEndpoint {
+    /// Human-readable label (e.g. "lair-cafe").
+    pub name: String,
+    /// Base URL of the cortex gateway (e.g. "https://cortex.example.com").
+    pub endpoint: String,
+    /// Optional region tag (e.g. "eu-west") for geo affinity in dispatch
+    /// (#73). `None` → no region preference applies to this cortex.
+    #[serde(default)]
+    pub region: Option<String>,
+    /// Path to a PEM trust anchor that **enrols** this cortex (#74): the
+    /// expected CA (or self-signed cert) the cortex's TLS cert must chain
+    /// to. When set on an `https://` endpoint, the router builds a client
+    /// that trusts **only** this anchor (platform roots disabled), so the
+    /// outbound router→cortex hop — which carries the client's bearer —
+    /// reaches a cert the router was told to expect, and a rogue endpoint
+    /// presenting any other (even publicly-valid) cert is rejected at the
+    /// TLS handshake. A rejected handshake surfaces as a connection error,
+    /// which the poller (#72) already treats as unreachable → excluded.
+    ///
+    /// `None` → standard platform-root validation (use for cortexes behind
+    /// a publicly-trusted cert, or plaintext `http://` on a private network
+    /// where the WireGuard mesh is the trust boundary).
+    #[serde(default)]
+    pub tls_ca: Option<String>,
+}
+
+impl RouterConfig {
+    /// Load configuration from a TOML file, with environment variable
+    /// overrides prefixed with `HELEXA_ROUTER_` and `__` as the separator
+    /// (e.g. `HELEXA_ROUTER_ROUTER__LISTEN=0.0.0.0:8088`).
+    pub fn load(path: impl AsRef<Path>) -> Result<Self, Box<figment::Error>> {
+        Figment::new()
+            .merge(Toml::file(path))
+            .merge(Env::prefixed("HELEXA_ROUTER_").split("__"))
+            .extract()
+            .map_err(Box::new)
+    }
+}
+
+impl Default for RouterConfig {
+    fn default() -> Self {
+        Self {
+            router: RouterSettings {
+                listen: "0.0.0.0:8088".into(),
+                poll_interval_secs: default_poll_interval_secs(),
+                region: None,
+            },
+            cortexes: vec![],
+        }
+    }
+}

crates/helexa-router/src/dispatch.rs

@@ -0,0 +1,221 @@
+//! Capacity-aware dispatch (#73) — the router's data path.
+//!
+//! Given an inbound request's `model`, pick a reachable cortex that can
+//! serve it (preferring warm/loaded, region-affine, higher-headroom),
+//! forward the client's bearer **unchanged** (auth stays at cortex), and
+//! stream the response back verbatim via the shared [`helexa_stream`]
+//! module. Cortex's #63-shaped rejections (`429 rate_limit_exceeded`,
+//! `400 context_length_exceeded`, …) pass through untouched. Transport
+//! failures fail over to the next feasible cortex; a genuine HTTP response —
+//! any status — is returned as-is and never retried away.
+//!
+//! The router holds **no entitlement logic**: it routes on capacity, not
+//! budget.
+
+use crate::config::CortexEndpoint;
+use crate::error::envelope_response;
+use crate::state::RouterState;
+use axum::body::Bytes;
+use axum::http::HeaderMap;
+use axum::response::Response;
+use cortex_core::error_envelope::OpenAiError;
+use helexa_stream::{ChunkObserver, StreamError};
+use std::cmp::Reverse;
+use std::collections::HashMap;
+
+/// Retry-After hint (seconds) on the router's own transient rejections.
+const RETRY_AFTER_SECS: u64 = 5;
+
+/// Outcome of choosing where to send a request.
+#[derive(Debug, PartialEq, Eq)]
+pub enum Selection {
+    /// Feasible reachable cortexes, best-first (failover order).
+    Candidates(Vec<CortexEndpoint>),
+    /// Some cortex knows the model but none are reachable right now → 503.
+    NoReachableCapacity,
+    /// No configured cortex serves the model at all → 404.
+    UnknownModel,
+}
+
+/// Rank the reachable cortexes that can serve `model`, best-first.
+///
+/// Ordering (each a tie-break for the next): loaded/warm before cold-loadable
+/// · region match before not · more healthy nodes before fewer · name for
+/// determinism.
+pub async fn select_cortexes(state: &RouterState, model: &str) -> Selection {
+    let topo = state.topology.read().await;
+    let by_name: HashMap<&str, &CortexEndpoint> = state
+        .cortexes
+        .iter()
+        .map(|c| (c.name.as_str(), c))
+        .collect();
+
+    let mut ranked: Vec<Ranked> = Vec::new();
+    let mut known_anywhere = false;
+
+    for (name, t) in topo.iter() {
+        let Some(entry) = t.models.get(model) else {
+            continue;
+        };
+        if !crate::state::entry_feasible(entry) {
+            continue;
+        }
+        // Known even via an unreachable cortex's last-good poll — lets us
+        // tell "temporarily down" (503) from "nobody serves it" (404).
+        known_anywhere = true;
+        if !t.reachable {
+            continue;
+        }
+        let Some(ep) = by_name.get(name.as_str()) else {
+            continue;
+        };
+        let region_match = match (&state.region, &ep.region) {
+            (Some(r), Some(cr)) => r == cr,
+            _ => false,
+        };
+        ranked.push(Ranked {
+            loaded: entry.loaded,
+            region_match,
+            healthy_nodes: t.healthy_nodes,
+            endpoint: (*ep).clone(),
+        });
+    }
+
+    if ranked.is_empty() {
+        return if known_anywhere {
+            Selection::NoReachableCapacity
+        } else {
+            Selection::UnknownModel
+        };
+    }
+
+    ranked.sort_by(|a, b| {
+        // false < true, so negate the "good" booleans to sort good first.
+        (
+            !a.loaded,
+            !a.region_match,
+            Reverse(a.healthy_nodes),
+            &a.endpoint.name,
+        )
+            .cmp(&(
+                !b.loaded,
+                !b.region_match,
+                Reverse(b.healthy_nodes),
+                &b.endpoint.name,
+            ))
+    });
+
+    Selection::Candidates(ranked.into_iter().map(|r| r.endpoint).collect())
+}
+
+struct Ranked {
+    loaded: bool,
+    region_match: bool,
+    healthy_nodes: u32,
+    endpoint: CortexEndpoint,
+}
+
+/// Proxy an inbound inference request to a capacity-bearing cortex.
+///
+/// `path` is the inference path to forward to (same on the cortex, e.g.
+/// `/v1/chat/completions`). The body is parsed only to extract `model`.
+pub async fn dispatch(
+    state: &RouterState,
+    path: &str,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    let Some(model) = extract_model(&body) else {
+        return envelope_response(OpenAiError::new(
+            400,
+            "invalid_request_error",
+            "missing_model_field",
+            "missing 'model' field in request body",
+        ));
+    };
+
+    let candidates = match select_cortexes(state, &model).await {
+        Selection::Candidates(c) => c,
+        Selection::UnknownModel => {
+            return envelope_response(
+                OpenAiError::new(
+                    404,
+                    "invalid_request_error",
+                    "model_not_found",
+                    format!("no operator serves model '{model}'"),
+                )
+                .with_param("model"),
+            );
+        }
+        Selection::NoReachableCapacity => {
+            return envelope_response(OpenAiError::service_unavailable(
+                format!("model '{model}' is temporarily unavailable on all operators"),
+                Some(RETRY_AFTER_SECS),
+            ));
+        }
+    };
+
+    // Try candidates in order, failing over only on transport errors. A
+    // genuine HTTP response (any status — including cortex's #63 429/400)
+    // is returned verbatim and never retried away.
+    for ep in &candidates {
+        // A candidate whose pinned TLS client failed to build (#74) is
+        // disabled — skip it and fail over, same as an unreachable cortex.
+        let Some(client) = state.client_for(&ep.name) else {
+            tracing::warn!(cortex = %ep.name, "no TLS client (disabled); skipping candidate");
+            continue;
+        };
+        let url = format!("{}{}", ep.endpoint, path);
+        tracing::info!(cortex = %ep.name, url = %url, model = %model, "dispatching");
+        match helexa_stream::forward_streaming(
+            client,
+            &url,
+            headers.clone(),
+            body.clone(),
+            NoopObserver,
+        )
+        .await
+        {
+            Ok(resp) => return resp,
+            Err(StreamError::Upstream(e)) => {
+                tracing::warn!(
+                    cortex = %ep.name,
+                    url = %url,
+                    error = %e,
+                    "cortex unreachable; failing over"
+                );
+                continue;
+            }
+            Err(StreamError::ResponseBuild(msg)) => {
+                tracing::error!(cortex = %ep.name, error = %msg, "failed to build proxied response");
+                return envelope_response(OpenAiError::without_code(
+                    500,
+                    "api_error",
+                    "failed to build proxied response",
+                ));
+            }
+        }
+    }
+
+    // Every feasible cortex failed to connect.
+    tracing::warn!(model = %model, tried = candidates.len(), "all feasible operators unreachable");
+    envelope_response(OpenAiError::service_unavailable(
+        format!("all operators able to serve '{model}' are unreachable"),
+        Some(RETRY_AFTER_SECS),
+    ))
+}
+
+/// Pull the `model` field out of a request body without re-serialising it.
+fn extract_model(body: &Bytes) -> Option<String> {
+    let v: serde_json::Value = serde_json::from_slice(body).ok()?;
+    v.get("model")?.as_str().map(str::to_string)
+}
+
+/// The router proxies bytes verbatim and keeps no per-request policy, so it
+/// needs no observation hooks. (Token metrics/metering stay at cortex.)
+struct NoopObserver;
+
+impl ChunkObserver for NoopObserver {
+    fn observe(&mut self, _chunk: &[u8]) {}
+    fn finish(&mut self) {}
+}

crates/helexa-router/src/error.rs

@@ -0,0 +1,27 @@
+//! Router adapter from the shared, axum-agnostic
+//! [`cortex_core::error_envelope::OpenAiError`] (#60/#63) to an axum
+//! [`Response`], setting `Retry-After` when the envelope carries one.
+//!
+//! cortex-core owns the envelope shape; this is the only place the router
+//! crosses from that data into axum. Mirrors cortex-gateway's adapter so
+//! the router's own rejections (no feasible operator, all unreachable) are
+//! the same #63-shaped envelopes clients already understand — distinct from
+//! cortex's rejections, which the router proxies through verbatim.
+
+use axum::http::{HeaderValue, StatusCode, header};
+use axum::response::{IntoResponse, Json, Response};
+use cortex_core::error_envelope::OpenAiError;
+
+/// Render an [`OpenAiError`] as an axum response (status + JSON envelope +
+/// optional `Retry-After`).
+pub fn envelope_response(err: OpenAiError) -> Response {
+    let status = StatusCode::from_u16(err.status).unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
+    let retry_after = err.retry_after_secs;
+    let mut response = (status, Json(err.body())).into_response();
+    if let Some(secs) = retry_after
+        && let Ok(value) = HeaderValue::from_str(&secs.to_string())
+    {
+        response.headers_mut().insert(header::RETRY_AFTER, value);
+    }
+    response
+}

crates/helexa-router/src/handlers.rs

@@ -0,0 +1,89 @@
+use crate::state::RouterState;
+use crate::{catalogue, dispatch};
+use axum::body::Bytes;
+use axum::http::HeaderMap;
+use axum::response::Response;
+use axum::{Json, Router, extract::State, routing::get, routing::post};
+use serde_json::{Value, json};
+use std::sync::Arc;
+
+/// Routes served by the router. Inference paths are capacity-aware-dispatched
+/// (#73) to a downstream cortex; `/health` and a stub `/v1/models` are local.
+pub fn api_routes() -> Router<Arc<RouterState>> {
+    Router::new()
+        .route("/v1/chat/completions", post(chat_completions))
+        .route("/v1/completions", post(completions))
+        .route("/v1/responses", post(responses))
+        .route("/v1/messages", post(messages))
+        .route("/v1/models", get(list_models))
+        .route("/health", get(health))
+        .route("/", get(health))
+}
+
+// ── Inference paths — forwarded verbatim to a chosen cortex ──────────
+//
+// Each handler dispatches to the same path on a capacity-bearing cortex.
+// The body is parsed only to read `model`; the bearer and bytes are
+// forwarded unchanged, and the SSE response streams back verbatim.
+
+async fn chat_completions(
+    State(state): State<Arc<RouterState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    dispatch::dispatch(&state, "/v1/chat/completions", headers, body).await
+}
+
+async fn completions(
+    State(state): State<Arc<RouterState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    dispatch::dispatch(&state, "/v1/completions", headers, body).await
+}
+
+async fn responses(
+    State(state): State<Arc<RouterState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    dispatch::dispatch(&state, "/v1/responses", headers, body).await
+}
+
+async fn messages(
+    State(state): State<Arc<RouterState>>,
+    headers: HeaderMap,
+    body: Bytes,
+) -> Response {
+    dispatch::dispatch(&state, "/v1/messages", headers, body).await
+}
+
+/// `GET /health` — router liveness plus a summary of downstream cortex
+/// reachability from the topology poller (#72). `status` reflects the
+/// router process itself (always `ok` if it answers); downstream health is
+/// the informational `cortexes` block, so a fully-degraded fleet doesn't
+/// make the router look dead to its own liveness probe.
+async fn health(State(state): State<Arc<RouterState>>) -> Json<Value> {
+    let topo = state.topology.read().await;
+    let reachable = topo.values().filter(|t| t.reachable).count();
+    Json(json!({
+        "status": "ok",
+        "cortexes": {
+            "configured": state.cortexes.len(),
+            "reachable": reachable,
+        }
+    }))
+}
+
+/// `GET /v1/models` — the federation catalogue (#75): the deduped union of
+/// every reachable cortex's `/v1/models`, so a client doing discovery
+/// against the router resolves the whole federation without knowing about
+/// operators or cortexes.
+async fn list_models(State(state): State<Arc<RouterState>>) -> Json<Value> {
+    let topo = state.topology.read().await;
+    let data: Vec<Value> = catalogue::aggregate_models(&topo)
+        .iter()
+        .map(|e| json!(e))
+        .collect();
+    Json(json!({ "object": "list", "data": data }))
+}

crates/helexa-router/src/lib.rs

@@ -0,0 +1,60 @@
+//! helexa-router — public multi-operator ingress proxy (router.helexa.ai).
+//!
+//! The router is the data-plane *ingress* tier: a geo-distributed,
+//! capacity-aware, OpenAI/Anthropic-compatible reverse proxy in front of
+//! many operator-run cortexes ("cortex-of-cortexes"). End users configure
+//! one `baseURL` and the router forwards their request to a cortex with
+//! capacity, proxying #63-shaped rejections back verbatim.
+//!
+//! It holds **zero entitlement logic** — auth/budget stays at cortex
+//! (epic #47); the router forwards the client bearer unchanged and routes
+//! on capacity (epic #69). A background [`poller`] keeps a live
+//! per-cortex topology (#72) that the dispatcher (#73) will route on.
+
+pub mod catalogue;
+pub mod config;
+pub mod dispatch;
+pub mod error;
+pub mod handlers;
+pub mod poller;
+pub mod state;
+
+use anyhow::Result;
+use config::RouterConfig;
+use std::sync::Arc;
+use tower_http::cors::CorsLayer;
+use tower_http::trace::TraceLayer;
+
+/// Build the axum application: handlers + CORS + tracing. No auth layer —
+/// the router asserts no identity of its own and forwards the client bearer
+/// to the downstream cortex, which authenticates it (#69).
+pub fn build_app(state: Arc<state::RouterState>) -> axum::Router {
+    axum::Router::new()
+        .merge(handlers::api_routes())
+        .layer(CorsLayer::permissive())
+        .layer(TraceLayer::new_for_http())
+        .with_state(state)
+}
+
+/// Start the router: build state from config and bind the plaintext HTTP
+/// listener. TLS is terminated by edge nginx ahead of this process.
+pub async fn run(config: RouterConfig) -> Result<()> {
+    let state = Arc::new(state::RouterState::from_config(&config));
+
+    // Background topology poller (#72): refresh each cortex's health +
+    // catalogue so routing decisions see live capacity.
+    let poller_state = Arc::clone(&state);
+    tokio::spawn(async move {
+        poller::poll_loop(poller_state).await;
+    });
+
+    let app = build_app(Arc::clone(&state));
+
+    let listen_addr = config.router.listen.parse::<std::net::SocketAddr>()?;
+    tracing::info!("helexa-router listening on {listen_addr}");
+
+    let listener = tokio::net::TcpListener::bind(listen_addr).await?;
+    axum::serve(listener, app).await?;
+
+    Ok(())
+}

crates/helexa-router/src/main.rs

@@ -0,0 +1,52 @@
+use anyhow::Result;
+use clap::{Parser, Subcommand};
+use helexa_router::config::RouterConfig;
+use tracing_subscriber::EnvFilter;
+
+#[derive(Parser)]
+#[command(name = "helexa-router")]
+#[command(about = "Public multi-operator ingress proxy for helexa")]
+#[command(version)]
+struct Cli {
+    #[command(subcommand)]
+    command: Commands,
+}
+
+#[derive(Subcommand)]
+enum Commands {
+    /// Start the router server.
+    Serve {
+        /// Path to the router config file.
+        #[arg(short, long, default_value = "helexa-router.toml")]
+        config: String,
+    },
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| EnvFilter::new("info,helexa_router=debug")),
+        )
+        .init();
+
+    let cli = Cli::parse();
+
+    match cli.command {
+        Commands::Serve { config } => {
+            let cfg = RouterConfig::load(&config)
+                .map_err(|e| anyhow::anyhow!("failed to load config from '{config}': {e}"))?;
+
+            tracing::info!(
+                cortexes = cfg.cortexes.len(),
+                listen = %cfg.router.listen,
+                "starting helexa-router"
+            );
+
+            helexa_router::run(cfg).await?;
+        }
+    }
+
+    Ok(())
+}

crates/helexa-router/src/poller.rs

@@ -0,0 +1,150 @@
+//! Background poller that refreshes the multi-operator topology (#72).
+//!
+//! The same pattern as cortex↔neuron, one tier up: periodically poll each
+//! configured cortex's `GET /v1/models` (catalogue × topology feasibility +
+//! loaded state) and `GET /health` (coarse node-health/load), building the
+//! live map the dispatcher (#73) routes on. An unreachable or erroring
+//! cortex is debounced over [`POLL_FAILURE_THRESHOLD`] consecutive misses,
+//! then flipped unhealthy and excluded from routing; it recovers on the
+//! next successful poll.
+
+use crate::state::RouterState;
+use chrono::Utc;
+use cortex_core::node::CortexModelEntry;
+use serde::Deserialize;
+use std::time::Duration;
+
+/// Per-cortex HTTP timeout for each poll request.
+const POLL_TIMEOUT: Duration = Duration::from_secs(5);
+
+/// Consecutive failed polls before a cortex is marked unreachable. Mirrors
+/// cortex's neuron-poll debounce: a single blip (a busy cortex briefly slow
+/// to answer) can't yank it — and all its models — out of routing.
+pub const POLL_FAILURE_THRESHOLD: u32 = 3;
+
+/// cortex's `/v1/models` envelope — `{ "object": "list", "data": [...] }`.
+#[derive(Debug, Deserialize)]
+struct ModelsEnvelope {
+    #[serde(default)]
+    data: Vec<CortexModelEntry>,
+}
+
+/// The subset of cortex's `/health` the router reads.
+#[derive(Debug, Deserialize)]
+struct CortexHealth {
+    nodes: CortexHealthNodes,
+}
+
+#[derive(Debug, Deserialize)]
+struct CortexHealthNodes {
+    healthy: u32,
+    total: u32,
+}
+
+/// Run forever, polling all cortexes on the configured interval.
+pub async fn poll_loop(state: std::sync::Arc<RouterState>) {
+    loop {
+        poll_once(&state).await;
+        tokio::time::sleep(state.poll_interval).await;
+    }
+}
+
+/// Poll every configured cortex once. Public for testing.
+pub async fn poll_once(state: &RouterState) {
+    for cortex in &state.cortexes {
+        poll_cortex(state, &cortex.name, &cortex.endpoint).await;
+    }
+}
+
+/// Poll one cortex: refresh its model map from `/v1/models`, then its node
+/// health from `/health`. A `/v1/models` failure debounces toward
+/// unreachable; the `/health` poll is best-effort and never flips
+/// reachability on its own (a cortex serving `/v1/models` is routable even
+/// if `/health` momentarily isn't).
+async fn poll_cortex(state: &RouterState, name: &str, endpoint: &str) {
+    // A cortex whose pinned TLS client failed to build (#74) is disabled:
+    // there is no client to poll with, so it stays unreachable.
+    let Some(client) = state.client_for(name) else {
+        let mut topo = state.topology.write().await;
+        if let Some(entry) = topo.get_mut(name) {
+            entry.consecutive_failures = entry.consecutive_failures.saturating_add(1);
+            entry.reachable = false;
+        }
+        tracing::warn!(cortex = name, "no TLS client (disabled); skipping poll");
+        return;
+    };
+
+    let models = fetch_models(client, endpoint).await;
+
+    let mut topo = state.topology.write().await;
+    let Some(entry) = topo.get_mut(name) else {
+        return; // not a configured cortex (shouldn't happen)
+    };
+
+    match models {
+        Ok(models) => {
+            entry.models = models.into_iter().map(|m| (m.id.clone(), m)).collect();
+            entry.reachable = true;
+            entry.consecutive_failures = 0;
+            entry.last_poll = Some(Utc::now());
+            tracing::debug!(cortex = name, models = entry.models.len(), "poll ok");
+        }
+        Err(reason) => {
+            entry.consecutive_failures = entry.consecutive_failures.saturating_add(1);
+            if entry.consecutive_failures >= POLL_FAILURE_THRESHOLD {
+                entry.reachable = false;
+            }
+            tracing::warn!(
+                cortex = name,
+                failures = entry.consecutive_failures,
+                reachable = entry.reachable,
+                reason,
+                "cortex poll failed"
+            );
+        }
+    }
+    drop(topo);
+
+    // Best-effort health (node counts). Never flips reachability.
+    if let Some((healthy, total)) = fetch_health(client, endpoint).await {
+        let mut topo = state.topology.write().await;
+        if let Some(entry) = topo.get_mut(name) {
+            entry.healthy_nodes = healthy;
+            entry.total_nodes = total;
+        }
+    }
+}
+
+/// GET `/v1/models`, returning the parsed entries or a short failure reason.
+async fn fetch_models(
+    client: &reqwest::Client,
+    endpoint: &str,
+) -> Result<Vec<CortexModelEntry>, &'static str> {
+    let url = format!("{endpoint}/v1/models");
+    let resp = client
+        .get(&url)
+        .timeout(POLL_TIMEOUT)
+        .send()
+        .await
+        .map_err(|_| "unreachable")?;
+    if !resp.status().is_success() {
+        return Err("non-success status");
+    }
+    let envelope = resp
+        .json::<ModelsEnvelope>()
+        .await
+        .map_err(|_| "bad json")?;
+    Ok(envelope.data)
+}
+
+/// GET `/health`, returning `(healthy, total)` node counts. `None` on any
+/// failure — the caller leaves the previous counts in place.
+async fn fetch_health(client: &reqwest::Client, endpoint: &str) -> Option<(u32, u32)> {
+    let url = format!("{endpoint}/health");
+    let resp = client.get(&url).timeout(POLL_TIMEOUT).send().await.ok()?;
+    if !resp.status().is_success() {
+        return None;
+    }
+    let health = resp.json::<CortexHealth>().await.ok()?;
+    Some((health.nodes.healthy, health.nodes.total))
+}

crates/helexa-router/src/state.rs

@@ -0,0 +1,144 @@
+use crate::config::{CortexEndpoint, RouterConfig};
+use chrono::{DateTime, Utc};
+use cortex_core::node::CortexModelEntry;
+use std::collections::HashMap;
+use std::time::Duration;
+use tokio::sync::RwLock;
+
+/// Shared router state: the configured cortex list plus the live topology
+/// map the poller (#72) maintains and the dispatcher (#73) will route on.
+///
+/// This is the router tier of the fractal neuron ← cortex ← router design:
+/// just as cortex polls each neuron for capacity/catalogue, the router
+/// polls each cortex's `/health` + `/v1/models`.
+#[derive(Debug)]
+pub struct RouterState {
+    /// Downstream cortex endpoints, as configured.
+    pub cortexes: Vec<CortexEndpoint>,
+    /// Per-cortex HTTP client, keyed by cortex name (#74). A cortex enrolled
+    /// with a `tls_ca` gets a client that trusts only that anchor; others
+    /// get a default client. A cortex whose `tls_ca` failed to load is
+    /// **absent** here — `client_for` returns `None` and it is never
+    /// polled or routed to (fail closed: a misconfigured pin must not
+    /// silently fall back to unpinned TLS).
+    clients: HashMap<String, reqwest::Client>,
+    /// This router instance's region, for dispatch geo affinity (#73).
+    pub region: Option<String>,
+    /// How often the poller refreshes the topology.
+    pub poll_interval: Duration,
+    /// Live per-cortex topology, keyed by cortex name. Pre-populated from
+    /// config (every configured cortex present, `reachable = false`) so the
+    /// poller and handlers always find an entry; the poller flips
+    /// reachability and fills the model map.
+    pub topology: RwLock<HashMap<String, CortexTopology>>,
+}
+
+/// Live view of one downstream cortex, refreshed each poll.
+#[derive(Debug, Clone, Default)]
+pub struct CortexTopology {
+    /// Whether the cortex is currently routable. Flipped `false` only after
+    /// [`crate::poller::POLL_FAILURE_THRESHOLD`] consecutive failed polls
+    /// (debounces transient blips); restored on the next successful poll.
+    pub reachable: bool,
+    /// Consecutive failed polls; reset to 0 on success.
+    pub consecutive_failures: u32,
+    /// Timestamp of the last successful poll.
+    pub last_poll: Option<DateTime<Utc>>,
+    /// Healthy / total neuron counts from the cortex's `/health` (coarse
+    /// load signal; #73 refines headroom). 0/0 until first health poll.
+    pub healthy_nodes: u32,
+    pub total_nodes: u32,
+    /// The cortex's full `/v1/models` entries, keyed by model id. Stored
+    /// whole (not distilled to a loaded/feasible bool) so the federation
+    /// catalogue (#75) can preserve per-model `limit`/`cost`/capabilities.
+    pub models: HashMap<String, CortexModelEntry>,
+}
+
+/// Whether a cortex can serve this model — loaded now, or feasible to
+/// cold-load (its catalogue × topology says some neuron can host it).
+pub fn entry_feasible(entry: &CortexModelEntry) -> bool {
+    entry.loaded || !entry.feasible_on.is_empty()
+}
+
+impl RouterState {
+    pub fn from_config(config: &RouterConfig) -> Self {
+        let topology = config
+            .cortexes
+            .iter()
+            .map(|c| (c.name.clone(), CortexTopology::default()))
+            .collect();
+
+        // One client per cortex. A `tls_ca` that fails to load omits the
+        // cortex from the map (fail closed) rather than degrading to an
+        // unpinned client.
+        let mut clients = HashMap::new();
+        for c in &config.cortexes {
+            match build_client(c.tls_ca.as_deref()) {
+                Ok(client) => {
+                    clients.insert(c.name.clone(), client);
+                }
+                Err(e) => {
+                    tracing::error!(
+                        cortex = %c.name,
+                        tls_ca = c.tls_ca.as_deref().unwrap_or(""),
+                        error = %e,
+                        "failed to build pinned TLS client; cortex disabled (fail closed)"
+                    );
+                }
+            }
+        }
+
+        Self {
+            cortexes: config.cortexes.clone(),
+            clients,
+            region: config.router.region.clone(),
+            poll_interval: Duration::from_secs(config.router.poll_interval_secs),
+            topology: RwLock::new(topology),
+        }
+    }
+
+    /// The HTTP client to use for `name`, or `None` if the cortex is
+    /// disabled (its `tls_ca` failed to load). Callers must treat `None` as
+    /// "not routable / not pollable".
+    pub fn client_for(&self, name: &str) -> Option<&reqwest::Client> {
+        self.clients.get(name)
+    }
+
+    /// Names of reachable cortexes that can serve `model_id` (loaded or
+    /// feasible to cold-load). Groundwork for capacity-aware dispatch (#73);
+    /// unreachable cortexes are excluded by construction.
+    pub async fn cortexes_serving(&self, model_id: &str) -> Vec<String> {
+        let topo = self.topology.read().await;
+        topo.iter()
+            .filter(|(_, t)| t.reachable)
+            .filter(|(_, t)| t.models.get(model_id).is_some_and(entry_feasible))
+            .map(|(name, _)| name.clone())
+            .collect()
+    }
+}
+
+/// Build a cortex HTTP client. With `tls_ca` set, the client trusts **only**
+/// that PEM anchor (platform roots disabled) — pinning the router→cortex hop
+/// to an enrolled cert (#74). Without it, standard platform-root validation.
+pub fn build_client(tls_ca: Option<&str>) -> Result<reqwest::Client, BuildClientError> {
+    let mut builder = reqwest::Client::builder();
+    if let Some(path) = tls_ca {
+        let pem = std::fs::read(path).map_err(|e| BuildClientError::Read(path.to_string(), e))?;
+        let cert = reqwest::Certificate::from_pem(&pem).map_err(BuildClientError::Parse)?;
+        builder = builder
+            .tls_built_in_root_certs(false)
+            .add_root_certificate(cert);
+    }
+    builder.build().map_err(BuildClientError::Build)
+}
+
+/// Why a cortex's pinned client could not be built (→ cortex disabled).
+#[derive(Debug, thiserror::Error)]
+pub enum BuildClientError {
+    #[error("reading TLS anchor '{0}'")]
+    Read(String, #[source] std::io::Error),
+    #[error("parsing TLS anchor PEM")]
+    Parse(#[source] reqwest::Error),
+    #[error("building HTTP client")]
+    Build(#[source] reqwest::Error),
+}

crates/helexa-router/tests/catalogue.rs

@@ -0,0 +1,132 @@
+//! End-to-end federation-catalogue test for #75: poll two mock cortexes
+//! that overlap on a model, then `GET /v1/models` on the router and verify
+//! the deduped union with merged availability and preserved limit/cost.
+
+use axum::Router;
+use axum::routing::get;
+use helexa_router::config::{CortexEndpoint, RouterConfig};
+use helexa_router::poller::poll_once;
+use helexa_router::state::RouterState;
+use serde_json::{Value, json};
+use std::sync::Arc;
+use tokio::net::TcpListener;
+
+/// Spawn a mock cortex serving the given `/v1/models` `data` array.
+async fn spawn_cortex(models: Value) -> String {
+    let models = Arc::new(models);
+    let app = Router::new()
+        .route(
+            "/v1/models",
+            get({
+                let models = Arc::clone(&models);
+                move || {
+                    let models = Arc::clone(&models);
+                    async move { axum::Json(json!({ "object": "list", "data": &*models })) }
+                }
+            }),
+        )
+        .route(
+            "/health",
+            get(|| async { axum::Json(json!({"status":"ok","nodes":{"healthy":1,"total":1}})) }),
+        );
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+/// Spawn the router (with poller) wired to the given cortex endpoints, and
+/// poll once synchronously so the topology is populated before we query.
+async fn spawn_router(cortexes: Vec<CortexEndpoint>) -> String {
+    let cfg = RouterConfig {
+        cortexes,
+        ..Default::default()
+    };
+    let state = Arc::new(RouterState::from_config(&cfg));
+    poll_once(&state).await; // deterministic: fill topology now
+
+    let app = helexa_router::build_app(Arc::clone(&state));
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+fn model(id: &str, loaded: bool, feasible_on: &[&str], ctx: u64, input_cost: f64) -> Value {
+    json!({
+        "id": id,
+        "object": "model",
+        "created": 0,
+        "owned_by": "helexa",
+        "loaded": loaded,
+        "feasible_on": feasible_on,
+        "locations": [],
+        "limit": { "context": ctx, "output": 4096 },
+        "cost": { "input": input_cost, "output": input_cost * 3.0 }
+    })
+}
+
+#[tokio::test]
+async fn federation_catalogue_dedupes_and_preserves_limit_cost() {
+    // cortex A: "shared" loaded (ctx 32768, $0.50) + "only-a" loaded.
+    let a = spawn_cortex(json!([
+        model("shared", true, &["beast"], 32_768, 0.50),
+        model("only-a", true, &["beast"], 8_192, 1.00),
+    ]))
+    .await;
+    // cortex B: "shared" cold-loadable, tighter ctx (16384), cheaper ($0.20).
+    let b = spawn_cortex(json!([model("shared", false, &["benjy"], 16_384, 0.20)])).await;
+
+    let router = spawn_router(vec![
+        CortexEndpoint {
+            name: "op-a".into(),
+            endpoint: a,
+            region: None,
+            tls_ca: None,
+        },
+        CortexEndpoint {
+            name: "op-b".into(),
+            endpoint: b,
+            region: None,
+            tls_ca: None,
+        },
+    ])
+    .await;
+
+    let body: Value = reqwest::get(format!("{router}/v1/models"))
+        .await
+        .unwrap()
+        .json()
+        .await
+        .unwrap();
+
+    assert_eq!(body["object"], "list");
+    let data = body["data"].as_array().unwrap();
+    // Deduped union: "shared" once + "only-a".
+    assert_eq!(data.len(), 2);
+
+    let shared = data.iter().find(|m| m["id"] == "shared").unwrap();
+    // Loaded somewhere (op-a) → loaded.
+    assert_eq!(shared["loaded"], true);
+    // feasible_on re-tiered to operator names, both present, sorted.
+    let feasible: Vec<&str> = shared["feasible_on"]
+        .as_array()
+        .unwrap()
+        .iter()
+        .map(|v| v.as_str().unwrap())
+        .collect();
+    assert_eq!(feasible, vec!["op-a", "op-b"]);
+    // Tightest limit (16384) and cheapest cost ($0.20) win.
+    assert_eq!(shared["limit"]["context"], 16_384);
+    assert_eq!(shared["cost"]["input"], 0.20);
+    // Loaded location named by operator, no neuron VRAM leaked.
+    let locs = shared["locations"].as_array().unwrap();
+    assert_eq!(locs.len(), 1);
+    assert_eq!(locs[0]["node"], "op-a");
+
+    assert!(data.iter().any(|m| m["id"] == "only-a"));
+}

crates/helexa-router/tests/dispatch.rs

@@ -0,0 +1,304 @@
+//! Capacity-aware dispatch acceptance tests for #73.
+//!
+//! Covers: a request routes to a cortex serving the model; the client's
+//! bearer reaches the cortex; cortex's #63 rejections pass through verbatim
+//! and are NOT retried away; transport failure fails over to another
+//! feasible cortex; unknown model → 404, no reachable capacity → 503; and
+//! the selection ranking (warm/region/headroom).
+
+use axum::body::Bytes;
+use axum::extract::State;
+use axum::http::{HeaderMap, StatusCode};
+use axum::response::{IntoResponse, Response};
+use axum::routing::post;
+use axum::{Json, Router};
+use cortex_core::node::CortexModelEntry;
+use helexa_router::config::{CortexEndpoint, RouterConfig};
+use helexa_router::dispatch::{Selection, dispatch, select_cortexes};
+use helexa_router::state::{CortexTopology, RouterState};
+use serde_json::{Value, json};
+use std::collections::HashMap;
+use tokio::net::TcpListener;
+
+/// A minimal `CortexModelEntry` for MODEL with the given serveability.
+fn model_entry(loaded: bool, feasible: bool) -> CortexModelEntry {
+    CortexModelEntry {
+        id: MODEL.into(),
+        object: "model".into(),
+        created: 0,
+        owned_by: "helexa".into(),
+        loaded,
+        feasible_on: if feasible || loaded {
+            vec!["n".into()]
+        } else {
+            vec![]
+        },
+        locations: vec![],
+        capabilities: vec![],
+        limit: None,
+        cost: None,
+        tool_call: false,
+        reasoning: false,
+        max_model_len: None,
+        max_input_tokens: None,
+        max_output_tokens: None,
+    }
+}
+
+const MODEL: &str = "Qwen/Qwen3-Coder-30B";
+
+// ── Mock cortex backend ──────────────────────────────────────────────
+
+/// Behaviour of a mock cortex, carried in axum State.
+#[derive(Clone)]
+struct MockCortex {
+    /// Identifies which cortex answered, echoed in the 200 body.
+    name: &'static str,
+    /// When true, return a genuine #63-shaped `429 rate_limit_exceeded`.
+    rate_limited: bool,
+}
+
+async fn mock_handler(State(m): State<MockCortex>, headers: HeaderMap) -> Response {
+    if m.rate_limited {
+        return (
+            StatusCode::TOO_MANY_REQUESTS,
+            Json(json!({"error":{"type":"rate_limit_error","code":"rate_limit_exceeded","message":"slow down","param":null}})),
+        )
+            .into_response();
+    }
+    let auth = headers
+        .get("authorization")
+        .and_then(|v| v.to_str().ok())
+        .unwrap_or("")
+        .to_string();
+    Json(json!({ "served_by": m.name, "auth_seen": auth })).into_response()
+}
+
+async fn spawn_cortex(mock: MockCortex) -> String {
+    let app = Router::new()
+        .route("/v1/chat/completions", post(mock_handler))
+        .with_state(mock);
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+fn ok_cortex(name: &'static str) -> MockCortex {
+    MockCortex {
+        name,
+        rate_limited: false,
+    }
+}
+
+// ── Helpers to build state with a hand-set topology ──────────────────
+
+fn state_with(cortexes: Vec<CortexEndpoint>, region: Option<String>) -> RouterState {
+    let cfg = RouterConfig {
+        cortexes,
+        ..Default::default()
+    };
+    let mut state = RouterState::from_config(&cfg);
+    state.region = region;
+    state
+}
+
+/// Overwrite the topology entry for `name` so tests control reachability and
+/// model serveability directly (no live poll).
+async fn set_topology(
+    state: &RouterState,
+    name: &str,
+    reachable: bool,
+    loaded: bool,
+    feasible: bool,
+    healthy_nodes: u32,
+) {
+    let mut topo = state.topology.write().await;
+    let mut models = HashMap::new();
+    models.insert(MODEL.to_string(), model_entry(loaded, feasible));
+    topo.insert(
+        name.to_string(),
+        CortexTopology {
+            reachable,
+            consecutive_failures: 0,
+            last_poll: None,
+            healthy_nodes,
+            total_nodes: healthy_nodes,
+            models,
+        },
+    );
+}
+
+fn ep(name: &str, endpoint: &str, region: Option<&str>) -> CortexEndpoint {
+    CortexEndpoint {
+        name: name.into(),
+        endpoint: endpoint.into(),
+        region: region.map(str::to_string),
+        tls_ca: None,
+    }
+}
+
+fn chat_body() -> Bytes {
+    Bytes::from(format!("{{\"model\":\"{MODEL}\",\"stream\":false}}"))
+}
+
+async fn body_json(resp: Response) -> (StatusCode, Value) {
+    let status = resp.status();
+    let bytes = axum::body::to_bytes(resp.into_body(), usize::MAX)
+        .await
+        .unwrap();
+    let v = serde_json::from_slice(&bytes).unwrap_or(Value::Null);
+    (status, v)
+}
+
+// ── Tests ────────────────────────────────────────────────────────────
+
+#[tokio::test]
+async fn routes_to_serving_cortex_and_forwards_bearer() {
+    let url = spawn_cortex(ok_cortex("c1")).await;
+    let state = state_with(vec![ep("c1", &url, None)], None);
+    set_topology(&state, "c1", true, true, true, 2).await;
+
+    let mut headers = HeaderMap::new();
+    headers.insert("authorization", "Bearer sk-test-123".parse().unwrap());
+
+    let resp = dispatch(&state, "/v1/chat/completions", headers, chat_body()).await;
+    let (status, body) = body_json(resp).await;
+
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["served_by"], "c1");
+    // Bearer reached the cortex unchanged.
+    assert_eq!(body["auth_seen"], "Bearer sk-test-123");
+}
+
+#[tokio::test]
+async fn cortex_429_passes_through_and_is_not_retried() {
+    // c1 (ranked first: loaded) returns a genuine 429; c2 would return 200.
+    let c1 = spawn_cortex(MockCortex {
+        name: "c1",
+        rate_limited: true,
+    })
+    .await;
+    let c2 = spawn_cortex(ok_cortex("c2")).await;
+    let state = state_with(vec![ep("c1", &c1, None), ep("c2", &c2, None)], None);
+    // Both reachable + loaded; c1 has more headroom so it ranks first.
+    set_topology(&state, "c1", true, true, true, 5).await;
+    set_topology(&state, "c2", true, true, true, 1).await;
+
+    let resp = dispatch(
+        &state,
+        "/v1/chat/completions",
+        HeaderMap::new(),
+        chat_body(),
+    )
+    .await;
+    let (status, body) = body_json(resp).await;
+
+    // The genuine 4xx is returned verbatim — NOT retried to c2.
+    assert_eq!(status, StatusCode::TOO_MANY_REQUESTS);
+    assert_eq!(body["error"]["code"], "rate_limit_exceeded");
+    assert!(body.get("served_by").is_none(), "must not have hit c2");
+}
+
+#[tokio::test]
+async fn fails_over_to_next_cortex_on_transport_error() {
+    // c_dead ranks first (more headroom) but its endpoint is a closed port;
+    // c_live is the fallback. The router must fail over and c_live serves.
+    let live = spawn_cortex(ok_cortex("c_live")).await;
+    let state = state_with(
+        vec![
+            ep("c_dead", "http://127.0.0.1:1", None),
+            ep("c_live", &live, None),
+        ],
+        None,
+    );
+    set_topology(&state, "c_dead", true, true, true, 9).await;
+    set_topology(&state, "c_live", true, true, true, 1).await;
+
+    let resp = dispatch(
+        &state,
+        "/v1/chat/completions",
+        HeaderMap::new(),
+        chat_body(),
+    )
+    .await;
+    let (status, body) = body_json(resp).await;
+
+    assert_eq!(status, StatusCode::OK);
+    assert_eq!(body["served_by"], "c_live");
+}
+
+#[tokio::test]
+async fn unknown_model_is_404() {
+    let state = state_with(vec![ep("c1", "http://127.0.0.1:1", None)], None);
+    // Topology has no entry for MODEL at all.
+    let resp = dispatch(
+        &state,
+        "/v1/chat/completions",
+        HeaderMap::new(),
+        chat_body(),
+    )
+    .await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::NOT_FOUND);
+    assert_eq!(body["error"]["code"], "model_not_found");
+}
+
+#[tokio::test]
+async fn known_but_all_unreachable_is_503() {
+    let state = state_with(vec![ep("c1", "http://127.0.0.1:1", None)], None);
+    // Cortex knows the model (from a prior good poll) but is now unreachable.
+    set_topology(&state, "c1", false, true, true, 2).await;
+    let resp = dispatch(
+        &state,
+        "/v1/chat/completions",
+        HeaderMap::new(),
+        chat_body(),
+    )
+    .await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::SERVICE_UNAVAILABLE);
+    assert_eq!(body["error"]["code"], "service_unavailable");
+}
+
+#[tokio::test]
+async fn missing_model_field_is_400() {
+    let state = state_with(vec![ep("c1", "http://127.0.0.1:1", None)], None);
+    let resp = dispatch(
+        &state,
+        "/v1/chat/completions",
+        HeaderMap::new(),
+        Bytes::from_static(b"{\"messages\":[]}"),
+    )
+    .await;
+    let (status, body) = body_json(resp).await;
+    assert_eq!(status, StatusCode::BAD_REQUEST);
+    assert_eq!(body["error"]["code"], "missing_model_field");
+}
+
+#[tokio::test]
+async fn ranking_prefers_loaded_then_region_then_headroom() {
+    // Router is in eu-west. Candidates:
+    //   warm-eu : loaded, region match, 1 node   → best
+    //   warm-us : loaded, no region,    9 nodes
+    //   cold-eu : feasible only, region match     → worst (cold)
+    let state = state_with(
+        vec![
+            ep("warm-eu", "http://127.0.0.1:1", Some("eu-west")),
+            ep("warm-us", "http://127.0.0.1:1", Some("us-east")),
+            ep("cold-eu", "http://127.0.0.1:1", Some("eu-west")),
+        ],
+        Some("eu-west".into()),
+    );
+    set_topology(&state, "warm-eu", true, true, true, 1).await;
+    set_topology(&state, "warm-us", true, true, true, 9).await;
+    set_topology(&state, "cold-eu", true, false, true, 5).await;
+
+    let Selection::Candidates(order) = select_cortexes(&state, MODEL).await else {
+        panic!("expected candidates");
+    };
+    let names: Vec<&str> = order.iter().map(|e| e.name.as_str()).collect();
+    assert_eq!(names, vec!["warm-eu", "warm-us", "cold-eu"]);
+}

crates/helexa-router/tests/skeleton.rs

@@ -0,0 +1,97 @@
+//! Skeleton acceptance tests for #70: the router builds, serves `/health`
+//! and `/v1/models` on a plaintext port, and loads its cortex-endpoint list
+//! from TOML with env overrides.
+
+use helexa_router::config::{CortexEndpoint, RouterConfig};
+use helexa_router::state::RouterState;
+use std::sync::Arc;
+use tokio::net::TcpListener;
+
+/// Bind the router app on an ephemeral port and return its base URL.
+async fn spawn_router(cortexes: Vec<CortexEndpoint>) -> String {
+    let cfg = RouterConfig {
+        cortexes,
+        ..Default::default()
+    };
+    let state = Arc::new(RouterState::from_config(&cfg));
+    let app = helexa_router::build_app(state);
+
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+
+    format!("http://{addr}")
+}
+
+#[tokio::test]
+async fn health_reports_configured_cortex_count() {
+    let base = spawn_router(vec![
+        CortexEndpoint {
+            name: "a".into(),
+            endpoint: "https://a.example.com".into(),
+            region: None,
+            tls_ca: None,
+        },
+        CortexEndpoint {
+            name: "b".into(),
+            endpoint: "https://b.example.com".into(),
+            region: None,
+            tls_ca: None,
+        },
+    ])
+    .await;
+
+    let body: serde_json::Value = reqwest::get(format!("{base}/health"))
+        .await
+        .unwrap()
+        .json()
+        .await
+        .unwrap();
+
+    assert_eq!(body["status"], "ok");
+    assert_eq!(body["cortexes"]["configured"], 2);
+}
+
+#[tokio::test]
+async fn models_returns_empty_openai_list() {
+    let base = spawn_router(vec![]).await;
+
+    let resp = reqwest::get(format!("{base}/v1/models")).await.unwrap();
+    assert!(resp.status().is_success());
+
+    let body: serde_json::Value = resp.json().await.unwrap();
+    assert_eq!(body["object"], "list");
+    assert_eq!(body["data"].as_array().unwrap().len(), 0);
+}
+
+#[test]
+#[allow(clippy::result_large_err)]
+fn config_loads_from_toml_with_env_override() {
+    figment::Jail::expect_with(|jail| {
+        jail.create_file(
+            "helexa-router.toml",
+            r#"
+[router]
+listen = "127.0.0.1:8088"
+
+[[cortexes]]
+name = "lair-cafe"
+endpoint = "https://cortex.lair.cafe"
+"#,
+        )?;
+
+        // Env override wins over the TOML value.
+        jail.set_env("HELEXA_ROUTER_ROUTER__LISTEN", "0.0.0.0:9099");
+
+        let cfg = RouterConfig::load("helexa-router.toml").expect("load config");
+
+        assert_eq!(cfg.router.listen, "0.0.0.0:9099");
+        assert_eq!(cfg.cortexes.len(), 1);
+        assert_eq!(cfg.cortexes[0].name, "lair-cafe");
+        assert_eq!(cfg.cortexes[0].endpoint, "https://cortex.lair.cafe");
+
+        Ok(())
+    });
+}

crates/helexa-router/tests/tls.rs

@@ -0,0 +1,210 @@
+//! Outbound TLS pinning tests for #74.
+//!
+//! Proves the router, as a TLS client to cortexes, reaches a cortex
+//! presenting its **enrolled** cert and rejects one presenting an
+//! unexpected (or untrusted) cert — and that a rejected handshake flows
+//! through the existing reachability path (#72) to exclude the cortex.
+//!
+//! A minimal `tokio-rustls` HTTPS server presents a self-signed cert; the
+//! router's `reqwest` client (native-tls) validates against the PEM anchor
+//! enrolled in config. Server (rustls) and client (native-tls) interoperate
+//! at the protocol level — what matters is the trust decision.
+
+use helexa_router::config::{CortexEndpoint, RouterConfig};
+use helexa_router::poller::poll_once;
+use helexa_router::state::{RouterState, build_client};
+use std::io::Write;
+use std::sync::Arc;
+use tokio::io::{AsyncReadExt, AsyncWriteExt};
+use tokio::net::TcpListener;
+use tokio_rustls::TlsAcceptor;
+
+/// A self-signed cert: PEM (for the reqwest pin file) + DER cert/key (for
+/// the rustls server).
+struct TestCert {
+    cert_pem: String,
+    cert_der: rustls::pki_types::CertificateDer<'static>,
+    key_der: Vec<u8>,
+}
+
+fn make_cert() -> TestCert {
+    let key = rcgen::generate_simple_self_signed(vec!["127.0.0.1".to_string()]).unwrap();
+    TestCert {
+        cert_pem: key.cert.pem(),
+        cert_der: key.cert.der().clone(),
+        key_der: key.key_pair.serialize_der(),
+    }
+}
+
+/// Write a cert PEM to a unique temp file (named by `tag`) and return the
+/// path. `tag` is caller-unique (we use the bound port), so no randomness.
+fn write_pem(tag: &str, pem: &str) -> String {
+    let path = std::env::temp_dir().join(format!("helexa-router-tls-{tag}.pem"));
+    let mut f = std::fs::File::create(&path).unwrap();
+    f.write_all(pem.as_bytes()).unwrap();
+    path.to_string_lossy().into_owned()
+}
+
+/// Spawn a minimal HTTPS server presenting `cert`, answering every request
+/// with a canned `/v1/models`-shaped 200. Returns its `https://` base URL.
+async fn spawn_https(cert: &TestCert) -> String {
+    let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
+
+    let key = rustls::pki_types::PrivateKeyDer::Pkcs8(rustls::pki_types::PrivatePkcs8KeyDer::from(
+        cert.key_der.clone(),
+    ));
+    let config = rustls::ServerConfig::builder()
+        .with_no_client_auth()
+        .with_single_cert(vec![cert.cert_der.clone()], key)
+        .unwrap();
+    let acceptor = TlsAcceptor::from(Arc::new(config));
+
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        loop {
+            let Ok((stream, _)) = listener.accept().await else {
+                continue;
+            };
+            let acceptor = acceptor.clone();
+            tokio::spawn(async move {
+                if let Ok(mut tls) = acceptor.accept(stream).await {
+                    let mut buf = [0u8; 2048];
+                    let _ = tls.read(&mut buf).await; // consume request line/headers
+                    let body = "{\"object\":\"list\",\"data\":[]}";
+                    let resp = format!(
+                        "HTTP/1.1 200 OK\r\ncontent-type: application/json\r\ncontent-length: {}\r\nconnection: close\r\n\r\n{}",
+                        body.len(),
+                        body
+                    );
+                    let _ = tls.write_all(resp.as_bytes()).await;
+                    let _ = tls.shutdown().await;
+                }
+            });
+        }
+    });
+    format!("https://{addr}")
+}
+
+fn tag_for(url: &str) -> String {
+    url.rsplit(':').next().unwrap_or("0").to_string()
+}
+
+#[tokio::test]
+async fn pinned_client_accepts_enrolled_cert_and_rejects_others() {
+    let server_cert = make_cert();
+    let other_cert = make_cert();
+    let url = spawn_https(&server_cert).await;
+    let tag = tag_for(&url);
+
+    let good_pin = write_pem(&format!("{tag}-good"), &server_cert.cert_pem);
+    let bad_pin = write_pem(&format!("{tag}-bad"), &other_cert.cert_pem);
+
+    // Enrolled with the server's own cert → handshake trusted → 200.
+    let good = build_client(Some(&good_pin)).unwrap();
+    let resp = good.get(format!("{url}/v1/models")).send().await;
+    assert!(resp.is_ok(), "enrolled cert must be accepted: {resp:?}");
+    assert_eq!(resp.unwrap().status(), 200);
+
+    // Enrolled with a different cert → server's cert is unexpected → reject.
+    let bad = build_client(Some(&bad_pin)).unwrap();
+    assert!(
+        bad.get(format!("{url}/v1/models")).send().await.is_err(),
+        "unexpected cert must be rejected"
+    );
+
+    // No enrollment (default platform roots) → self-signed cert untrusted.
+    let default = build_client(None).unwrap();
+    assert!(
+        default
+            .get(format!("{url}/v1/models"))
+            .send()
+            .await
+            .is_err(),
+        "un-enrolled self-signed cert must be rejected by default roots"
+    );
+}
+
+#[tokio::test]
+async fn poller_excludes_cortex_with_unexpected_cert() {
+    let server_cert = make_cert();
+    let other_cert = make_cert();
+    let url = spawn_https(&server_cert).await;
+    let tag = tag_for(&url);
+
+    let good_pin = write_pem(&format!("{tag}-pgood"), &server_cert.cert_pem);
+    let bad_pin = write_pem(&format!("{tag}-pbad"), &other_cert.cert_pem);
+
+    // Cortex A enrolled correctly → reachable. Cortex B enrolled with the
+    // wrong cert → TLS handshake fails → excluded.
+    let cfg = RouterConfig {
+        cortexes: vec![
+            CortexEndpoint {
+                name: "good".into(),
+                endpoint: url.clone(),
+                region: None,
+                tls_ca: Some(good_pin),
+            },
+            CortexEndpoint {
+                name: "bad".into(),
+                endpoint: url.clone(),
+                region: None,
+                tls_ca: Some(bad_pin),
+            },
+        ],
+        ..Default::default()
+    };
+    let state = RouterState::from_config(&cfg);
+    poll_once(&state).await;
+
+    let topo = state.topology.read().await;
+    assert!(
+        topo["good"].reachable,
+        "correctly-enrolled cortex reachable"
+    );
+    assert!(
+        !topo["bad"].reachable,
+        "cortex presenting an unexpected cert is excluded"
+    );
+}
+
+#[tokio::test]
+async fn misconfigured_pin_disables_cortex_fail_closed() {
+    // A `tls_ca` pointing at a nonexistent file must NOT fall back to an
+    // unpinned client — the cortex is disabled entirely.
+    let cfg = RouterConfig {
+        cortexes: vec![
+            CortexEndpoint {
+                name: "broken".into(),
+                endpoint: "https://127.0.0.1:1".into(),
+                region: None,
+                tls_ca: Some("/no/such/anchor.pem".into()),
+            },
+            CortexEndpoint {
+                name: "plain".into(),
+                endpoint: "http://127.0.0.1:1".into(),
+                region: None,
+                tls_ca: None,
+            },
+        ],
+        ..Default::default()
+    };
+    let state = RouterState::from_config(&cfg);
+    assert!(
+        state.client_for("broken").is_none(),
+        "a cortex with an unloadable pin is disabled (fail closed)"
+    );
+    assert!(
+        state.client_for("plain").is_some(),
+        "an un-pinned cortex still gets a client"
+    );
+}
+
+#[test]
+fn build_client_rejects_garbage_pem() {
+    let path = write_pem(
+        "garbage",
+        "-----BEGIN CERTIFICATE-----\nnope\n-----END CERTIFICATE-----",
+    );
+    assert!(build_client(Some(&path)).is_err());
+}

crates/helexa-router/tests/topology.rs

@@ -0,0 +1,172 @@
+//! Topology-poller acceptance tests for #72: the router maintains a live
+//! map of which cortexes serve which models, marks an unreachable/erroring
+//! cortex unhealthy and excludes it from routing, and recovers it once
+//! reachable again.
+
+use axum::extract::State;
+use axum::http::StatusCode;
+use axum::routing::get;
+use axum::{Json, Router};
+use helexa_router::config::{CortexEndpoint, RouterConfig};
+use helexa_router::poller::{POLL_FAILURE_THRESHOLD, poll_once};
+use helexa_router::state::{RouterState, entry_feasible};
+use serde_json::{Value, json};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use tokio::net::TcpListener;
+
+/// Shared "is this mock cortex up?" flag, toggled by tests to simulate
+/// outage and recovery.
+#[derive(Clone)]
+struct MockState {
+    up: Arc<AtomicBool>,
+}
+
+async fn mock_models(State(s): State<MockState>) -> Result<Json<Value>, StatusCode> {
+    if !s.up.load(Ordering::SeqCst) {
+        return Err(StatusCode::SERVICE_UNAVAILABLE);
+    }
+    Ok(Json(json!({
+        "object": "list",
+        "data": [
+            {
+                "id": "Qwen/Qwen3-Coder-30B",
+                "object": "model",
+                "created": 0,
+                "owned_by": "helexa",
+                "loaded": true,
+                "feasible_on": ["beast"],
+                "locations": [{"node": "beast", "status": "loaded", "vram_estimate_mb": 19000}]
+            },
+            {
+                "id": "Qwen/Qwen3-VL-8B",
+                "object": "model",
+                "created": 0,
+                "owned_by": "helexa",
+                "loaded": false,
+                "feasible_on": ["beast"],
+                "locations": []
+            }
+        ]
+    })))
+}
+
+async fn mock_health(State(s): State<MockState>) -> Result<Json<Value>, StatusCode> {
+    if !s.up.load(Ordering::SeqCst) {
+        return Err(StatusCode::SERVICE_UNAVAILABLE);
+    }
+    Ok(Json(json!({
+        "status": "ok",
+        "nodes": { "healthy": 2, "total": 3 }
+    })))
+}
+
+/// Spawn a mock cortex; returns (base_url, up_flag).
+async fn spawn_mock_cortex() -> (String, Arc<AtomicBool>) {
+    let up = Arc::new(AtomicBool::new(true));
+    let state = MockState { up: up.clone() };
+    let app = Router::new()
+        .route("/v1/models", get(mock_models))
+        .route("/health", get(mock_health))
+        .with_state(state);
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    (format!("http://{addr}"), up)
+}
+
+fn state_for(name: &str, endpoint: &str) -> RouterState {
+    let cfg = RouterConfig {
+        cortexes: vec![CortexEndpoint {
+            name: name.into(),
+            endpoint: endpoint.into(),
+            region: None,
+            tls_ca: None,
+        }],
+        ..Default::default()
+    };
+    RouterState::from_config(&cfg)
+}
+
+#[tokio::test]
+async fn poll_builds_live_topology() {
+    let (base, _up) = spawn_mock_cortex().await;
+    let state = state_for("c1", &base);
+
+    poll_once(&state).await;
+
+    let topo = state.topology.read().await;
+    let c1 = topo.get("c1").expect("cortex present");
+    assert!(c1.reachable, "should be reachable after a good poll");
+    assert_eq!(c1.consecutive_failures, 0);
+    assert!(c1.last_poll.is_some());
+    assert_eq!((c1.healthy_nodes, c1.total_nodes), (2, 3));
+
+    // Loaded model: loaded + feasible. Catalogue-only model: feasible only
+    // (not loaded, but feasible_on non-empty).
+    let coder = c1.models.get("Qwen/Qwen3-Coder-30B").unwrap();
+    assert!(coder.loaded && entry_feasible(coder));
+    let vl = c1.models.get("Qwen/Qwen3-VL-8B").unwrap();
+    assert!(!vl.loaded && entry_feasible(vl));
+    drop(topo);
+
+    // The routing helper sees both serveable models on the reachable cortex.
+    assert_eq!(
+        state.cortexes_serving("Qwen/Qwen3-VL-8B").await,
+        vec!["c1".to_string()]
+    );
+}
+
+#[tokio::test]
+async fn unreachable_cortex_excluded_then_recovers() {
+    let (base, up) = spawn_mock_cortex().await;
+    let state = state_for("c1", &base);
+
+    // Healthy first.
+    poll_once(&state).await;
+    assert!(state.topology.read().await["c1"].reachable);
+
+    // Take it down. The first failures debounce (stay reachable) until the
+    // threshold; only then is it excluded.
+    up.store(false, Ordering::SeqCst);
+    for i in 1..POLL_FAILURE_THRESHOLD {
+        poll_once(&state).await;
+        assert!(
+            state.topology.read().await["c1"].reachable,
+            "still reachable after {i} failure(s) (below threshold)"
+        );
+    }
+    poll_once(&state).await; // crosses the threshold
+    {
+        let topo = state.topology.read().await;
+        assert!(!topo["c1"].reachable, "excluded after threshold failures");
+        assert!(topo["c1"].consecutive_failures >= POLL_FAILURE_THRESHOLD);
+    }
+    // Excluded from routing.
+    assert!(
+        state
+            .cortexes_serving("Qwen/Qwen3-Coder-30B")
+            .await
+            .is_empty()
+    );
+
+    // Bring it back: the next successful poll restores it.
+    up.store(true, Ordering::SeqCst);
+    poll_once(&state).await;
+    let topo = state.topology.read().await;
+    assert!(topo["c1"].reachable, "recovered after a good poll");
+    assert_eq!(topo["c1"].consecutive_failures, 0);
+}
+
+#[tokio::test]
+async fn unconfigured_endpoint_is_unreachable() {
+    // Nothing listening on this port → polls fail; below threshold it stays
+    // at its initial unreachable state, and never panics.
+    let state = state_for("dead", "http://127.0.0.1:1");
+    poll_once(&state).await;
+    let topo = state.topology.read().await;
+    assert!(!topo["dead"].reachable);
+    assert_eq!(topo["dead"].consecutive_failures, 1);
+}

crates/helexa-stream/Cargo.toml

@@ -0,0 +1,21 @@
+[package]
+name = "helexa-stream"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[lib]
+name = "helexa_stream"
+path = "src/lib.rs"
+
+[dependencies]
+axum = { workspace = true }
+reqwest = { workspace = true }
+futures = { workspace = true }
+thiserror = { workspace = true }
+
+[dev-dependencies]
+tokio = { workspace = true }
+tokio-stream = { workspace = true }
+async-stream = "0.3"

crates/helexa-stream/src/lib.rs

@@ -0,0 +1,290 @@
+//! Shared streaming reverse-proxy mechanism (#71).
+//!
+//! cortex and helexa-router both need to proxy an OpenAI/Anthropic SSE
+//! response from a downstream backend **verbatim** — chunks forwarded as
+//! they arrive, never buffering the full body — while observing the bytes
+//! for metrics/metering. This crate owns that mechanism so there is one
+//! implementation, not one per tier.
+//!
+//! The split is mechanism vs policy:
+//!
+//! - **Mechanism (here):** [`forward_streaming`] POSTs to a backend and
+//!   streams the response body back through an [`ObservedStream`], which
+//!   feeds every chunk to a caller-supplied [`ChunkObserver`] and calls
+//!   [`ChunkObserver::finish`] exactly once on clean end-of-stream or on
+//!   drop (client disconnect mid-stream). [`BodyTail`] and
+//!   [`last_count_for`] are the reusable pieces an observer uses to pull
+//!   the trailing OpenAI `usage` object out of the streamed bytes.
+//! - **Policy (caller):** what to *do* with the observed bytes — which
+//!   metric names to emit, which labels, whether to settle a per-principal
+//!   reservation — lives in the consumer's `ChunkObserver` impl, not here.
+//!
+//! The proxy is status-agnostic: a non-2xx upstream response (e.g. a
+//! cortex `429 rate_limit_exceeded`) is streamed back with its status and
+//! headers intact, so honest backpressure reaches the client unchanged.
+//! Only a network failure or a malformed response build is an error.
+
+use axum::body::{Body, Bytes};
+use axum::http::{HeaderMap, StatusCode};
+use axum::response::Response;
+use futures::Stream;
+use futures::stream::BoxStream;
+use reqwest::Client;
+use std::pin::Pin;
+use std::task::{Context, Poll};
+
+/// Observes the bytes of a streamed proxy response without altering them.
+///
+/// `observe` is called for each forwarded chunk; `finish` is called
+/// exactly once — on clean end-of-stream or on drop — and implementations
+/// must be idempotent (the [`ObservedStream`] guards against a double call,
+/// but a `finish` that runs side effects should still self-guard).
+pub trait ChunkObserver: Send + Unpin + 'static {
+    /// A body chunk has been forwarded downstream. The slice is the exact
+    /// bytes the client receives.
+    fn observe(&mut self, chunk: &[u8]);
+
+    /// The stream has ended (cleanly or via client disconnect). Called once.
+    fn finish(&mut self);
+}
+
+/// A bounded accumulator for the tail of a streamed body.
+///
+/// The OpenAI `usage` object rides on the final SSE chunk (and sits at the
+/// end of a non-streaming JSON body), so retaining a generous tail is
+/// enough to recover token counts via [`last_count_for`]; the cap bounds
+/// memory on huge bodies. Appends are char-boundary-safe.
+#[derive(Debug)]
+pub struct BodyTail {
+    tail: String,
+    cap: usize,
+}
+
+impl BodyTail {
+    /// Create a tail retaining at most `cap` bytes.
+    pub fn new(cap: usize) -> Self {
+        Self {
+            tail: String::new(),
+            cap,
+        }
+    }
+
+    /// Append a chunk, trimming from the front past the cap. When trimming,
+    /// the newest half is kept (the usage object is always at the very end).
+    pub fn push(&mut self, chunk: &[u8]) {
+        self.tail.push_str(&String::from_utf8_lossy(chunk));
+        if self.tail.len() > self.cap {
+            let mut cut = self.tail.len() - self.cap / 2;
+            while !self.tail.is_char_boundary(cut) {
+                cut += 1;
+            }
+            self.tail.drain(..cut);
+        }
+    }
+
+    /// The retained tail text.
+    pub fn as_str(&self) -> &str {
+        &self.tail
+    }
+}
+
+/// Find the value of the LAST `"key": <integer>` occurrence in `tail`.
+///
+/// Pure and chunk-boundary-safe (the tail is contiguous appended text).
+/// The quoted-needle form means `completion_tokens` never matches
+/// `completion_tokens_details`, and taking the last occurrence means the
+/// final `usage` object wins even if content earlier in the stream echoed
+/// a usage-shaped string.
+pub fn last_count_for(tail: &str, key: &str) -> Option<u64> {
+    let needle = format!("\"{key}\"");
+    let mut result = None;
+    for (idx, _) in tail.match_indices(&needle) {
+        let rest = tail[idx + needle.len()..].trim_start();
+        let Some(rest) = rest.strip_prefix(':') else {
+            continue;
+        };
+        let rest = rest.trim_start();
+        let digits: &str = &rest[..rest
+            .char_indices()
+            .find(|(_, c)| !c.is_ascii_digit())
+            .map(|(i, _)| i)
+            .unwrap_or(rest.len())];
+        if let Ok(v) = digits.parse::<u64>() {
+            result = Some(v);
+        }
+    }
+    result
+}
+
+/// Error from [`forward_streaming`]. Distinguishes a network/transport
+/// failure reaching the backend from a failure assembling the downstream
+/// response. A non-2xx upstream *status* is not an error — it is streamed
+/// through verbatim.
+#[derive(Debug, thiserror::Error)]
+pub enum StreamError {
+    #[error("upstream request failed")]
+    Upstream(reqwest::Error),
+    #[error("failed to build response")]
+    ResponseBuild(String),
+}
+
+/// POST `body` to `url` and stream the response back verbatim through
+/// `observer`.
+///
+/// Request headers are forwarded except `host` / `content-length` (reqwest
+/// sets these). The returned [`Response`] carries the upstream status and
+/// headers unchanged — including non-2xx — with a body that streams the
+/// upstream bytes chunk-for-chunk, feeding each chunk to `observer`.
+pub async fn forward_streaming<O: ChunkObserver>(
+    client: &Client,
+    url: &str,
+    headers: HeaderMap,
+    body: Bytes,
+    observer: O,
+) -> Result<Response, StreamError> {
+    let mut req_builder = client.post(url).body(body);
+    for (key, value) in headers.iter() {
+        if key == "host" || key == "content-length" {
+            continue; // reqwest sets these
+        }
+        req_builder = req_builder.header(key, value);
+    }
+
+    let upstream = req_builder.send().await.map_err(StreamError::Upstream)?;
+
+    let status =
+        StatusCode::from_u16(upstream.status().as_u16()).unwrap_or(StatusCode::BAD_GATEWAY);
+    let resp_headers = upstream.headers().clone();
+
+    let stream = ObservedStream::new(Box::pin(upstream.bytes_stream()), observer);
+    let body = Body::from_stream(stream);
+
+    let mut response = Response::builder().status(status);
+    for (key, value) in resp_headers.iter() {
+        response = response.header(key, value);
+    }
+    response
+        .body(body)
+        .map_err(|e| StreamError::ResponseBuild(e.to_string()))
+}
+
+/// Pass-through stream wrapper that feeds a [`ChunkObserver`]. Forwards
+/// each chunk verbatim, calls `observe` per chunk, and `finish` once on
+/// clean end-of-stream; the `Drop` impl covers client disconnects.
+pub struct ObservedStream<O: ChunkObserver> {
+    inner: BoxStream<'static, Result<Bytes, reqwest::Error>>,
+    observer: O,
+    finished: bool,
+}
+
+impl<O: ChunkObserver> ObservedStream<O> {
+    /// Wrap a byte stream with an observer.
+    pub fn new(inner: BoxStream<'static, Result<Bytes, reqwest::Error>>, observer: O) -> Self {
+        Self {
+            inner,
+            observer,
+            finished: false,
+        }
+    }
+
+    fn finish(&mut self) {
+        if self.finished {
+            return;
+        }
+        self.finished = true;
+        self.observer.finish();
+    }
+}
+
+impl<O: ChunkObserver> Stream for ObservedStream<O> {
+    type Item = Result<Bytes, reqwest::Error>;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.get_mut();
+        match this.inner.as_mut().poll_next(cx) {
+            Poll::Ready(Some(Ok(chunk))) => {
+                this.observer.observe(&chunk);
+                Poll::Ready(Some(Ok(chunk)))
+            }
+            Poll::Ready(Some(Err(e))) => Poll::Ready(Some(Err(e))),
+            Poll::Ready(None) => {
+                this.finish();
+                Poll::Ready(None)
+            }
+            Poll::Pending => Poll::Pending,
+        }
+    }
+}
+
+impl<O: ChunkObserver> Drop for ObservedStream<O> {
+    fn drop(&mut self) {
+        self.finish();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn extracts_counts_from_final_sse_usage_chunk() {
+        let tail = concat!(
+            "data: {\"choices\":[{\"delta\":{\"content\":\"hi\"}}]}\n\n",
+            "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":225,",
+            "\"completion_tokens\":42,\"total_tokens\":267}}\n\n",
+            "data: [DONE]\n\n"
+        );
+        assert_eq!(last_count_for(tail, "prompt_tokens"), Some(225));
+        assert_eq!(last_count_for(tail, "completion_tokens"), Some(42));
+    }
+
+    #[test]
+    fn extracts_counts_from_non_streaming_body() {
+        let tail = "{\"choices\":[{\"message\":{\"content\":\"hi\"}}],\
+                    \"usage\":{\"prompt_tokens\": 12, \"completion_tokens\": 7}}";
+        assert_eq!(last_count_for(tail, "prompt_tokens"), Some(12));
+        assert_eq!(last_count_for(tail, "completion_tokens"), Some(7));
+    }
+
+    #[test]
+    fn ignores_details_variants_and_takes_last_occurrence() {
+        // completion_tokens_details must not shadow completion_tokens,
+        // and the LAST usage object wins (matters when content echoes
+        // a usage-shaped string earlier in the stream).
+        let tail = concat!(
+            "data: {\"usage\":{\"completion_tokens\":1}}\n\n",
+            "data: {\"usage\":{\"completion_tokens\":99,",
+            "\"completion_tokens_details\":{\"reasoning_tokens\":3}}}\n\n"
+        );
+        assert_eq!(last_count_for(tail, "completion_tokens"), Some(99));
+    }
+
+    #[test]
+    fn absent_keys_yield_none() {
+        assert_eq!(
+            last_count_for("data: [DONE]\n\n", "completion_tokens"),
+            None
+        );
+        assert_eq!(last_count_for("", "prompt_tokens"), None);
+        // key present but non-numeric value
+        assert_eq!(
+            last_count_for("\"completion_tokens\": null", "completion_tokens"),
+            None
+        );
+    }
+
+    #[test]
+    fn body_tail_retains_usage_after_cap_trim() {
+        // Cap small enough that the filler forces several front-trims, but
+        // (as in production, where cap ≫ the usage object) large enough that
+        // the trailing usage object survives the newest-half retention.
+        let mut tail = BodyTail::new(512);
+        for _ in 0..100 {
+            tail.push(b"data: {\"choices\":[{\"delta\":{\"content\":\"x\"}}]}\n\n");
+        }
+        assert!(tail.as_str().len() <= 512, "cap must bound the tail");
+        tail.push(b"data: {\"usage\":{\"prompt_tokens\":5,\"completion_tokens\":9}}\n\n");
+        assert_eq!(last_count_for(tail.as_str(), "prompt_tokens"), Some(5));
+        assert_eq!(last_count_for(tail.as_str(), "completion_tokens"), Some(9));
+    }
+}

crates/helexa-stream/tests/streaming.rs

@@ -0,0 +1,162 @@
+//! Integration tests for the shared streaming proxy (#71): proves a backend
+//! SSE response is forwarded chunk-for-chunk (no buffering), the observer
+//! sees every byte and finishes once, and non-2xx is streamed through with
+//! its status intact — the behaviours both cortex and helexa-router rely on.
+
+use axum::Router;
+use axum::body::Body;
+use axum::http::{HeaderMap, StatusCode};
+use axum::response::Response;
+use axum::routing::post;
+use helexa_stream::{BodyTail, ChunkObserver, forward_streaming, last_count_for};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+use tokio::net::TcpListener;
+
+/// Observer that records what it saw, for assertions.
+#[derive(Clone, Default)]
+struct RecordingObserver {
+    inner: Arc<Mutex<Recorded>>,
+}
+
+#[derive(Default)]
+struct Recorded {
+    chunks: usize,
+    finished: usize,
+    tail: String,
+}
+
+impl ChunkObserver for RecordingObserver {
+    fn observe(&mut self, chunk: &[u8]) {
+        let mut r = self.inner.lock().unwrap();
+        r.chunks += 1;
+        r.tail.push_str(&String::from_utf8_lossy(chunk));
+    }
+    fn finish(&mut self) {
+        self.inner.lock().unwrap().finished += 1;
+    }
+}
+
+/// Mock backend that streams 5 SSE chunks with 30ms gaps, then a usage
+/// chunk and `[DONE]`.
+async fn sse_handler() -> Response {
+    let chunks: Vec<&'static str> = vec![
+        "data: {\"choices\":[{\"delta\":{\"content\":\"a\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"b\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"c\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"d\"}}]}\n\n",
+        "data: {\"choices\":[{\"delta\":{\"content\":\"e\"}}]}\n\n",
+        "data: {\"choices\":[],\"usage\":{\"prompt_tokens\":11,\"completion_tokens\":5}}\n\n",
+        "data: [DONE]\n\n",
+    ];
+    let stream = async_stream::stream! {
+        for c in chunks {
+            tokio::time::sleep(Duration::from_millis(30)).await;
+            yield Ok::<_, std::io::Error>(axum::body::Bytes::from_static(c.as_bytes()));
+        }
+    };
+    Response::new(Body::from_stream(stream))
+}
+
+async fn rate_limited_handler() -> Response {
+    Response::builder()
+        .status(StatusCode::TOO_MANY_REQUESTS)
+        .body(Body::from("{\"error\":{\"type\":\"rate_limit_exceeded\"}}"))
+        .unwrap()
+}
+
+async fn spawn_backend(router: Router) -> String {
+    let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, router).await.unwrap();
+    });
+    format!("http://{addr}")
+}
+
+#[tokio::test]
+async fn streams_chunks_incrementally_and_observes_usage() {
+    let base = spawn_backend(Router::new().route("/v1/chat/completions", post(sse_handler))).await;
+    let observer = RecordingObserver::default();
+    let probe = observer.clone();
+
+    let client = reqwest::Client::new();
+    let resp = forward_streaming(
+        &client,
+        &format!("{base}/v1/chat/completions"),
+        HeaderMap::new(),
+        axum::body::Bytes::from_static(b"{\"model\":\"x\",\"stream\":true}"),
+        observer,
+    )
+    .await
+    .expect("forward ok");
+
+    assert_eq!(resp.status(), StatusCode::OK);
+
+    // Read the proxied body as a stream, timestamping arrivals.
+    let mut body = resp.into_body().into_data_stream();
+    let mut arrivals: Vec<Instant> = Vec::new();
+    let mut collected = String::new();
+    use futures::StreamExt;
+    while let Some(item) = body.next().await {
+        let bytes = item.unwrap();
+        arrivals.push(Instant::now());
+        collected.push_str(&String::from_utf8_lossy(&bytes));
+    }
+
+    // Incremental delivery: first and last chunk are meaningfully apart
+    // (5×30ms gaps), proving no full-response buffering.
+    let spread = *arrivals.last().unwrap() - arrivals[0];
+    assert!(
+        spread >= Duration::from_millis(100),
+        "expected incremental delivery, spread was {spread:?}"
+    );
+
+    // The client received the terminator and the usage object verbatim.
+    assert!(collected.contains("data: [DONE]"));
+
+    // The observer saw the bytes and finished exactly once.
+    let r = probe.inner.lock().unwrap();
+    assert!(r.chunks >= 5, "observer saw {} chunks", r.chunks);
+    assert_eq!(r.finished, 1, "finish must run exactly once");
+    assert_eq!(last_count_for(&r.tail, "prompt_tokens"), Some(11));
+    assert_eq!(last_count_for(&r.tail, "completion_tokens"), Some(5));
+}
+
+#[tokio::test]
+async fn non_2xx_is_streamed_through_verbatim() {
+    let base =
+        spawn_backend(Router::new().route("/v1/chat/completions", post(rate_limited_handler)))
+            .await;
+    let observer = RecordingObserver::default();
+    let probe = observer.clone();
+
+    let client = reqwest::Client::new();
+    let resp = forward_streaming(
+        &client,
+        &format!("{base}/v1/chat/completions"),
+        HeaderMap::new(),
+        axum::body::Bytes::new(),
+        observer,
+    )
+    .await
+    .expect("forward ok");
+
+    // Backpressure status reaches the client unchanged.
+    assert_eq!(resp.status(), StatusCode::TOO_MANY_REQUESTS);
+    let body = axum::body::to_bytes(resp.into_body(), usize::MAX)
+        .await
+        .unwrap();
+    assert!(String::from_utf8_lossy(&body).contains("rate_limit_exceeded"));
+
+    // finish still runs once even with a tiny non-streaming body.
+    assert_eq!(probe.inner.lock().unwrap().finished, 1);
+}
+
+#[test]
+fn body_tail_smoke() {
+    let mut tail = BodyTail::new(128);
+    tail.push(b"hello ");
+    tail.push(b"world");
+    assert_eq!(tail.as_str(), "hello world");
+}

crates/helexa-upstream/Cargo.toml

@@ -0,0 +1,63 @@
+[package]
+name = "helexa-upstream"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[[bin]]
+name = "helexa-upstream"
+path = "src/main.rs"
+
+[lib]
+name = "helexa_upstream"
+path = "src/lib.rs"
+
+[dependencies]
+tokio = { workspace = true }
+axum = { workspace = true }
+tower-http = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true }
+figment = { workspace = true }
+anyhow = { workspace = true }
+thiserror = { workspace = true }
+clap = { workspace = true }
+tracing = { workspace = true }
+tracing-subscriber = { workspace = true }
+chrono = { workspace = true }
+
+# PostgreSQL — the mesh authority's system of record. Runtime query API
+# (not the compile-time `query!` macros) so the crate builds in CI without a
+# live database or a committed `.sqlx` offline cache; correctness is covered
+# by the gated integration tests. (Macro adoption is a later refinement once
+# a dev DB + offline cache exist.)
+sqlx = { version = "0.8", default-features = false, features = [
+    "runtime-tokio",
+    "tls-rustls",
+    "postgres",
+    "macros",
+    "migrate",
+    "chrono",
+    "uuid",
+] }
+uuid = { version = "1", features = ["v4", "serde"] }
+sha2 = "0.10"
+subtle = "2.6"
+# Web auth (B4): argon2id password hashing, JWT sessions, CSPRNG secrets,
+# transactional email.
+argon2 = "0.5"
+jsonwebtoken = "9"
+rand = "0.8"
+lettre = { version = "0.11", default-features = false, features = [
+    "tokio1-rustls-tls",
+    "smtp-transport",
+    "builder",
+] }
+
+# cortex-core for the shared #63 OpenAiError envelope on the authz surface.
+cortex-core = { workspace = true }
+
+[dev-dependencies]
+figment = { workspace = true, features = ["test"] }
+reqwest = { workspace = true }

crates/helexa-upstream/migrations/0001_init.sql

@@ -0,0 +1,137 @@
+-- helexa-upstream initial schema (#59): accounts, keys, ledger, top-up
+-- codes, served-usage. The mesh-level authority's system of record.
+--
+-- Token amounts are BIGINT (i64) throughout; the cortex EntitlementProvider
+-- carries u64 but mesh allocations sit comfortably inside i64 and Postgres
+-- has no unsigned type.
+
+CREATE EXTENSION IF NOT EXISTS citext;
+CREATE EXTENSION IF NOT EXISTS pgcrypto;
+
+-- ── Users (web auth: email + password) ──────────────────────────────
+CREATE TABLE users (
+    id                       UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    email                    CITEXT NOT NULL UNIQUE,
+    password_hash            TEXT NOT NULL,                 -- argon2id PHC string
+    email_verified           BOOLEAN NOT NULL DEFAULT FALSE,
+    -- Browser fingerprint captured at registration (#abuse). Best-effort,
+    -- client-supplied; the primary signal for silent multi-account
+    -- detection. NULL when the client could not produce one.
+    registration_fingerprint TEXT,
+    created_at               TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+CREATE INDEX users_registration_fingerprint_idx
+    ON users (registration_fingerprint)
+    WHERE registration_fingerprint IS NOT NULL;
+
+-- Single-use email tokens for verification and password reset. Only the
+-- sha256 of the emailed secret is stored.
+CREATE TABLE email_tokens (
+    token_hash BYTEA PRIMARY KEY,
+    user_id    UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
+    kind       TEXT NOT NULL CHECK (kind IN ('verify', 'reset')),
+    expires_at TIMESTAMPTZ NOT NULL,
+    consumed_at TIMESTAMPTZ
+);
+CREATE INDEX email_tokens_user_idx ON email_tokens (user_id);
+
+-- ── Accounts (the billable allocation ledger) ───────────────────────
+CREATE TABLE accounts (
+    id                 UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    owner_user_id      UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
+    allocation_total   BIGINT NOT NULL DEFAULT 0,
+    allocation_spent   BIGINT NOT NULL DEFAULT 0,
+    allocation_reserved BIGINT NOT NULL DEFAULT 0,
+    -- 'deactivated' is the SILENT abuse flag: keys stop authorizing but no
+    -- surface ever tells the user why (see resolve → 401).
+    status             TEXT NOT NULL DEFAULT 'active'
+                       CHECK (status IN ('active', 'deactivated')),
+    -- This account shares a registration fingerprint with >= 1 other.
+    fingerprint_flagged BOOLEAN NOT NULL DEFAULT FALSE,
+    created_at         TIMESTAMPTZ NOT NULL DEFAULT now(),
+    -- The no-overshoot backstop to the atomic reserve UPDATE.
+    CONSTRAINT accounts_no_overshoot
+        CHECK (allocation_spent + allocation_reserved <= allocation_total),
+    CONSTRAINT accounts_nonneg
+        CHECK (allocation_spent >= 0 AND allocation_reserved >= 0)
+);
+CREATE INDEX accounts_owner_idx ON accounts (owner_user_id);
+
+-- ── API keys (Principal.key_id = api_keys.id) ───────────────────────
+CREATE TABLE api_keys (
+    id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+    account_id  UUID NOT NULL REFERENCES accounts(id) ON DELETE CASCADE,
+    key_hash    BYTEA NOT NULL,                 -- sha256(raw key)
+    key_prefix  TEXT NOT NULL,                  -- non-secret display prefix
+    label       TEXT NOT NULL DEFAULT '',
+    status      TEXT NOT NULL DEFAULT 'active'
+                CHECK (status IN ('active', 'archived')),
+    -- Per-key sub-cap: 'hardcap' = absolute tokens; 'percent' = % of the
+    -- account's allocation_total (resolved to an absolute at reserve time).
+    limit_kind  TEXT NOT NULL DEFAULT 'percent'
+                CHECK (limit_kind IN ('percent', 'hardcap')),
+    limit_value BIGINT NOT NULL DEFAULT 100,
+    -- serde of cortex_core::entitlements::CapWindow (Balance | Rolling).
+    cap_window  JSONB NOT NULL DEFAULT '{"kind":"balance"}'::jsonb,
+    -- Per-key running ledger (mirrors the account ledger; Balance semantics
+    -- in this migration — rolling-window reset lands with the authz API).
+    key_spent    BIGINT NOT NULL DEFAULT 0,
+    key_reserved BIGINT NOT NULL DEFAULT 0,
+    created_at  TIMESTAMPTZ NOT NULL DEFAULT now(),
+    CONSTRAINT api_keys_key_nonneg
+        CHECK (key_spent >= 0 AND key_reserved >= 0)
+);
+-- A raw key resolves only while active; the hash is unique among active keys.
+CREATE UNIQUE INDEX api_keys_active_hash_idx
+    ON api_keys (key_hash) WHERE status = 'active';
+CREATE INDEX api_keys_account_idx ON api_keys (account_id);
+
+-- ── Reservations (reserve → settle/release) ─────────────────────────
+-- id is BIGSERIAL so it maps to the cortex Reservation.id (u64) verbatim,
+-- with the Postgres sequence as the sole global authority.
+CREATE TABLE reservations (
+    id          BIGSERIAL PRIMARY KEY,
+    account_id  UUID NOT NULL REFERENCES accounts(id) ON DELETE CASCADE,
+    key_id      UUID NOT NULL REFERENCES api_keys(id) ON DELETE CASCADE,
+    reserved    BIGINT NOT NULL,
+    actual      BIGINT,
+    state       TEXT NOT NULL DEFAULT 'open'
+                CHECK (state IN ('open', 'settled', 'released')),
+    created_at  TIMESTAMPTZ NOT NULL DEFAULT now(),
+    settled_at  TIMESTAMPTZ
+);
+-- The sweeper scans open reservations by age.
+CREATE INDEX reservations_open_idx
+    ON reservations (created_at) WHERE state = 'open';
+
+-- ── Top-up codes (hybrid allocation) ────────────────────────────────
+CREATE TABLE top_up_codes (
+    code_hash   BYTEA PRIMARY KEY,             -- sha256(raw code)
+    value       BIGINT NOT NULL,              -- tokens this code grants
+    denomination TEXT,                         -- human label (e.g. "small")
+    created_at  TIMESTAMPTZ NOT NULL DEFAULT now(),
+    redeemed_by UUID REFERENCES accounts(id) ON DELETE SET NULL,
+    redeemed_at TIMESTAMPTZ
+);
+
+-- ── Served-usage ledger (#58 reconciliation) ────────────────────────
+-- Absolute per-(operator, account, key, period) served tokens, upserted by
+-- each cortex; reconciliation rolls these up for operator compensation.
+CREATE TABLE served_usage (
+    operator_id   TEXT NOT NULL,
+    account_id    UUID NOT NULL,
+    key_id        UUID NOT NULL,
+    period        DATE NOT NULL,
+    served_tokens BIGINT NOT NULL DEFAULT 0,
+    reconciled_at TIMESTAMPTZ,
+    PRIMARY KEY (operator_id, account_id, key_id, period)
+);
+
+-- ── Web sessions (DB-backed; alt/complement to stateless JWT) ───────
+CREATE TABLE sessions (
+    token_hash BYTEA PRIMARY KEY,             -- sha256(session token)
+    user_id    UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
+    expires_at TIMESTAMPTZ NOT NULL,
+    created_at TIMESTAMPTZ NOT NULL DEFAULT now()
+);
+CREATE INDEX sessions_user_idx ON sessions (user_id);

crates/helexa-upstream/src/authz.rs

@@ -0,0 +1,340 @@
+//! `/authz/v1` — the machine surface cortex's `UpstreamEntitlementProvider`
+//! (#57) consumes. It mirrors the `cortex_core::entitlements::EntitlementProvider`
+//! trait 1:1 (resolve / reserve / settle / release / snapshot) over the B1
+//! ledger.
+//!
+//! Contract notes for the cortex client:
+//! - A **non-2xx** response means the authority could not give an
+//!   authoritative answer (bad caller auth, malformed request, server
+//!   error) → the client should **fail closed**.
+//! - `reserve` returns **200** whether granted or budget-refused: the body
+//!   carries either `reservation_id` or a `rejected` discriminant. A budget
+//!   refusal is an authoritative answer, not a transport failure.
+//! - Rejections that are genuinely auth failures use the #63 `OpenAiError`
+//!   envelope so they can be surfaced verbatim.
+
+use crate::crypto::sha256;
+use crate::error::envelope_response;
+use crate::ledger::{self, LedgerError};
+use crate::state::AppState;
+use axum::extract::{Request, State};
+use axum::http::{StatusCode, header};
+use axum::middleware::Next;
+use axum::response::{IntoResponse, Response};
+use axum::routing::post;
+use axum::{Extension, Json, Router};
+use cortex_core::error_envelope::OpenAiError;
+use serde::{Deserialize, Serialize};
+use subtle::ConstantTimeEq;
+use uuid::Uuid;
+
+/// The operator a validated client bearer identifies (served-usage
+/// attribution, #58). Inserted into request extensions by [`client_auth`].
+#[derive(Debug, Clone)]
+pub struct OperatorId(pub String);
+
+/// Build the `/authz/v1` router with the client-auth layer applied.
+pub fn router(state: &AppState) -> Router<AppState> {
+    Router::new()
+        .route("/authz/v1/resolve", post(resolve))
+        .route("/authz/v1/reserve", post(reserve))
+        .route("/authz/v1/settle", post(settle))
+        .route("/authz/v1/release", post(release))
+        .route("/authz/v1/snapshot", post(snapshot))
+        .route("/authz/v1/served-usage", post(served_usage))
+        .layer(axum::middleware::from_fn_with_state(
+            state.clone(),
+            client_auth,
+        ))
+}
+
+// ── client auth (shared bearer → operator_id) ───────────────────────
+
+/// Validate the caller's `Authorization: Bearer` against the configured
+/// client tokens (constant-time) and stamp the `operator_id`. When no tokens
+/// are configured the surface is open (dev) and a synthetic operator is
+/// used.
+async fn client_auth(State(state): State<AppState>, mut req: Request, next: Next) -> Response {
+    let tokens = &state.config.client_auth.tokens;
+    if tokens.is_empty() {
+        req.extensions_mut().insert(OperatorId("dev".into()));
+        return next.run(req).await;
+    }
+    let presented = req
+        .headers()
+        .get(header::AUTHORIZATION)
+        .and_then(|v| v.to_str().ok())
+        .and_then(|v| v.strip_prefix("Bearer "))
+        .map(str::trim)
+        .unwrap_or("");
+    let matched = tokens
+        .iter()
+        .find(|t| t.token.as_bytes().ct_eq(presented.as_bytes()).into());
+    match matched {
+        Some(t) => {
+            req.extensions_mut()
+                .insert(OperatorId(t.operator_id.clone()));
+            next.run(req).await
+        }
+        None => envelope_response(OpenAiError::invalid_api_key(
+            "missing or invalid client credentials",
+        )),
+    }
+}
+
+// ── DTOs ────────────────────────────────────────────────────────────
+
+#[derive(Deserialize)]
+struct ResolveReq {
+    api_key: String,
+}
+
+#[derive(Serialize)]
+struct PrincipalDto {
+    account_id: String,
+    key_id: String,
+}
+
+#[derive(Serialize)]
+struct SnapshotDto {
+    hard_cap: Option<i64>,
+    spent: i64,
+    reserved: i64,
+}
+
+#[derive(Serialize)]
+struct ResolveResp {
+    principal: PrincipalDto,
+    snapshot: SnapshotDto,
+}
+
+#[derive(Deserialize)]
+struct ReserveReq {
+    account_id: String,
+    key_id: String,
+    max_tokens: i64,
+}
+
+#[derive(Serialize, Default)]
+struct ReserveResp {
+    #[serde(skip_serializing_if = "Option::is_none")]
+    reservation_id: Option<i64>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    rejected: Option<Rejection>,
+}
+
+#[derive(Serialize)]
+#[serde(tag = "kind", rename_all = "snake_case")]
+enum Rejection {
+    InsufficientQuota {
+        requested: i64,
+        available: i64,
+    },
+    // Part of the frozen wire contract so the cortex client (#57) can map it
+    // without a later breaking change. Not yet constructed: the B1 ledger
+    // implements Balance caps only; rolling-window key sub-caps (which yield
+    // this) land in a follow-up.
+    #[allow(dead_code)]
+    RateLimited {
+        requested: i64,
+        available: i64,
+        retry_after_secs: u64,
+    },
+}
+
+#[derive(Deserialize)]
+struct SettleReq {
+    reservation_id: i64,
+    actual_tokens: i64,
+}
+
+#[derive(Deserialize)]
+struct ReservationRef {
+    reservation_id: i64,
+}
+
+#[derive(Deserialize)]
+struct SnapshotReq {
+    account_id: String,
+    key_id: String,
+}
+
+// ── handlers ────────────────────────────────────────────────────────
+
+/// `POST /authz/v1/resolve` — bearer key → principal + snapshot, or
+/// `401 invalid_api_key` (also for a deactivated account: no clue).
+async fn resolve(State(state): State<AppState>, Json(req): Json<ResolveReq>) -> Response {
+    match ledger::resolve_key(&state.pool, &sha256(&req.api_key)).await {
+        Ok(Some(p)) => Json(ResolveResp {
+            principal: PrincipalDto {
+                account_id: p.account_id.to_string(),
+                key_id: p.key_id.to_string(),
+            },
+            snapshot: SnapshotDto {
+                hard_cap: Some(p.hard_cap),
+                spent: p.key_spent,
+                reserved: p.key_reserved,
+            },
+        })
+        .into_response(),
+        Ok(None) => envelope_response(OpenAiError::invalid_api_key("invalid or unknown API key")),
+        Err(e) => {
+            tracing::error!(error = %e, "resolve query failed");
+            envelope_response(OpenAiError::service_unavailable("authority error", Some(5)))
+        }
+    }
+}
+
+/// `POST /authz/v1/reserve` — 200 with `reservation_id` (granted) or
+/// `rejected` (budget). Non-2xx only for bad input / server error.
+async fn reserve(State(state): State<AppState>, Json(req): Json<ReserveReq>) -> Response {
+    let (Ok(account_id), Ok(key_id)) = (
+        Uuid::parse_str(&req.account_id),
+        Uuid::parse_str(&req.key_id),
+    ) else {
+        return bad_request("account_id and key_id must be UUIDs");
+    };
+    match ledger::reserve(&state.pool, account_id, key_id, req.max_tokens).await {
+        Ok(reservation_id) => Json(ReserveResp {
+            reservation_id: Some(reservation_id),
+            rejected: None,
+        })
+        .into_response(),
+        Err(LedgerError::InsufficientQuota {
+            requested,
+            available,
+        }) => Json(ReserveResp {
+            reservation_id: None,
+            rejected: Some(Rejection::InsufficientQuota {
+                requested,
+                available,
+            }),
+        })
+        .into_response(),
+        Err(LedgerError::AccountNotFound | LedgerError::KeyNotFound) => {
+            // Resolve succeeded earlier; the principal vanished (archived /
+            // deactivated). Treat as no budget — fail closed at the client.
+            Json(ReserveResp {
+                reservation_id: None,
+                rejected: Some(Rejection::InsufficientQuota {
+                    requested: req.max_tokens,
+                    available: 0,
+                }),
+            })
+            .into_response()
+        }
+        Err(LedgerError::Db(e)) => {
+            tracing::error!(error = %e, "reserve failed");
+            envelope_response(OpenAiError::service_unavailable("authority error", Some(5)))
+        }
+    }
+}
+
+/// `POST /authz/v1/settle` — idempotent; `204`.
+async fn settle(State(state): State<AppState>, Json(req): Json<SettleReq>) -> Response {
+    match ledger::settle(&state.pool, req.reservation_id, req.actual_tokens).await {
+        Ok(()) => StatusCode::NO_CONTENT.into_response(),
+        Err(e) => {
+            tracing::error!(error = %e, "settle failed");
+            envelope_response(OpenAiError::service_unavailable("authority error", Some(5)))
+        }
+    }
+}
+
+/// `POST /authz/v1/release` — idempotent; `204`.
+async fn release(State(state): State<AppState>, Json(req): Json<ReservationRef>) -> Response {
+    match ledger::release(&state.pool, req.reservation_id).await {
+        Ok(()) => StatusCode::NO_CONTENT.into_response(),
+        Err(e) => {
+            tracing::error!(error = %e, "release failed");
+            envelope_response(OpenAiError::service_unavailable("authority error", Some(5)))
+        }
+    }
+}
+
+/// `POST /authz/v1/snapshot` — `{hard_cap, spent, reserved}` or `404`.
+async fn snapshot(State(state): State<AppState>, Json(req): Json<SnapshotReq>) -> Response {
+    let (Ok(account_id), Ok(key_id)) = (
+        Uuid::parse_str(&req.account_id),
+        Uuid::parse_str(&req.key_id),
+    ) else {
+        return bad_request("account_id and key_id must be UUIDs");
+    };
+    match ledger::snapshot(&state.pool, account_id, key_id).await {
+        Ok(Some((hard_cap, spent, reserved))) => Json(SnapshotDto {
+            hard_cap: Some(hard_cap),
+            spent,
+            reserved,
+        })
+        .into_response(),
+        Ok(None) => StatusCode::NOT_FOUND.into_response(),
+        Err(e) => {
+            tracing::error!(error = %e, "snapshot failed");
+            envelope_response(OpenAiError::service_unavailable("authority error", Some(5)))
+        }
+    }
+}
+
+// ── served-usage report (#58) ───────────────────────────────────────
+
+#[derive(Deserialize)]
+struct ServedUsageReport {
+    rows: Vec<ServedUsageRow>,
+}
+
+#[derive(Deserialize)]
+struct ServedUsageRow {
+    account_id: String,
+    key_id: String,
+    period: String, // YYYY-MM-DD
+    served_tokens: i64,
+}
+
+/// `POST /authz/v1/served-usage` — a cortex reports the absolute served-token
+/// counters it has accrued for the current period. Upsert is monotonic
+/// (`GREATEST`) so re-sends and races are idempotent and never regress.
+/// `operator_id` comes from the validated client bearer (request extension).
+async fn served_usage(
+    State(state): State<AppState>,
+    Extension(operator): Extension<OperatorId>,
+    Json(req): Json<ServedUsageReport>,
+) -> Response {
+    for row in &req.rows {
+        let (Ok(account_id), Ok(key_id)) = (
+            Uuid::parse_str(&row.account_id),
+            Uuid::parse_str(&row.key_id),
+        ) else {
+            continue; // skip malformed ids rather than fail the whole batch
+        };
+        let Ok(period) = chrono::NaiveDate::parse_from_str(&row.period, "%Y-%m-%d") else {
+            continue;
+        };
+        let res = sqlx::query(
+            "INSERT INTO served_usage (operator_id, account_id, key_id, period, served_tokens) \
+             VALUES ($1, $2, $3, $4, $5) \
+             ON CONFLICT (operator_id, account_id, key_id, period) \
+             DO UPDATE SET served_tokens = GREATEST(served_usage.served_tokens, EXCLUDED.served_tokens)",
+        )
+        .bind(&operator.0)
+        .bind(account_id)
+        .bind(key_id)
+        .bind(period)
+        .bind(row.served_tokens.max(0))
+        .execute(&state.pool)
+        .await;
+        if let Err(e) = res {
+            tracing::error!(error = %e, "served-usage upsert failed");
+            return envelope_response(OpenAiError::service_unavailable("authority error", Some(5)));
+        }
+    }
+    StatusCode::NO_CONTENT.into_response()
+}
+
+fn bad_request(msg: &str) -> Response {
+    envelope_response(OpenAiError::new(
+        400,
+        "invalid_request_error",
+        "invalid_request",
+        msg,
+    ))
+}

crates/helexa-upstream/src/config.rs

@@ -0,0 +1,261 @@
+//! helexa-upstream configuration: loaded from `helexa-upstream.toml` with
+//! figment, `UPSTREAM_`-prefixed env overrides (mirrors the cortex/router
+//! convention, e.g. `UPSTREAM_SERVER__LISTEN`, `UPSTREAM_DB__URL`).
+
+use figment::{
+    Figment,
+    providers::{Env, Format, Toml},
+};
+use serde::{Deserialize, Serialize};
+use std::path::Path;
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct UpstreamConfig {
+    #[serde(default)]
+    pub server: ServerSettings,
+    pub db: DbSettings,
+    #[serde(default)]
+    pub grant: GrantSettings,
+    #[serde(default)]
+    pub abuse: AbuseSettings,
+    #[serde(default)]
+    pub client_auth: ClientAuthSettings,
+    #[serde(default)]
+    pub authz: AuthzSettings,
+    #[serde(default)]
+    pub auth: AuthSettings,
+    #[serde(default)]
+    pub email: EmailSettings,
+}
+
+/// `[auth]` — web-session signing + token lifetimes (B4). Web sessions are
+/// JWTs, distinct from inference API keys.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AuthSettings {
+    /// HMAC secret for signing session JWTs. MUST be overridden in prod
+    /// (env `UPSTREAM_AUTH__JWT_SECRET`); the default is dev-only.
+    #[serde(default = "default_jwt_secret")]
+    pub jwt_secret: String,
+    /// Session token lifetime (seconds).
+    #[serde(default = "default_session_ttl")]
+    pub session_ttl_secs: u64,
+    /// Email verification / password-reset token lifetime (seconds).
+    #[serde(default = "default_email_token_ttl")]
+    pub email_token_ttl_secs: u64,
+    /// Public base URL of the frontend, used to build verify/reset links.
+    #[serde(default = "default_app_base_url")]
+    pub app_base_url: String,
+}
+
+impl Default for AuthSettings {
+    fn default() -> Self {
+        Self {
+            jwt_secret: default_jwt_secret(),
+            session_ttl_secs: default_session_ttl(),
+            email_token_ttl_secs: default_email_token_ttl(),
+            app_base_url: default_app_base_url(),
+        }
+    }
+}
+
+/// `[email]` — transactional email transport for verify/reset.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct EmailSettings {
+    /// `"log"` (dev — logs the link) or `"smtp"`.
+    #[serde(default = "default_email_provider")]
+    pub provider: String,
+    /// SMTP relay URL (e.g. "smtp://user:pass@host:587") when provider=smtp.
+    #[serde(default)]
+    pub smtp_url: Option<String>,
+    /// `From:` address.
+    #[serde(default = "default_from_addr")]
+    pub from_addr: String,
+}
+
+impl Default for EmailSettings {
+    fn default() -> Self {
+        Self {
+            provider: default_email_provider(),
+            smtp_url: None,
+            from_addr: default_from_addr(),
+        }
+    }
+}
+
+/// `[client_auth]` — credentials operators' cortexes present to `/authz/v1`.
+/// Each token maps to an `operator_id` (served-usage attribution, #58). This
+/// transport credential is distinct from end-user API keys (which ride in
+/// the `resolve` body). v2 adds mTLS.
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub struct ClientAuthSettings {
+    /// When empty the authz surface is **open** (dev only; logged at warn).
+    #[serde(default)]
+    pub tokens: Vec<ClientToken>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ClientToken {
+    /// Shared bearer a cortex presents.
+    pub token: String,
+    /// Operator this token identifies.
+    pub operator_id: String,
+}
+
+/// `[authz]` — reservation lifecycle knobs.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AuthzSettings {
+    /// Open reservations older than this are swept (released), self-healing
+    /// a reservation whose settle/release from cortex was lost.
+    #[serde(default = "default_reservation_ttl")]
+    pub reservation_ttl_secs: u64,
+    /// How often the sweeper runs.
+    #[serde(default = "default_sweep_interval")]
+    pub sweep_interval_secs: u64,
+}
+
+impl Default for AuthzSettings {
+    fn default() -> Self {
+        Self {
+            reservation_ttl_secs: default_reservation_ttl(),
+            sweep_interval_secs: default_sweep_interval(),
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ServerSettings {
+    /// Address to listen on (e.g. "0.0.0.0:8090"). Plaintext — edge nginx
+    /// terminates TLS, consistent with the rest of the stack.
+    #[serde(default = "default_listen")]
+    pub listen: String,
+}
+
+impl Default for ServerSettings {
+    fn default() -> Self {
+        Self {
+            listen: default_listen(),
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct DbSettings {
+    /// PostgreSQL connection URL (e.g. "postgres://user:pass@host/helexa").
+    pub url: String,
+    /// Max pool connections.
+    #[serde(default = "default_max_connections")]
+    pub max_connections: u32,
+}
+
+/// `[grant]` — the flat free token grant every email-verified account
+/// receives (the floor of the hybrid allocation model; top-up codes extend
+/// it).
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct GrantSettings {
+    #[serde(default = "default_free_grant")]
+    pub free_token_grant: i64,
+}
+
+impl Default for GrantSettings {
+    fn default() -> Self {
+        Self {
+            free_token_grant: default_free_grant(),
+        }
+    }
+}
+
+/// `[abuse]` — silent multi-account abuse detection. When at least
+/// `fingerprint_account_threshold` accounts share one registration
+/// fingerprint, all of them are silently deactivated (no notice to the
+/// user; deactivation only surfaces as ordinary inference rejections).
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AbuseSettings {
+    #[serde(default = "default_fingerprint_threshold")]
+    pub fingerprint_account_threshold: i64,
+}
+
+impl Default for AbuseSettings {
+    fn default() -> Self {
+        Self {
+            fingerprint_account_threshold: default_fingerprint_threshold(),
+        }
+    }
+}
+
+fn default_listen() -> String {
+    "0.0.0.0:8090".into()
+}
+fn default_max_connections() -> u32 {
+    16
+}
+fn default_free_grant() -> i64 {
+    1_000_000
+}
+fn default_fingerprint_threshold() -> i64 {
+    5
+}
+fn default_reservation_ttl() -> u64 {
+    120
+}
+fn default_sweep_interval() -> u64 {
+    60
+}
+fn default_jwt_secret() -> String {
+    "dev-insecure-change-me".into()
+}
+fn default_session_ttl() -> u64 {
+    7 * 24 * 3600
+}
+fn default_email_token_ttl() -> u64 {
+    24 * 3600
+}
+fn default_app_base_url() -> String {
+    "http://localhost:5173".into()
+}
+fn default_email_provider() -> String {
+    "log".into()
+}
+fn default_from_addr() -> String {
+    "helexa <no-reply@helexa.ai>".into()
+}
+
+impl UpstreamConfig {
+    /// Load from a TOML file with `UPSTREAM_`-prefixed env overrides
+    /// (`__` nesting separator).
+    pub fn load(path: impl AsRef<Path>) -> Result<Self, Box<figment::Error>> {
+        Figment::new()
+            .merge(Toml::file(path))
+            .merge(Env::prefixed("UPSTREAM_").split("__"))
+            .extract()
+            .map_err(Box::new)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    #[allow(clippy::result_large_err)]
+    fn loads_toml_with_env_override_and_defaults() {
+        figment::Jail::expect_with(|jail| {
+            jail.create_file(
+                "helexa-upstream.toml",
+                r#"
+[db]
+url = "postgres://localhost/helexa"
+"#,
+            )?;
+            jail.set_env("UPSTREAM_SERVER__LISTEN", "127.0.0.1:9099");
+
+            let cfg = UpstreamConfig::load("helexa-upstream.toml").expect("load");
+            assert_eq!(cfg.server.listen, "127.0.0.1:9099");
+            assert_eq!(cfg.db.url, "postgres://localhost/helexa");
+            // Defaults applied when sections omitted.
+            assert_eq!(cfg.grant.free_token_grant, 1_000_000);
+            assert_eq!(cfg.abuse.fingerprint_account_threshold, 5);
+            assert_eq!(cfg.db.max_connections, 16);
+            Ok(())
+        });
+    }
+}

crates/helexa-upstream/src/crypto.rs

@@ -0,0 +1,113 @@
+//! Hashing + secret-generation helpers.
+//!
+//! - **Passwords** (low-entropy) → argon2id PHC strings.
+//! - **API keys / top-up codes / email + session tokens** (high-entropy
+//!   secrets minted here) → stored only as their sha256; sha256 is the fast,
+//!   sufficient choice for high-entropy material.
+
+use argon2::Argon2;
+use argon2::password_hash::rand_core::OsRng as ArgonOsRng;
+use argon2::password_hash::{PasswordHash, PasswordHasher, PasswordVerifier, SaltString};
+use rand::RngCore;
+use sha2::{Digest, Sha256};
+
+/// sha256 of `input`, as raw bytes (matches the `BYTEA` columns).
+pub fn sha256(input: &str) -> Vec<u8> {
+    let mut h = Sha256::new();
+    h.update(input.as_bytes());
+    h.finalize().to_vec()
+}
+
+/// Hash a password with argon2id, returning a PHC string for storage.
+pub fn hash_password(password: &str) -> Result<String, argon2::password_hash::Error> {
+    let salt = SaltString::generate(&mut ArgonOsRng);
+    Ok(Argon2::default()
+        .hash_password(password.as_bytes(), &salt)?
+        .to_string())
+}
+
+/// Verify a password against a stored PHC hash. `false` on any mismatch or
+/// malformed hash (never panics).
+pub fn verify_password(password: &str, phc: &str) -> bool {
+    match PasswordHash::new(phc) {
+        Ok(parsed) => Argon2::default()
+            .verify_password(password.as_bytes(), &parsed)
+            .is_ok(),
+        Err(_) => false,
+    }
+}
+
+/// A fresh URL-safe high-entropy secret (256 bits) for email/session/reset
+/// tokens. The caller stores only `sha256` of this and emails/returns the
+/// raw value.
+pub fn random_token() -> String {
+    let mut bytes = [0u8; 32];
+    rand::rngs::OsRng.fill_bytes(&mut bytes);
+    base62(&bytes)
+}
+
+/// Mint a new API key: `(raw, prefix)`. `raw` is shown to the user once;
+/// only `sha256(raw)` is stored. The prefix is a non-secret display tag.
+pub fn generate_api_key() -> (String, String) {
+    let mut bytes = [0u8; 32];
+    rand::rngs::OsRng.fill_bytes(&mut bytes);
+    let raw = format!("sk-helexa-{}", base62(&bytes));
+    // Non-secret prefix for the dashboard list (scheme + first few chars).
+    let prefix: String = raw.chars().take(14).collect();
+    (raw, prefix)
+}
+
+/// base62 encode (0-9A-Za-z) — URL/clipboard friendly, no padding.
+fn base62(bytes: &[u8]) -> String {
+    const ALPHABET: &[u8] = b"0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+    // Treat the bytes as a big-endian integer and base62 it. 32 bytes → ~43
+    // chars. Simple repeated-division over a big-uint built from the bytes.
+    let mut digits: Vec<u8> = vec![0];
+    for &byte in bytes {
+        let mut carry = byte as u32;
+        for d in digits.iter_mut() {
+            let v = (*d as u32) * 256 + carry;
+            *d = (v % 62) as u8;
+            carry = v / 62;
+        }
+        while carry > 0 {
+            digits.push((carry % 62) as u8);
+            carry /= 62;
+        }
+    }
+    digits
+        .iter()
+        .rev()
+        .map(|&d| ALPHABET[d as usize] as char)
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn password_round_trips_and_rejects_wrong() {
+        let phc = hash_password("correct horse").unwrap();
+        assert!(verify_password("correct horse", &phc));
+        assert!(!verify_password("wrong", &phc));
+        assert!(!verify_password("correct horse", "not-a-phc-string"));
+    }
+
+    #[test]
+    fn api_key_has_scheme_prefix_and_unique_body() {
+        let (raw, prefix) = generate_api_key();
+        assert!(raw.starts_with("sk-helexa-"));
+        assert!(prefix.starts_with("sk-helexa-"));
+        let (raw2, _) = generate_api_key();
+        assert_ne!(raw, raw2, "keys are unique");
+    }
+
+    #[test]
+    fn random_tokens_are_unique_and_nonempty() {
+        let a = random_token();
+        let b = random_token();
+        assert!(!a.is_empty());
+        assert_ne!(a, b);
+    }
+}

crates/helexa-upstream/src/db.rs

@@ -0,0 +1,20 @@
+//! PostgreSQL pool + embedded migrations.
+
+use anyhow::{Context, Result};
+use sqlx::postgres::{PgPool, PgPoolOptions};
+
+/// Connect to Postgres and run embedded migrations (`./migrations`).
+pub async fn connect_and_migrate(url: &str, max_connections: u32) -> Result<PgPool> {
+    let pool = PgPoolOptions::new()
+        .max_connections(max_connections)
+        .connect(url)
+        .await
+        .with_context(|| "connecting to PostgreSQL")?;
+
+    sqlx::migrate!("./migrations")
+        .run(&pool)
+        .await
+        .with_context(|| "running migrations")?;
+
+    Ok(pool)
+}

crates/helexa-upstream/src/email.rs

@@ -0,0 +1,64 @@
+//! Transactional email for verification + password-reset links.
+//!
+//! Two transports: `Log` (dev — writes the link to the log so flows are
+//! testable without a relay) and `Smtp` (lettre over rustls). Built from
+//! `[email]` config.
+
+use crate::config::EmailSettings;
+use anyhow::{Context, Result};
+use lettre::message::Mailbox;
+use lettre::{AsyncSmtpTransport, AsyncTransport, Message, Tokio1Executor};
+
+#[derive(Clone)]
+pub enum EmailSender {
+    /// Dev: log the message instead of sending.
+    Log { from: String },
+    Smtp {
+        from: String,
+        transport: AsyncSmtpTransport<Tokio1Executor>,
+    },
+}
+
+impl EmailSender {
+    pub fn from_config(cfg: &EmailSettings) -> Result<Self> {
+        match cfg.provider.as_str() {
+            "smtp" => {
+                let url = cfg
+                    .smtp_url
+                    .as_deref()
+                    .context("[email].smtp_url required when provider = \"smtp\"")?;
+                let transport = AsyncSmtpTransport::<Tokio1Executor>::from_url(url)
+                    .context("parsing [email].smtp_url")?
+                    .build();
+                Ok(EmailSender::Smtp {
+                    from: cfg.from_addr.clone(),
+                    transport,
+                })
+            }
+            _ => Ok(EmailSender::Log {
+                from: cfg.from_addr.clone(),
+            }),
+        }
+    }
+
+    /// Send a plaintext email. Errors are returned but the caller treats
+    /// send failures as non-fatal to the request (the user can re-request).
+    pub async fn send(&self, to: &str, subject: &str, body: &str) -> Result<()> {
+        match self {
+            EmailSender::Log { from } => {
+                tracing::info!(%from, %to, %subject, body, "EMAIL (log transport)");
+                Ok(())
+            }
+            EmailSender::Smtp { from, transport } => {
+                let msg = Message::builder()
+                    .from(from.parse::<Mailbox>().context("parsing from_addr")?)
+                    .to(to.parse::<Mailbox>().context("parsing recipient")?)
+                    .subject(subject)
+                    .body(body.to_string())
+                    .context("building message")?;
+                transport.send(msg).await.context("sending email")?;
+                Ok(())
+            }
+        }
+    }
+}

crates/helexa-upstream/src/error.rs

@@ -0,0 +1,21 @@
+//! Adapter from the shared, axum-agnostic
+//! [`cortex_core::error_envelope::OpenAiError`] (#60/#63) to an axum
+//! response, with `Retry-After`. The `/authz/v1` surface speaks the #63
+//! envelope so cortex (an OpenAI-compatible proxy) can forward rejections
+//! verbatim. (The future `/web/v1` surface uses a plain JSON error shape.)
+
+use axum::http::{HeaderValue, StatusCode, header};
+use axum::response::{IntoResponse, Json, Response};
+use cortex_core::error_envelope::OpenAiError;
+
+pub fn envelope_response(err: OpenAiError) -> Response {
+    let status = StatusCode::from_u16(err.status).unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
+    let retry_after = err.retry_after_secs;
+    let mut response = (status, Json(err.body())).into_response();
+    if let Some(secs) = retry_after
+        && let Ok(value) = HeaderValue::from_str(&secs.to_string())
+    {
+        response.headers_mut().insert(header::RETRY_AFTER, value);
+    }
+    response
+}

crates/helexa-upstream/src/handlers.rs

@@ -0,0 +1,21 @@
+//! HTTP handlers. B1 ships `/health`; the authz (`/authz/v1`) and web
+//! (`/web/v1`) surfaces land in later phases.
+
+use crate::state::AppState;
+use axum::{Json, Router, extract::State, routing::get};
+use serde_json::{Value, json};
+
+pub fn routes() -> Router<AppState> {
+    Router::new()
+        .route("/health", get(health))
+        .route("/", get(health))
+}
+
+/// `GET /health` — liveness + a database round-trip (`SELECT 1`).
+async fn health(State(state): State<AppState>) -> Json<Value> {
+    let db_ok = sqlx::query("SELECT 1").execute(&state.pool).await.is_ok();
+    Json(json!({
+        "status": if db_ok { "ok" } else { "degraded" },
+        "db": if db_ok { "ok" } else { "unreachable" },
+    }))
+}

crates/helexa-upstream/src/ledger.rs

@@ -0,0 +1,338 @@
+//! The allocation ledger: reserve → settle/release with the no-overshoot
+//! guarantee enforced by a row-locked transaction.
+//!
+//! Each reserve takes `SELECT … FOR UPDATE` on the account (and key) row, so
+//! concurrent reserves from many cortexes serialize and `spent + reserved`
+//! can never exceed the effective cap. The `accounts_no_overshoot` CHECK is
+//! the DB-level backstop. Settle/release are idempotent (they only act on a
+//! reservation still in `open`).
+//!
+//! Per-key effective cap = `min(resolved key cap, remaining account
+//! allocation)`. The key cap is resolved from its `limit_kind`:
+//! `hardcap` → the value verbatim; `percent` → that % of the account's
+//! `allocation_total`.
+//!
+//! Cap-window semantics: this module implements **Balance** (non-resetting)
+//! caps. Rolling-window key sub-caps (and the `RateLimited` rejection that
+//! rides them) land with the authz API (B2); today an over-cap is always
+//! `InsufficientQuota`.
+
+use sqlx::postgres::PgPool;
+use uuid::Uuid;
+
+/// A bearer key resolved to its principal + a budget snapshot.
+#[derive(Debug, Clone)]
+pub struct ResolvedPrincipal {
+    pub account_id: Uuid,
+    pub key_id: Uuid,
+    /// Effective per-key absolute cap (the key sub-cap; the account cap
+    /// still binds at reserve time).
+    pub hard_cap: i64,
+    pub key_spent: i64,
+    pub key_reserved: i64,
+}
+
+/// Resolve a key by its `sha256` hash to its principal, or `None` when the
+/// key is unknown/archived **or its account is deactivated** (the silent
+/// abuse flag — indistinguishable from an unknown key, by design: no clue).
+pub async fn resolve_key(
+    pool: &PgPool,
+    key_hash: &[u8],
+) -> Result<Option<ResolvedPrincipal>, sqlx::Error> {
+    let row = sqlx::query(
+        "SELECT k.id AS key_id, k.account_id, k.limit_kind, k.limit_value, \
+                k.key_spent, k.key_reserved, a.allocation_total \
+         FROM api_keys k JOIN accounts a ON a.id = k.account_id \
+         WHERE k.key_hash = $1 AND k.status = 'active' AND a.status = 'active'",
+    )
+    .bind(key_hash)
+    .fetch_optional(pool)
+    .await?;
+    Ok(row.map(|r| {
+        let total: i64 = sqlx::Row::get(&r, "allocation_total");
+        let limit_kind: String = sqlx::Row::get(&r, "limit_kind");
+        let limit_value: i64 = sqlx::Row::get(&r, "limit_value");
+        ResolvedPrincipal {
+            account_id: sqlx::Row::get(&r, "account_id"),
+            key_id: sqlx::Row::get(&r, "key_id"),
+            hard_cap: resolve_abs_cap(&limit_kind, limit_value, total),
+            key_spent: sqlx::Row::get(&r, "key_spent"),
+            key_reserved: sqlx::Row::get(&r, "key_reserved"),
+        }
+    }))
+}
+
+/// Per-key budget snapshot `(hard_cap, spent, reserved)`, or `None` if the
+/// key/account isn't an active pair.
+pub async fn snapshot(
+    pool: &PgPool,
+    account_id: Uuid,
+    key_id: Uuid,
+) -> Result<Option<(i64, i64, i64)>, sqlx::Error> {
+    let row = sqlx::query(
+        "SELECT k.limit_kind, k.limit_value, k.key_spent, k.key_reserved, a.allocation_total \
+         FROM api_keys k JOIN accounts a ON a.id = k.account_id \
+         WHERE k.id = $1 AND k.account_id = $2 AND k.status = 'active' AND a.status = 'active'",
+    )
+    .bind(key_id)
+    .bind(account_id)
+    .fetch_optional(pool)
+    .await?;
+    Ok(row.map(|r| {
+        let total: i64 = sqlx::Row::get(&r, "allocation_total");
+        let limit_kind: String = sqlx::Row::get(&r, "limit_kind");
+        let limit_value: i64 = sqlx::Row::get(&r, "limit_value");
+        let cap = resolve_abs_cap(&limit_kind, limit_value, total);
+        (
+            cap,
+            sqlx::Row::get::<i64, _>(&r, "key_spent"),
+            sqlx::Row::get::<i64, _>(&r, "key_reserved"),
+        )
+    }))
+}
+
+/// Release every `open` reservation older than `max_age_secs`, returning
+/// each one's reserved tokens to its account and key in a single statement.
+/// The lost-settle self-heal. Returns the number swept.
+pub async fn sweep_stale(pool: &PgPool, max_age_secs: i64) -> Result<u64, sqlx::Error> {
+    // Data-modifying CTEs: release stale rows, then fold their reserved sums
+    // back into accounts and api_keys. All in one atomic statement.
+    let result = sqlx::query(
+        "WITH stale AS ( \
+             UPDATE reservations SET state = 'released', settled_at = now() \
+             WHERE state = 'open' AND created_at < now() - make_interval(secs => $1) \
+             RETURNING account_id, key_id, reserved \
+         ), acct AS ( \
+             UPDATE accounts a SET allocation_reserved = allocation_reserved - s.total \
+             FROM (SELECT account_id, SUM(reserved) AS total FROM stale GROUP BY account_id) s \
+             WHERE a.id = s.account_id \
+         ) \
+         UPDATE api_keys k SET key_reserved = key_reserved - s.total \
+         FROM (SELECT key_id, SUM(reserved) AS total FROM stale GROUP BY key_id) s \
+         WHERE k.id = s.key_id",
+    )
+    .bind(max_age_secs as f64)
+    .execute(pool)
+    .await?;
+    Ok(result.rows_affected())
+}
+
+/// Resolve a key's per-key cap to an absolute token count.
+///
+/// `percent` is `floor(allocation_total * limit_value / 100)`; `hardcap` is
+/// `limit_value` verbatim. Computed in i128 to avoid overflow, floored at 0.
+pub fn resolve_abs_cap(limit_kind: &str, limit_value: i64, allocation_total: i64) -> i64 {
+    let cap = match limit_kind {
+        "percent" => (allocation_total as i128 * limit_value as i128) / 100,
+        _ => limit_value as i128, // "hardcap" (and any unknown → treat as absolute)
+    };
+    cap.clamp(0, i64::MAX as i128) as i64
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum LedgerError {
+    #[error("account not found")]
+    AccountNotFound,
+    #[error("api key not found or not active")]
+    KeyNotFound,
+    /// Account balance or a Balance-window key sub-cap is exhausted.
+    #[error("insufficient quota: requested {requested}, available {available}")]
+    InsufficientQuota { requested: i64, available: i64 },
+    #[error(transparent)]
+    Db(#[from] sqlx::Error),
+}
+
+/// Reserve `max_tokens` against `account_id`/`key_id`. Returns the
+/// reservation id (the `BIGSERIAL`, mapped to the cortex `Reservation.id`).
+pub async fn reserve(
+    pool: &PgPool,
+    account_id: Uuid,
+    key_id: Uuid,
+    max_tokens: i64,
+) -> Result<i64, LedgerError> {
+    let mut tx = pool.begin().await?;
+
+    // Lock the account row — serializes concurrent reserves on this account.
+    let acct = sqlx::query(
+        "SELECT allocation_total, allocation_spent, allocation_reserved \
+         FROM accounts WHERE id = $1 AND status = 'active' FOR UPDATE",
+    )
+    .bind(account_id)
+    .fetch_optional(&mut *tx)
+    .await?;
+    let Some(acct) = acct else {
+        return Err(LedgerError::AccountNotFound);
+    };
+    let total: i64 = sqlx::Row::get(&acct, "allocation_total");
+    let spent: i64 = sqlx::Row::get(&acct, "allocation_spent");
+    let reserved: i64 = sqlx::Row::get(&acct, "allocation_reserved");
+    let account_avail = total - spent - reserved;
+
+    // Lock the key row and resolve its absolute sub-cap.
+    let key = sqlx::query(
+        "SELECT limit_kind, limit_value, key_spent, key_reserved \
+         FROM api_keys WHERE id = $1 AND account_id = $2 AND status = 'active' FOR UPDATE",
+    )
+    .bind(key_id)
+    .bind(account_id)
+    .fetch_optional(&mut *tx)
+    .await?;
+    let Some(key) = key else {
+        return Err(LedgerError::KeyNotFound);
+    };
+    let limit_kind: String = sqlx::Row::get(&key, "limit_kind");
+    let limit_value: i64 = sqlx::Row::get(&key, "limit_value");
+    let key_spent: i64 = sqlx::Row::get(&key, "key_spent");
+    let key_reserved: i64 = sqlx::Row::get(&key, "key_reserved");
+    let key_cap = resolve_abs_cap(&limit_kind, limit_value, total);
+    let key_avail = key_cap - key_spent - key_reserved;
+
+    let available = account_avail.min(key_avail).max(0);
+    if max_tokens > available {
+        // tx rolls back on drop
+        return Err(LedgerError::InsufficientQuota {
+            requested: max_tokens,
+            available,
+        });
+    }
+
+    let id: i64 = sqlx::Row::get(
+        &sqlx::query(
+            "INSERT INTO reservations (account_id, key_id, reserved, state) \
+             VALUES ($1, $2, $3, 'open') RETURNING id",
+        )
+        .bind(account_id)
+        .bind(key_id)
+        .bind(max_tokens)
+        .fetch_one(&mut *tx)
+        .await?,
+        "id",
+    );
+    sqlx::query("UPDATE accounts SET allocation_reserved = allocation_reserved + $1 WHERE id = $2")
+        .bind(max_tokens)
+        .bind(account_id)
+        .execute(&mut *tx)
+        .await?;
+    sqlx::query("UPDATE api_keys SET key_reserved = key_reserved + $1 WHERE id = $2")
+        .bind(max_tokens)
+        .bind(key_id)
+        .execute(&mut *tx)
+        .await?;
+
+    tx.commit().await?;
+    Ok(id)
+}
+
+/// Settle a reservation with the actual tokens used (clamped to
+/// `[0, reserved]`). Idempotent: a second settle (or settle after release)
+/// is a no-op.
+pub async fn settle(
+    pool: &PgPool,
+    reservation_id: i64,
+    actual_tokens: i64,
+) -> Result<(), LedgerError> {
+    let mut tx = pool.begin().await?;
+    let row = sqlx::query(
+        "UPDATE reservations SET state = 'settled', settled_at = now(), \
+         actual = LEAST(GREATEST($2, 0), reserved) \
+         WHERE id = $1 AND state = 'open' \
+         RETURNING reserved, account_id, key_id, actual",
+    )
+    .bind(reservation_id)
+    .bind(actual_tokens)
+    .fetch_optional(&mut *tx)
+    .await?;
+    let Some(row) = row else {
+        return Ok(()); // already settled/released, or unknown → idempotent no-op
+    };
+    let reserved: i64 = sqlx::Row::get(&row, "reserved");
+    let actual: i64 = sqlx::Row::get(&row, "actual");
+    let account_id: Uuid = sqlx::Row::get(&row, "account_id");
+    let key_id: Uuid = sqlx::Row::get(&row, "key_id");
+
+    sqlx::query(
+        "UPDATE accounts SET allocation_reserved = allocation_reserved - $1, \
+         allocation_spent = allocation_spent + $2 WHERE id = $3",
+    )
+    .bind(reserved)
+    .bind(actual)
+    .bind(account_id)
+    .execute(&mut *tx)
+    .await?;
+    sqlx::query(
+        "UPDATE api_keys SET key_reserved = key_reserved - $1, key_spent = key_spent + $2 WHERE id = $3",
+    )
+    .bind(reserved)
+    .bind(actual)
+    .bind(key_id)
+    .execute(&mut *tx)
+    .await?;
+
+    tx.commit().await?;
+    Ok(())
+}
+
+/// Release a reservation, returning its full reserved amount to the
+/// allocation. Idempotent.
+pub async fn release(pool: &PgPool, reservation_id: i64) -> Result<(), LedgerError> {
+    let mut tx = pool.begin().await?;
+    let row = sqlx::query(
+        "UPDATE reservations SET state = 'released', settled_at = now() \
+         WHERE id = $1 AND state = 'open' \
+         RETURNING reserved, account_id, key_id",
+    )
+    .bind(reservation_id)
+    .fetch_optional(&mut *tx)
+    .await?;
+    let Some(row) = row else {
+        return Ok(());
+    };
+    let reserved: i64 = sqlx::Row::get(&row, "reserved");
+    let account_id: Uuid = sqlx::Row::get(&row, "account_id");
+    let key_id: Uuid = sqlx::Row::get(&row, "key_id");
+
+    sqlx::query("UPDATE accounts SET allocation_reserved = allocation_reserved - $1 WHERE id = $2")
+        .bind(reserved)
+        .bind(account_id)
+        .execute(&mut *tx)
+        .await?;
+    sqlx::query("UPDATE api_keys SET key_reserved = key_reserved - $1 WHERE id = $2")
+        .bind(reserved)
+        .bind(key_id)
+        .execute(&mut *tx)
+        .await?;
+
+    tx.commit().await?;
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::resolve_abs_cap;
+
+    #[test]
+    fn hardcap_is_verbatim() {
+        assert_eq!(resolve_abs_cap("hardcap", 50_000, 1_000_000), 50_000);
+    }
+
+    #[test]
+    fn percent_is_fraction_of_allocation() {
+        assert_eq!(resolve_abs_cap("percent", 25, 1_000_000), 250_000);
+        assert_eq!(resolve_abs_cap("percent", 100, 1_000_000), 1_000_000);
+        // floor
+        assert_eq!(resolve_abs_cap("percent", 33, 10), 3);
+    }
+
+    #[test]
+    fn percent_does_not_overflow_on_large_allocation() {
+        // total * value would overflow i64 if not widened to i128.
+        let cap = resolve_abs_cap("percent", 100, i64::MAX);
+        assert_eq!(cap, i64::MAX);
+    }
+
+    #[test]
+    fn negative_or_zero_clamps_to_zero() {
+        assert_eq!(resolve_abs_cap("hardcap", -5, 100), 0);
+        assert_eq!(resolve_abs_cap("percent", 0, 1_000_000), 0);
+    }
+}

crates/helexa-upstream/src/lib.rs

@@ -0,0 +1,90 @@
+//! helexa-upstream — the mesh-level account/authorization authority (#59).
+//!
+//! The clearing house above cortex: it issues accounts and API keys, holds
+//! the real token-allocation ledger, authorizes inference in real time
+//! (reserve → settle, fail-closed), and tracks served usage for operator
+//! reconciliation. cortex's `UpstreamEntitlementProvider` (#57) is a client
+//! of the `/authz/v1` surface; the helexa.ai frontend is a client of the
+//! `/web/v1` surface.
+//!
+//! Landed so far: B1 — schema + reserve→settle [`ledger`] (no-overshoot) +
+//! `/health`. B2 — the `/authz/v1` [`authz`] surface (resolve/reserve/
+//! settle/release/snapshot) with shared-bearer client auth and a
+//! stale-reservation sweeper.
+
+pub mod authz;
+pub mod config;
+pub mod crypto;
+pub mod db;
+pub mod email;
+pub mod error;
+pub mod handlers;
+pub mod ledger;
+pub mod reconcile;
+pub mod state;
+pub mod topup;
+pub mod web;
+
+use anyhow::Result;
+use config::UpstreamConfig;
+use email::EmailSender;
+use state::AppState;
+use std::time::Duration;
+use tower_http::cors::CorsLayer;
+use tower_http::trace::TraceLayer;
+
+/// Build the axum application.
+pub fn build_app(state: AppState) -> axum::Router {
+    axum::Router::new()
+        .merge(handlers::routes())
+        .merge(authz::router(&state))
+        .merge(web::router(&state))
+        // The /web/v1 surface is called cross-origin by the browser SPA in
+        // dev; same-origin behind nginx in prod. Permissive is fine — these
+        // endpoints authenticate via bearer/JWT, not cookies.
+        .layer(CorsLayer::permissive())
+        .layer(TraceLayer::new_for_http())
+        .with_state(state)
+}
+
+/// Start the service: connect Postgres, run migrations, spawn the
+/// reservation sweeper, bind the listener.
+pub async fn run(config: UpstreamConfig) -> Result<()> {
+    let pool = db::connect_and_migrate(&config.db.url, config.db.max_connections).await?;
+    let email = EmailSender::from_config(&config.email)?;
+    let listen = config.server.listen.clone();
+    let state = AppState::new(pool, config, email);
+
+    if state.config.client_auth.tokens.is_empty() {
+        tracing::warn!(
+            "no [client_auth] tokens configured — the /authz/v1 surface is OPEN (dev only)"
+        );
+    }
+
+    // Stale-reservation sweeper: releases open reservations whose
+    // settle/release from cortex was lost, self-healing allocation_reserved.
+    spawn_sweeper(&state);
+
+    let addr = listen.parse::<std::net::SocketAddr>()?;
+    tracing::info!("helexa-upstream listening on {addr}");
+
+    let listener = tokio::net::TcpListener::bind(addr).await?;
+    axum::serve(listener, build_app(state)).await?;
+    Ok(())
+}
+
+fn spawn_sweeper(state: &AppState) {
+    let pool = state.pool.clone();
+    let ttl = state.config.authz.reservation_ttl_secs as i64;
+    let interval = Duration::from_secs(state.config.authz.sweep_interval_secs);
+    tokio::spawn(async move {
+        loop {
+            tokio::time::sleep(interval).await;
+            match ledger::sweep_stale(&pool, ttl).await {
+                Ok(n) if n > 0 => tracing::info!(swept = n, "released stale reservations"),
+                Ok(_) => {}
+                Err(e) => tracing::warn!(error = %e, "reservation sweep failed"),
+            }
+        }
+    });
+}

crates/helexa-upstream/src/main.rs

@@ -0,0 +1,99 @@
+use anyhow::Result;
+use clap::{Parser, Subcommand};
+use helexa_upstream::config::UpstreamConfig;
+use tracing_subscriber::EnvFilter;
+
+#[derive(Parser)]
+#[command(name = "helexa-upstream")]
+#[command(about = "Mesh-level account & authorization authority for helexa")]
+#[command(version)]
+struct Cli {
+    #[command(subcommand)]
+    command: Commands,
+}
+
+#[derive(Subcommand)]
+enum Commands {
+    /// Start the upstream server.
+    Serve {
+        /// Path to the config file.
+        #[arg(short, long, default_value = "helexa-upstream.toml")]
+        config: String,
+    },
+    /// Mint single-use top-up codes and print them (one per line). The raw
+    /// codes are shown only here — only their hash is stored. (The future
+    /// faucet bot calls the same path.)
+    Mint {
+        #[arg(short, long, default_value = "helexa-upstream.toml")]
+        config: String,
+        /// Tokens each code grants.
+        #[arg(long)]
+        value: i64,
+        /// How many codes to mint.
+        #[arg(long, default_value_t = 1)]
+        count: u32,
+        /// Optional human label (e.g. "small", "beta-launch").
+        #[arg(long)]
+        denomination: Option<String>,
+    },
+    /// Roll up not-yet-reconciled served usage per operator/period (#58),
+    /// stamp it reconciled, and print the totals. Payout is out of scope.
+    Reconcile {
+        #[arg(short, long, default_value = "helexa-upstream.toml")]
+        config: String,
+    },
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| EnvFilter::new("info,helexa_upstream=debug")),
+        )
+        .init();
+
+    let cli = Cli::parse();
+
+    match cli.command {
+        Commands::Serve { config } => {
+            let cfg = UpstreamConfig::load(&config)
+                .map_err(|e| anyhow::anyhow!("failed to load config from '{config}': {e}"))?;
+            tracing::info!(listen = %cfg.server.listen, "starting helexa-upstream");
+            helexa_upstream::run(cfg).await?;
+        }
+        Commands::Mint {
+            config,
+            value,
+            count,
+            denomination,
+        } => {
+            let cfg = UpstreamConfig::load(&config)
+                .map_err(|e| anyhow::anyhow!("failed to load config from '{config}': {e}"))?;
+            let pool =
+                helexa_upstream::db::connect_and_migrate(&cfg.db.url, cfg.db.max_connections)
+                    .await?;
+            let codes =
+                helexa_upstream::topup::mint(&pool, value, count, denomination.as_deref()).await?;
+            // Raw codes to stdout (one per line) for the operator to distribute;
+            // logs/diagnostics go to stderr via tracing.
+            for code in codes {
+                println!("{code}");
+            }
+        }
+        Commands::Reconcile { config } => {
+            let cfg = UpstreamConfig::load(&config)
+                .map_err(|e| anyhow::anyhow!("failed to load config from '{config}': {e}"))?;
+            let pool =
+                helexa_upstream::db::connect_and_migrate(&cfg.db.url, cfg.db.max_connections)
+                    .await?;
+            let rollup = helexa_upstream::reconcile::reconcile(&pool).await?;
+            for r in &rollup {
+                println!("{}\t{}\t{}", r.operator_id, r.period, r.total_served_tokens);
+            }
+            tracing::info!(operators_periods = rollup.len(), "reconciliation complete");
+        }
+    }
+
+    Ok(())
+}

crates/helexa-upstream/src/reconcile.rs

@@ -0,0 +1,43 @@
+//! Reconciliation rollup (#58): aggregate the served-usage ledger per
+//! operator and period for operator compensation, stamping rows
+//! `reconciled_at` so each window is settled once. The payout mechanism
+//! itself is out of scope — this produces the authoritative per-operator
+//! totals a settlement process consumes.
+
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct RollupRow {
+    pub operator_id: String,
+    pub period: chrono::NaiveDate,
+    pub total_served_tokens: i64,
+}
+
+/// Roll up all not-yet-reconciled served-usage into per-(operator, period)
+/// totals, then stamp those rows `reconciled_at`. Returns the rollup.
+/// Idempotent: a second run finds nothing unreconciled and returns empty.
+pub async fn reconcile(pool: &PgPool) -> Result<Vec<RollupRow>, sqlx::Error> {
+    let mut tx = pool.begin().await?;
+    let rows = sqlx::query(
+        // SUM(bigint) is numeric in Postgres — cast back to bigint for i64.
+        "SELECT operator_id, period, SUM(served_tokens)::bigint AS total \
+         FROM served_usage WHERE reconciled_at IS NULL \
+         GROUP BY operator_id, period ORDER BY operator_id, period",
+    )
+    .fetch_all(&mut *tx)
+    .await?;
+    let rollup: Vec<RollupRow> = rows
+        .iter()
+        .map(|r| RollupRow {
+            operator_id: r.get("operator_id"),
+            period: r.get("period"),
+            total_served_tokens: r.get::<i64, _>("total"),
+        })
+        .collect();
+    sqlx::query("UPDATE served_usage SET reconciled_at = now() WHERE reconciled_at IS NULL")
+        .execute(&mut *tx)
+        .await?;
+    tx.commit().await?;
+    Ok(rollup)
+}

crates/helexa-upstream/src/state.rs

@@ -0,0 +1,23 @@
+//! Shared application state.
+
+use crate::config::UpstreamConfig;
+use crate::email::EmailSender;
+use sqlx::postgres::PgPool;
+use std::sync::Arc;
+
+#[derive(Clone)]
+pub struct AppState {
+    pub pool: PgPool,
+    pub config: Arc<UpstreamConfig>,
+    pub email: EmailSender,
+}
+
+impl AppState {
+    pub fn new(pool: PgPool, config: UpstreamConfig, email: EmailSender) -> Self {
+        Self {
+            pool,
+            config: Arc::new(config),
+            email,
+        }
+    }
+}

crates/helexa-upstream/src/topup.rs

@@ -0,0 +1,82 @@
+//! Single-use top-up codes (#B5) — the second half of the hybrid allocation
+//! model. Each code grants `value` tokens to the account that redeems it,
+//! raising `accounts.allocation_total`. Minting codes is operator/CLI side
+//! (the future faucet bot calls the same `mint` path); redemption is a
+//! `/web/v1` action.
+//!
+//! Security: only `sha256(code)` is stored. Redemption is **timing-safe and
+//! single-use** — a conditional `UPDATE … WHERE redeemed_by IS NULL` does
+//! the claim atomically (concurrent double-redeem → exactly one winner), and
+//! a not-found code and an already-redeemed code return the **same** generic
+//! failure with the same code path (no oracle for "valid but spent").
+
+use crate::crypto::{random_token, sha256};
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+use uuid::Uuid;
+
+#[derive(Debug, thiserror::Error)]
+pub enum TopUpError {
+    /// Code unknown OR already redeemed — deliberately indistinguishable.
+    #[error("invalid or already-redeemed code")]
+    Invalid,
+    #[error(transparent)]
+    Db(#[from] sqlx::Error),
+}
+
+/// Redeem `raw_code` for `account_id`, raising the account's
+/// `allocation_total` by the code's value. Returns the new total.
+pub async fn redeem(pool: &PgPool, account_id: Uuid, raw_code: &str) -> Result<i64, TopUpError> {
+    let mut tx = pool.begin().await?;
+    // Atomic single-use claim. `redeemed_by IS NULL` is the guarantee: under
+    // concurrent redemption exactly one UPDATE touches the row.
+    let claimed = sqlx::query(
+        "UPDATE top_up_codes SET redeemed_by = $1, redeemed_at = now() \
+         WHERE code_hash = $2 AND redeemed_by IS NULL RETURNING value",
+    )
+    .bind(account_id)
+    .bind(sha256(raw_code))
+    .fetch_optional(&mut *tx)
+    .await?;
+    let Some(row) = claimed else {
+        // Not found or already redeemed — same path, same error.
+        return Err(TopUpError::Invalid);
+    };
+    let value: i64 = row.get("value");
+    let new_total: i64 = sqlx::query(
+        "UPDATE accounts SET allocation_total = allocation_total + $1 WHERE id = $2 \
+         RETURNING allocation_total",
+    )
+    .bind(value)
+    .bind(account_id)
+    .fetch_one(&mut *tx)
+    .await?
+    .get("allocation_total");
+    tx.commit().await?;
+    Ok(new_total)
+}
+
+/// Mint `count` codes each worth `value` tokens, optionally tagged with a
+/// `denomination` label. Returns the raw codes (shown once — only their
+/// hash is stored). The CLI prints these; the future faucet bot calls this.
+pub async fn mint(
+    pool: &PgPool,
+    value: i64,
+    count: u32,
+    denomination: Option<&str>,
+) -> Result<Vec<String>, sqlx::Error> {
+    let mut codes = Vec::with_capacity(count as usize);
+    for _ in 0..count {
+        let raw = format!("helexa-topup-{}", random_token());
+        sqlx::query(
+            "INSERT INTO top_up_codes (code_hash, value, denomination) VALUES ($1, $2, $3)",
+        )
+        .bind(sha256(&raw))
+        .bind(value)
+        .bind(denomination)
+        .execute(pool)
+        .await?;
+        codes.push(raw);
+    }
+    Ok(codes)
+}

crates/helexa-upstream/src/web.rs

@@ -0,0 +1,594 @@
+//! `/web/v1` — the human-facing account API the helexa.ai frontend (#F4)
+//! consumes: email+password auth (register / verify / login / reset),
+//! API-key CRUD with per-key limits, and the account balance. Web sessions
+//! are JWTs, **distinct** from inference API keys.
+//!
+//! Errors use a plain JSON shape `{ "error": { "message", "code" } }` (web
+//! clients, not OpenAI clients — the #63 envelope is the authz surface).
+//!
+//! Silent fingerprint abuse (no clue to the abuser): registration captures
+//! the browser fingerprint and always succeeds; when ≥ threshold accounts
+//! share one fingerprint, all are silently `deactivated` (keys then resolve
+//! as ordinary `401`s at the authz surface — never a "banned" signal).
+
+use crate::crypto::{generate_api_key, hash_password, random_token, sha256, verify_password};
+use crate::state::AppState;
+use axum::extract::{Path, Request, State};
+use axum::http::{StatusCode, header};
+use axum::middleware::Next;
+use axum::response::{IntoResponse, Json, Response};
+use axum::routing::{get, post};
+use axum::{Extension, Router};
+use chrono::{DateTime, Duration, Utc};
+use jsonwebtoken::{DecodingKey, EncodingKey, Header, Validation, decode, encode};
+use serde::{Deserialize, Serialize};
+use serde_json::json;
+use sqlx::Row;
+use uuid::Uuid;
+
+pub fn router(state: &AppState) -> Router<AppState> {
+    let protected = Router::new()
+        .route("/web/v1/account", get(account))
+        .route("/web/v1/keys", get(list_keys).post(create_key))
+        .route("/web/v1/keys/{id}/archive", post(archive_key))
+        .route(
+            "/web/v1/keys/{id}/limit",
+            axum::routing::patch(update_key_limit),
+        )
+        .route("/web/v1/redeem", post(redeem))
+        .layer(axum::middleware::from_fn_with_state(
+            state.clone(),
+            require_session,
+        ));
+
+    Router::new()
+        .route("/web/v1/register", post(register))
+        .route("/web/v1/verify", post(verify))
+        .route("/web/v1/login", post(login))
+        .route("/web/v1/password-reset/request", post(reset_request))
+        .route("/web/v1/password-reset/confirm", post(reset_confirm))
+        .merge(protected)
+}
+
+// ── errors ──────────────────────────────────────────────────────────
+
+enum WebError {
+    BadRequest(&'static str),
+    Unauthorized,
+    Internal,
+}
+
+impl IntoResponse for WebError {
+    fn into_response(self) -> Response {
+        let (status, code, message) = match self {
+            WebError::BadRequest(m) => (StatusCode::BAD_REQUEST, "bad_request", m),
+            WebError::Unauthorized => (StatusCode::UNAUTHORIZED, "unauthorized", "unauthorized"),
+            WebError::Internal => (
+                StatusCode::INTERNAL_SERVER_ERROR,
+                "internal_error",
+                "internal error",
+            ),
+        };
+        (
+            status,
+            Json(json!({"error": {"message": message, "code": code}})),
+        )
+            .into_response()
+    }
+}
+
+impl From<sqlx::Error> for WebError {
+    fn from(e: sqlx::Error) -> Self {
+        tracing::error!(error = %e, "web db error");
+        WebError::Internal
+    }
+}
+
+type WebResult<T> = Result<T, WebError>;
+
+// ── sessions (JWT) ──────────────────────────────────────────────────
+
+#[derive(Serialize, Deserialize)]
+struct Claims {
+    sub: String, // user id
+    exp: usize,
+}
+
+fn mint_session(state: &AppState, user_id: Uuid) -> WebResult<String> {
+    let exp = (Utc::now() + Duration::seconds(state.config.auth.session_ttl_secs as i64))
+        .timestamp() as usize;
+    let claims = Claims {
+        sub: user_id.to_string(),
+        exp,
+    };
+    encode(
+        &Header::default(),
+        &claims,
+        &EncodingKey::from_secret(state.config.auth.jwt_secret.as_bytes()),
+    )
+    .map_err(|_| WebError::Internal)
+}
+
+/// Authenticated user id, injected by [`require_session`].
+#[derive(Clone)]
+struct AuthUser(Uuid);
+
+async fn require_session(State(state): State<AppState>, mut req: Request, next: Next) -> Response {
+    let token = req
+        .headers()
+        .get(header::AUTHORIZATION)
+        .and_then(|v| v.to_str().ok())
+        .and_then(|v| v.strip_prefix("Bearer "))
+        .map(str::trim);
+    let Some(token) = token else {
+        return WebError::Unauthorized.into_response();
+    };
+    let decoded = decode::<Claims>(
+        token,
+        &DecodingKey::from_secret(state.config.auth.jwt_secret.as_bytes()),
+        &Validation::default(),
+    );
+    match decoded
+        .ok()
+        .and_then(|d| Uuid::parse_str(&d.claims.sub).ok())
+    {
+        Some(uid) => {
+            req.extensions_mut().insert(AuthUser(uid));
+            next.run(req).await
+        }
+        None => WebError::Unauthorized.into_response(),
+    }
+}
+
+/// The caller's single account id.
+async fn account_id_for(state: &AppState, user_id: Uuid) -> WebResult<Uuid> {
+    let row = sqlx::query("SELECT id FROM accounts WHERE owner_user_id = $1")
+        .bind(user_id)
+        .fetch_optional(&state.pool)
+        .await?;
+    row.map(|r| r.get::<Uuid, _>("id"))
+        .ok_or(WebError::Internal)
+}
+
+// ── auth lifecycle ──────────────────────────────────────────────────
+
+#[derive(Deserialize)]
+struct RegisterReq {
+    email: String,
+    password: String,
+    #[serde(default)]
+    fingerprint: Option<String>,
+}
+
+/// `POST /web/v1/register` — always returns `202`, regardless of whether the
+/// email was new, already taken, or fingerprint-flagged (no enumeration, no
+/// abuse clue).
+async fn register(State(state): State<AppState>, Json(req): Json<RegisterReq>) -> Response {
+    match register_inner(&state, req).await {
+        Ok(()) | Err(WebError::BadRequest(_)) => {}
+        Err(e) => return e.into_response(),
+    }
+    // Generic 202 whatever happened above (except hard server errors).
+    StatusCode::ACCEPTED.into_response()
+}
+
+async fn register_inner(state: &AppState, req: RegisterReq) -> WebResult<()> {
+    if !req.email.contains('@') {
+        return Err(WebError::BadRequest("invalid email"));
+    }
+    if req.password.len() < 8 {
+        return Err(WebError::BadRequest("password too short (min 8)"));
+    }
+    let phc = hash_password(&req.password).map_err(|_| WebError::Internal)?;
+
+    // Insert the user; a duplicate email silently no-ops (no enumeration).
+    let user_id: Option<Uuid> = sqlx::query(
+        "INSERT INTO users (email, password_hash, registration_fingerprint) \
+         VALUES ($1, $2, $3) ON CONFLICT (email) DO NOTHING RETURNING id",
+    )
+    .bind(&req.email)
+    .bind(&phc)
+    .bind(&req.fingerprint)
+    .fetch_optional(&state.pool)
+    .await?
+    .map(|r| r.get("id"));
+
+    let Some(user_id) = user_id else {
+        return Ok(()); // email already registered — say nothing
+    };
+
+    // Account with the flat free grant.
+    sqlx::query("INSERT INTO accounts (owner_user_id, allocation_total) VALUES ($1, $2)")
+        .bind(user_id)
+        .bind(state.config.grant.free_token_grant)
+        .execute(&state.pool)
+        .await?;
+
+    // Silent fingerprint abuse handling.
+    if let Some(fp) = req.fingerprint.as_deref().filter(|f| !f.is_empty()) {
+        apply_fingerprint_policy(state, fp).await?;
+    }
+
+    // Email verification link.
+    let token = random_token();
+    let expires: DateTime<Utc> =
+        Utc::now() + Duration::seconds(state.config.auth.email_token_ttl_secs as i64);
+    sqlx::query(
+        "INSERT INTO email_tokens (token_hash, user_id, kind, expires_at) \
+         VALUES ($1, $2, 'verify', $3)",
+    )
+    .bind(sha256(&token))
+    .bind(user_id)
+    .bind(expires)
+    .execute(&state.pool)
+    .await?;
+
+    let link = format!("{}/verify?token={token}", state.config.auth.app_base_url);
+    let _ = state
+        .email
+        .send(
+            &req.email,
+            "Verify your helexa account",
+            &format!("Welcome to helexa. Verify your email:\n\n{link}\n"),
+        )
+        .await;
+    Ok(())
+}
+
+/// Count accounts sharing `fp`; flag them, and silently deactivate all once
+/// the count reaches the configured threshold. No response difference — the
+/// abuser gets no signal.
+async fn apply_fingerprint_policy(state: &AppState, fp: &str) -> WebResult<()> {
+    let count: i64 =
+        sqlx::query_scalar("SELECT count(*) FROM users WHERE registration_fingerprint = $1")
+            .bind(fp)
+            .fetch_one(&state.pool)
+            .await?;
+    if count > 1 {
+        sqlx::query(
+            "UPDATE accounts SET fingerprint_flagged = true \
+             WHERE owner_user_id IN (SELECT id FROM users WHERE registration_fingerprint = $1)",
+        )
+        .bind(fp)
+        .execute(&state.pool)
+        .await?;
+    }
+    if count >= state.config.abuse.fingerprint_account_threshold {
+        let res = sqlx::query(
+            "UPDATE accounts SET status = 'deactivated' \
+             WHERE owner_user_id IN (SELECT id FROM users WHERE registration_fingerprint = $1)",
+        )
+        .bind(fp)
+        .execute(&state.pool)
+        .await?;
+        tracing::warn!(
+            fingerprint = fp,
+            accounts = res.rows_affected(),
+            "silently deactivated fingerprint-abusing accounts"
+        );
+    }
+    Ok(())
+}
+
+#[derive(Deserialize)]
+struct TokenReq {
+    token: String,
+}
+
+/// `POST /web/v1/verify` — consume a verification token, mark verified.
+async fn verify(State(state): State<AppState>, Json(req): Json<TokenReq>) -> WebResult<Response> {
+    let row = sqlx::query(
+        "UPDATE email_tokens SET consumed_at = now() \
+         WHERE token_hash = $1 AND kind = 'verify' AND consumed_at IS NULL AND expires_at > now() \
+         RETURNING user_id",
+    )
+    .bind(sha256(&req.token))
+    .fetch_optional(&state.pool)
+    .await?;
+    let Some(row) = row else {
+        return Err(WebError::BadRequest("invalid or expired token"));
+    };
+    let user_id: Uuid = row.get("user_id");
+    sqlx::query("UPDATE users SET email_verified = true WHERE id = $1")
+        .bind(user_id)
+        .execute(&state.pool)
+        .await?;
+    Ok(StatusCode::OK.into_response())
+}
+
+#[derive(Deserialize)]
+struct LoginReq {
+    email: String,
+    password: String,
+}
+
+/// `POST /web/v1/login` — verify password + email-verified → session JWT.
+async fn login(State(state): State<AppState>, Json(req): Json<LoginReq>) -> WebResult<Response> {
+    let row = sqlx::query("SELECT id, password_hash, email_verified FROM users WHERE email = $1")
+        .bind(&req.email)
+        .fetch_optional(&state.pool)
+        .await?;
+    // Generic 401 for every failure mode (no enumeration).
+    let Some(row) = row else {
+        return Err(WebError::Unauthorized);
+    };
+    let phc: String = row.get("password_hash");
+    let verified: bool = row.get("email_verified");
+    if !verify_password(&req.password, &phc) || !verified {
+        return Err(WebError::Unauthorized);
+    }
+    let user_id: Uuid = row.get("id");
+    let token = mint_session(&state, user_id)?;
+    Ok(Json(json!({
+        "token": token,
+        "expires_in": state.config.auth.session_ttl_secs,
+    }))
+    .into_response())
+}
+
+#[derive(Deserialize)]
+struct EmailReq {
+    email: String,
+}
+
+/// `POST /web/v1/password-reset/request` — always `202` (no enumeration);
+/// mints + emails a reset token only if the account exists.
+async fn reset_request(State(state): State<AppState>, Json(req): Json<EmailReq>) -> Response {
+    // The inner only ever yields `Internal` (DB failure); a missing email is
+    // Ok(()) so there's no enumeration. Surface 500 on a real error, else 202.
+    match reset_request_inner(&state, &req.email).await {
+        Ok(()) => StatusCode::ACCEPTED.into_response(),
+        Err(e) => e.into_response(),
+    }
+}
+
+async fn reset_request_inner(state: &AppState, email: &str) -> WebResult<()> {
+    let row = sqlx::query("SELECT id FROM users WHERE email = $1")
+        .bind(email)
+        .fetch_optional(&state.pool)
+        .await?;
+    let Some(row) = row else { return Ok(()) };
+    let user_id: Uuid = row.get("id");
+    let token = random_token();
+    let expires: DateTime<Utc> =
+        Utc::now() + Duration::seconds(state.config.auth.email_token_ttl_secs as i64);
+    sqlx::query(
+        "INSERT INTO email_tokens (token_hash, user_id, kind, expires_at) \
+         VALUES ($1, $2, 'reset', $3)",
+    )
+    .bind(sha256(&token))
+    .bind(user_id)
+    .bind(expires)
+    .execute(&state.pool)
+    .await?;
+    let link = format!("{}/reset?token={token}", state.config.auth.app_base_url);
+    let _ = state
+        .email
+        .send(
+            email,
+            "Reset your helexa password",
+            &format!("Reset your password:\n\n{link}\n"),
+        )
+        .await;
+    Ok(())
+}
+
+#[derive(Deserialize)]
+struct ResetConfirmReq {
+    token: String,
+    new_password: String,
+}
+
+/// `POST /web/v1/password-reset/confirm` — consume reset token, rotate hash.
+async fn reset_confirm(
+    State(state): State<AppState>,
+    Json(req): Json<ResetConfirmReq>,
+) -> WebResult<Response> {
+    if req.new_password.len() < 8 {
+        return Err(WebError::BadRequest("password too short (min 8)"));
+    }
+    let row = sqlx::query(
+        "UPDATE email_tokens SET consumed_at = now() \
+         WHERE token_hash = $1 AND kind = 'reset' AND consumed_at IS NULL AND expires_at > now() \
+         RETURNING user_id",
+    )
+    .bind(sha256(&req.token))
+    .fetch_optional(&state.pool)
+    .await?;
+    let Some(row) = row else {
+        return Err(WebError::BadRequest("invalid or expired token"));
+    };
+    let user_id: Uuid = row.get("user_id");
+    let phc = hash_password(&req.new_password).map_err(|_| WebError::Internal)?;
+    sqlx::query("UPDATE users SET password_hash = $1 WHERE id = $2")
+        .bind(phc)
+        .bind(user_id)
+        .execute(&state.pool)
+        .await?;
+    Ok(StatusCode::OK.into_response())
+}
+
+// ── account + keys (protected) ──────────────────────────────────────
+
+async fn account(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+) -> WebResult<Response> {
+    let acct = account_id_for(&state, user.0).await?;
+    let row = sqlx::query(
+        "SELECT allocation_total, allocation_spent, allocation_reserved FROM accounts WHERE id = $1",
+    )
+    .bind(acct)
+    .fetch_one(&state.pool)
+    .await?;
+    Ok(Json(json!({
+        "account_id": acct.to_string(),
+        "allocation_total": row.get::<i64, _>("allocation_total"),
+        "allocation_spent": row.get::<i64, _>("allocation_spent"),
+        "allocation_reserved": row.get::<i64, _>("allocation_reserved"),
+    }))
+    .into_response())
+}
+
+async fn list_keys(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+) -> WebResult<Response> {
+    let acct = account_id_for(&state, user.0).await?;
+    let rows = sqlx::query(
+        "SELECT id, key_prefix, label, status, limit_kind, limit_value, key_spent, key_reserved, \
+                created_at \
+         FROM api_keys WHERE account_id = $1 ORDER BY created_at DESC",
+    )
+    .bind(acct)
+    .fetch_all(&state.pool)
+    .await?;
+    let keys: Vec<_> = rows
+        .iter()
+        .map(|r| {
+            json!({
+                "id": r.get::<Uuid, _>("id").to_string(),
+                "prefix": r.get::<String, _>("key_prefix"),
+                "label": r.get::<String, _>("label"),
+                "status": r.get::<String, _>("status"),
+                "limit_kind": r.get::<String, _>("limit_kind"),
+                "limit_value": r.get::<i64, _>("limit_value"),
+                "spent": r.get::<i64, _>("key_spent"),
+                "reserved": r.get::<i64, _>("key_reserved"),
+                "created_at": r.get::<DateTime<Utc>, _>("created_at").to_rfc3339(),
+            })
+        })
+        .collect();
+    Ok(Json(json!({ "keys": keys })).into_response())
+}
+
+#[derive(Deserialize)]
+struct CreateKeyReq {
+    #[serde(default)]
+    label: String,
+    /// "percent" | "hardcap" (default percent=100 → full allocation).
+    #[serde(default)]
+    limit_kind: Option<String>,
+    #[serde(default)]
+    limit_value: Option<i64>,
+}
+
+async fn create_key(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+    Json(req): Json<CreateKeyReq>,
+) -> WebResult<Response> {
+    let acct = account_id_for(&state, user.0).await?;
+    let limit_kind = match req.limit_kind.as_deref() {
+        Some("hardcap") => "hardcap",
+        _ => "percent",
+    };
+    let limit_value = req.limit_value.unwrap_or(100).max(0);
+    let (raw, prefix) = generate_api_key();
+    let id: Uuid = sqlx::query(
+        "INSERT INTO api_keys (account_id, key_hash, key_prefix, label, limit_kind, limit_value) \
+         VALUES ($1, $2, $3, $4, $5, $6) RETURNING id",
+    )
+    .bind(acct)
+    .bind(sha256(&raw))
+    .bind(&prefix)
+    .bind(&req.label)
+    .bind(limit_kind)
+    .bind(limit_value)
+    .fetch_one(&state.pool)
+    .await?
+    .get("id");
+    // The raw key is shown exactly once.
+    Ok((
+        StatusCode::CREATED,
+        Json(json!({
+            "id": id.to_string(),
+            "key": raw,
+            "prefix": prefix,
+            "limit_kind": limit_kind,
+            "limit_value": limit_value,
+        })),
+    )
+        .into_response())
+}
+
+async fn archive_key(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+    Path(id): Path<Uuid>,
+) -> WebResult<Response> {
+    let acct = account_id_for(&state, user.0).await?;
+    let res = sqlx::query(
+        "UPDATE api_keys SET status = 'archived' WHERE id = $1 AND account_id = $2 AND status = 'active'",
+    )
+    .bind(id)
+    .bind(acct)
+    .execute(&state.pool)
+    .await?;
+    if res.rows_affected() == 0 {
+        return Err(WebError::BadRequest("no such active key"));
+    }
+    Ok(StatusCode::NO_CONTENT.into_response())
+}
+
+#[derive(Deserialize)]
+struct UpdateLimitReq {
+    limit_kind: String,
+    limit_value: i64,
+}
+
+async fn update_key_limit(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+    Path(id): Path<Uuid>,
+    Json(req): Json<UpdateLimitReq>,
+) -> WebResult<Response> {
+    if req.limit_kind != "percent" && req.limit_kind != "hardcap" {
+        return Err(WebError::BadRequest(
+            "limit_kind must be percent or hardcap",
+        ));
+    }
+    if req.limit_value < 0 {
+        return Err(WebError::BadRequest("limit_value must be >= 0"));
+    }
+    let acct = account_id_for(&state, user.0).await?;
+    let res = sqlx::query(
+        "UPDATE api_keys SET limit_kind = $1, limit_value = $2 WHERE id = $3 AND account_id = $4",
+    )
+    .bind(&req.limit_kind)
+    .bind(req.limit_value)
+    .bind(id)
+    .bind(acct)
+    .execute(&state.pool)
+    .await?;
+    if res.rows_affected() == 0 {
+        return Err(WebError::BadRequest("no such key"));
+    }
+    Ok(StatusCode::NO_CONTENT.into_response())
+}
+
+#[derive(Deserialize)]
+struct RedeemReq {
+    code: String,
+}
+
+/// `POST /web/v1/redeem` — redeem a single-use top-up code, raising the
+/// account's allocation. Returns the new total. Generic 400 for an invalid
+/// or already-redeemed code (no oracle).
+async fn redeem(
+    State(state): State<AppState>,
+    Extension(user): Extension<AuthUser>,
+    Json(req): Json<RedeemReq>,
+) -> WebResult<Response> {
+    let acct = account_id_for(&state, user.0).await?;
+    match crate::topup::redeem(&state.pool, acct, &req.code).await {
+        Ok(new_total) => Ok(Json(json!({ "allocation_total": new_total })).into_response()),
+        Err(crate::topup::TopUpError::Invalid) => {
+            Err(WebError::BadRequest("invalid or already-redeemed code"))
+        }
+        Err(crate::topup::TopUpError::Db(e)) => {
+            tracing::error!(error = %e, "redeem db error");
+            Err(WebError::Internal)
+        }
+    }
+}

crates/helexa-upstream/tests/authz_pg.rs

@@ -0,0 +1,243 @@
+//! Integration tests for the `/authz/v1` surface against a real Postgres,
+//! driving the built axum app over HTTP. Gated on `UPSTREAM_TEST_DATABASE_URL`
+//! (skips cleanly when unset, so CI stays green without a DB):
+//!
+//!   UPSTREAM_TEST_DATABASE_URL=postgres://helexa:helexa@localhost/helexa_test \
+//!     cargo test -p helexa-upstream --test authz_pg
+
+use helexa_upstream::config::{ClientToken, UpstreamConfig};
+use helexa_upstream::crypto::sha256;
+use helexa_upstream::db::connect_and_migrate;
+use helexa_upstream::state::AppState;
+use serde_json::{Value, json};
+use sqlx::Executor;
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+use uuid::Uuid;
+
+const CLIENT_TOKEN: &str = "test-operator-token";
+
+async fn spawn_or_skip(test: &str) -> Option<(String, PgPool)> {
+    let Ok(url) = std::env::var("UPSTREAM_TEST_DATABASE_URL") else {
+        eprintln!("skipping {test}: UPSTREAM_TEST_DATABASE_URL not set");
+        return None;
+    };
+    let pool = connect_and_migrate(&url, 16).await.expect("migrate");
+
+    let mut config = UpstreamConfig {
+        server: Default::default(),
+        db: helexa_upstream::config::DbSettings {
+            url,
+            max_connections: 16,
+        },
+        grant: Default::default(),
+        abuse: Default::default(),
+        client_auth: Default::default(),
+        authz: Default::default(),
+        auth: Default::default(),
+        email: Default::default(),
+    };
+    config.client_auth.tokens.push(ClientToken {
+        token: CLIENT_TOKEN.into(),
+        operator_id: "op-test".into(),
+    });
+
+    let email = helexa_upstream::email::EmailSender::from_config(&config.email).unwrap();
+    let state = AppState::new(pool.clone(), config, email);
+    let app = helexa_upstream::build_app(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    Some((format!("http://{addr}"), pool))
+}
+
+/// Seed an account with `total` allocation and an active key with raw value
+/// `raw` (percent=100). Optionally deactivate the account. Returns
+/// (account_id, key_id).
+async fn seed_key(pool: &PgPool, total: i64, raw: &str, deactivated: bool) -> (Uuid, Uuid) {
+    let user_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO users (email, password_hash, email_verified) VALUES ($1,'x',true) RETURNING id",
+            )
+            .bind(format!("u-{}@t.local", Uuid::new_v4())),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let status = if deactivated { "deactivated" } else { "active" };
+    let account_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO accounts (owner_user_id, allocation_total, status) VALUES ($1,$2,$3) RETURNING id",
+            )
+            .bind(user_id)
+            .bind(total)
+            .bind(status),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let key_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO api_keys (account_id, key_hash, key_prefix, limit_kind, limit_value) \
+                 VALUES ($1,$2,'sk-test','percent',100) RETURNING id",
+            )
+            .bind(account_id)
+            .bind(sha256(raw)),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    (account_id, key_id)
+}
+
+fn client() -> reqwest::Client {
+    reqwest::Client::new()
+}
+
+async fn post(
+    c: &reqwest::Client,
+    url: String,
+    body: Value,
+    bearer: Option<&str>,
+) -> reqwest::Response {
+    let mut req = c.post(url).json(&body);
+    if let Some(b) = bearer {
+        req = req.bearer_auth(b);
+    }
+    req.send().await.unwrap()
+}
+
+#[tokio::test]
+async fn resolve_reserve_settle_round_trip() {
+    let Some((base, pool)) = spawn_or_skip("resolve_reserve_settle_round_trip").await else {
+        return;
+    };
+    let raw = format!("sk-{}", Uuid::new_v4());
+    let (account_id, key_id) = seed_key(&pool, 1000, &raw, false).await;
+    let c = client();
+
+    // resolve
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 200);
+    let body: Value = r.json().await.unwrap();
+    assert_eq!(body["principal"]["account_id"], account_id.to_string());
+    assert_eq!(body["principal"]["key_id"], key_id.to_string());
+    assert_eq!(body["snapshot"]["hard_cap"], 1000);
+
+    // reserve 400
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/reserve"),
+        json!({"account_id": account_id, "key_id": key_id, "max_tokens": 400}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 200);
+    let body: Value = r.json().await.unwrap();
+    let rid = body["reservation_id"].as_i64().expect("granted");
+
+    // settle 150
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/settle"),
+        json!({"reservation_id": rid, "actual_tokens": 150}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 204);
+
+    // snapshot reflects spend
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/snapshot"),
+        json!({"account_id": account_id, "key_id": key_id}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    let body: Value = r.json().await.unwrap();
+    assert_eq!(body["spent"], 150);
+    assert_eq!(body["reserved"], 0);
+}
+
+#[tokio::test]
+async fn over_cap_reserve_is_rejected_not_errored() {
+    let Some((base, pool)) = spawn_or_skip("over_cap_reserve_is_rejected_not_errored").await else {
+        return;
+    };
+    let raw = format!("sk-{}", Uuid::new_v4());
+    let (account_id, key_id) = seed_key(&pool, 100, &raw, false).await;
+    let c = client();
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/reserve"),
+        json!({"account_id": account_id, "key_id": key_id, "max_tokens": 999}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 200, "budget refusal is an authoritative 200");
+    let body: Value = r.json().await.unwrap();
+    assert!(body["reservation_id"].is_null());
+    assert_eq!(body["rejected"]["kind"], "insufficient_quota");
+    assert_eq!(body["rejected"]["available"], 100);
+}
+
+#[tokio::test]
+async fn deactivated_account_resolves_as_invalid_no_clue() {
+    let Some((base, pool)) = spawn_or_skip("deactivated_account_resolves_as_invalid_no_clue").await
+    else {
+        return;
+    };
+    let raw = format!("sk-{}", Uuid::new_v4());
+    seed_key(&pool, 1000, &raw, true).await; // deactivated
+    let c = client();
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    // Indistinguishable from an unknown key.
+    assert_eq!(r.status(), 401);
+    let body: Value = r.json().await.unwrap();
+    assert_eq!(body["error"]["code"], "invalid_api_key");
+}
+
+#[tokio::test]
+async fn missing_client_auth_is_401_before_db() {
+    let Some((base, pool)) = spawn_or_skip("missing_client_auth_is_401_before_db").await else {
+        return;
+    };
+    let raw = format!("sk-{}", Uuid::new_v4());
+    seed_key(&pool, 1000, &raw, false).await;
+    let c = client();
+    // No bearer → rejected by client_auth.
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw}),
+        None,
+    )
+    .await;
+    assert_eq!(r.status(), 401);
+    // Wrong bearer → also rejected.
+    let r = post(
+        &c,
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw}),
+        Some("wrong"),
+    )
+    .await;
+    assert_eq!(r.status(), 401);
+}

crates/helexa-upstream/tests/ledger_pg.rs

@@ -0,0 +1,204 @@
+//! Integration tests for the allocation ledger against a real PostgreSQL.
+//!
+//! Gated on `UPSTREAM_TEST_DATABASE_URL` — when unset (CI's generic runner,
+//! local builds without a DB), every test logs a skip and returns, so
+//! `cargo test --workspace` stays green without Postgres. Point the env var
+//! at a throwaway database to exercise the no-overshoot guarantee and
+//! settle/release idempotency:
+//!
+//!   UPSTREAM_TEST_DATABASE_URL=postgres://helexa:helexa@localhost/helexa_test \
+//!     cargo test -p helexa-upstream --test ledger_pg
+
+use helexa_upstream::db::connect_and_migrate;
+use helexa_upstream::ledger::{self, LedgerError};
+use sqlx::Executor;
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+use uuid::Uuid;
+
+/// Returns a migrated pool, or `None` (with a skip log) when the env var is
+/// unset.
+async fn pool_or_skip(test: &str) -> Option<PgPool> {
+    let Ok(url) = std::env::var("UPSTREAM_TEST_DATABASE_URL") else {
+        eprintln!("skipping {test}: UPSTREAM_TEST_DATABASE_URL not set");
+        return None;
+    };
+    Some(
+        connect_and_migrate(&url, 16)
+            .await
+            .expect("connect + migrate"),
+    )
+}
+
+/// Seed a verified user + account (with `total` allocation) + an active key
+/// (percent=100 so the account cap binds). Returns (account_id, key_id).
+async fn seed(pool: &PgPool, total: i64) -> (Uuid, Uuid) {
+    let user_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO users (email, password_hash, email_verified) \
+                 VALUES ($1, 'x', true) RETURNING id",
+            )
+            .bind(format!("u-{}@test.local", Uuid::new_v4())),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let account_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO accounts (owner_user_id, allocation_total) \
+                 VALUES ($1, $2) RETURNING id",
+            )
+            .bind(user_id)
+            .bind(total),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let key_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO api_keys (account_id, key_hash, key_prefix, limit_kind, limit_value) \
+                 VALUES ($1, $2, 'sk-test', 'percent', 100) RETURNING id",
+            )
+            .bind(account_id)
+            .bind(Uuid::new_v4().as_bytes().to_vec()),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    (account_id, key_id)
+}
+
+async fn account_cols(pool: &PgPool, account_id: Uuid) -> (i64, i64) {
+    let row = pool
+        .fetch_one(
+            sqlx::query("SELECT allocation_spent, allocation_reserved FROM accounts WHERE id = $1")
+                .bind(account_id),
+        )
+        .await
+        .unwrap();
+    (row.get("allocation_spent"), row.get("allocation_reserved"))
+}
+
+#[tokio::test]
+async fn concurrent_reserves_never_overshoot() {
+    let Some(pool) = pool_or_skip("concurrent_reserves_never_overshoot").await else {
+        return;
+    };
+    // Allocation admits exactly 5 reservations of 100 (cap 500).
+    let (account_id, key_id) = seed(&pool, 500).await;
+
+    let mut handles = Vec::new();
+    for _ in 0..20 {
+        let pool = pool.clone();
+        handles.push(tokio::spawn(async move {
+            ledger::reserve(&pool, account_id, key_id, 100).await
+        }));
+    }
+    let mut ok = 0;
+    let mut quota = 0;
+    for h in handles {
+        match h.await.unwrap() {
+            Ok(_) => ok += 1,
+            Err(LedgerError::InsufficientQuota { .. }) => quota += 1,
+            Err(e) => panic!("unexpected error: {e}"),
+        }
+    }
+    assert_eq!(ok, 5, "exactly 5 reserves of 100 fit in a 500 allocation");
+    assert_eq!(quota, 15);
+
+    let (spent, reserved) = account_cols(&pool, account_id).await;
+    assert_eq!(spent, 0);
+    assert_eq!(reserved, 500, "reserved exactly the cap, never over");
+}
+
+#[tokio::test]
+async fn settle_is_idempotent_and_reconciles_spend() {
+    let Some(pool) = pool_or_skip("settle_is_idempotent_and_reconciles_spend").await else {
+        return;
+    };
+    let (account_id, key_id) = seed(&pool, 1000).await;
+    let rid = ledger::reserve(&pool, account_id, key_id, 400)
+        .await
+        .unwrap();
+
+    // Settle actual=150 (< reserved 400): spent=150, reserved back to 0.
+    ledger::settle(&pool, rid, 150).await.unwrap();
+    let (spent, reserved) = account_cols(&pool, account_id).await;
+    assert_eq!((spent, reserved), (150, 0));
+
+    // Second settle is a no-op.
+    ledger::settle(&pool, rid, 999).await.unwrap();
+    let (spent2, reserved2) = account_cols(&pool, account_id).await;
+    assert_eq!((spent2, reserved2), (150, 0), "settle is idempotent");
+}
+
+#[tokio::test]
+async fn release_returns_reservation_and_is_idempotent() {
+    let Some(pool) = pool_or_skip("release_returns_reservation_and_is_idempotent").await else {
+        return;
+    };
+    let (account_id, key_id) = seed(&pool, 1000).await;
+    let rid = ledger::reserve(&pool, account_id, key_id, 300)
+        .await
+        .unwrap();
+    assert_eq!(account_cols(&pool, account_id).await, (0, 300));
+
+    ledger::release(&pool, rid).await.unwrap();
+    assert_eq!(account_cols(&pool, account_id).await, (0, 0));
+    // Idempotent; settle-after-release also a no-op.
+    ledger::release(&pool, rid).await.unwrap();
+    ledger::settle(&pool, rid, 100).await.unwrap();
+    assert_eq!(account_cols(&pool, account_id).await, (0, 0));
+}
+
+#[tokio::test]
+async fn hardcap_key_subcap_binds_below_account() {
+    let Some(pool) = pool_or_skip("hardcap_key_subcap_binds_below_account").await else {
+        return;
+    };
+    // Account has 1000 but the key is hard-capped at 200.
+    let user_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO users (email, password_hash, email_verified) \
+                 VALUES ($1, 'x', true) RETURNING id",
+            )
+            .bind(format!("u-{}@test.local", Uuid::new_v4())),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let account_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO accounts (owner_user_id, allocation_total) VALUES ($1, 1000) RETURNING id",
+            )
+            .bind(user_id),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let key_id: Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "INSERT INTO api_keys (account_id, key_hash, key_prefix, limit_kind, limit_value) \
+                 VALUES ($1, $2, 'sk-test', 'hardcap', 200) RETURNING id",
+            )
+            .bind(account_id)
+            .bind(Uuid::new_v4().as_bytes().to_vec()),
+        )
+        .await
+        .unwrap()
+        .get("id");
+
+    ledger::reserve(&pool, account_id, key_id, 200)
+        .await
+        .unwrap();
+    match ledger::reserve(&pool, account_id, key_id, 1).await {
+        Err(LedgerError::InsufficientQuota { available, .. }) => assert_eq!(available, 0),
+        other => panic!("expected InsufficientQuota, got {other:?}"),
+    }
+}

crates/helexa-upstream/tests/served_usage_pg.rs

@@ -0,0 +1,116 @@
+//! Integration test for the served-usage report (#58): the idempotent,
+//! monotonic upsert and the reconcile rollup. Gated on
+//! UPSTREAM_TEST_DATABASE_URL (skips cleanly when unset).
+
+use helexa_upstream::config::{ClientToken, UpstreamConfig};
+use helexa_upstream::db::connect_and_migrate;
+use helexa_upstream::email::EmailSender;
+use helexa_upstream::reconcile::reconcile;
+use helexa_upstream::state::AppState;
+use serde_json::{Value, json};
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+use uuid::Uuid;
+
+const CLIENT_TOKEN: &str = "su-test-token";
+const OPERATOR: &str = "op-su-test";
+
+async fn spawn_or_skip(test: &str) -> Option<(String, PgPool)> {
+    let Ok(url) = std::env::var("UPSTREAM_TEST_DATABASE_URL") else {
+        eprintln!("skipping {test}: UPSTREAM_TEST_DATABASE_URL not set");
+        return None;
+    };
+    let pool = connect_and_migrate(&url, 16).await.expect("migrate");
+    let mut config = UpstreamConfig {
+        server: Default::default(),
+        db: helexa_upstream::config::DbSettings {
+            url,
+            max_connections: 16,
+        },
+        grant: Default::default(),
+        abuse: Default::default(),
+        client_auth: Default::default(),
+        authz: Default::default(),
+        auth: Default::default(),
+        email: Default::default(),
+    };
+    config.client_auth.tokens.push(ClientToken {
+        token: CLIENT_TOKEN.into(),
+        operator_id: OPERATOR.into(),
+    });
+    let email = EmailSender::from_config(&config.email).unwrap();
+    let state = AppState::new(pool.clone(), config, email);
+    let app = helexa_upstream::build_app(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    Some((format!("http://{addr}"), pool))
+}
+
+async fn report(base: &str, rows: Value) -> u16 {
+    reqwest::Client::new()
+        .post(format!("{base}/authz/v1/served-usage"))
+        .bearer_auth(CLIENT_TOKEN)
+        .json(&json!({ "rows": rows }))
+        .send()
+        .await
+        .unwrap()
+        .status()
+        .as_u16()
+}
+
+async fn stored(pool: &PgPool, account: Uuid, key: Uuid) -> i64 {
+    sqlx::query(
+        "SELECT served_tokens FROM served_usage WHERE operator_id = $1 AND account_id = $2 AND key_id = $3",
+    )
+    .bind(OPERATOR)
+    .bind(account)
+    .bind(key)
+    .fetch_one(pool)
+    .await
+    .unwrap()
+    .get("served_tokens")
+}
+
+#[tokio::test]
+async fn served_usage_upsert_is_monotonic_and_reconciles() {
+    let Some((base, pool)) = spawn_or_skip("served_usage_upsert_is_monotonic_and_reconciles").await
+    else {
+        return;
+    };
+    let account = Uuid::new_v4();
+    let key = Uuid::new_v4();
+    let period = "2026-06-23";
+    let row = |n: i64| json!([{"account_id": account, "key_id": key, "period": period, "served_tokens": n}]);
+
+    // First report.
+    assert_eq!(report(&base, row(100)).await, 204);
+    assert_eq!(stored(&pool, account, key).await, 100);
+
+    // Re-send a higher absolute value → advances.
+    assert_eq!(report(&base, row(250)).await, 204);
+    assert_eq!(stored(&pool, account, key).await, 250);
+
+    // A lower value (e.g. a restarted cortex) must NOT regress (GREATEST).
+    assert_eq!(report(&base, row(50)).await, 204);
+    assert_eq!(stored(&pool, account, key).await, 250);
+
+    // Re-sending the same value is idempotent.
+    assert_eq!(report(&base, row(250)).await, 204);
+    assert_eq!(stored(&pool, account, key).await, 250);
+
+    // Reconcile rolls it up and stamps reconciled_at; a second run is empty.
+    let rollup = reconcile(&pool).await.unwrap();
+    let mine = rollup
+        .iter()
+        .find(|r| r.operator_id == OPERATOR)
+        .expect("operator in rollup");
+    assert!(mine.total_served_tokens >= 250);
+    let again = reconcile(&pool).await.unwrap();
+    assert!(
+        again.iter().all(|r| r.operator_id != OPERATOR),
+        "already reconciled"
+    );
+}

crates/helexa-upstream/tests/web_pg.rs

@@ -0,0 +1,425 @@
+//! Integration tests for the `/web/v1` account API + the silent fingerprint
+//! abuse policy, driving the built app over HTTP against a real Postgres.
+//! Gated on `UPSTREAM_TEST_DATABASE_URL` (skips cleanly when unset).
+
+use helexa_upstream::config::{ClientToken, UpstreamConfig};
+use helexa_upstream::crypto::sha256;
+use helexa_upstream::db::connect_and_migrate;
+use helexa_upstream::email::EmailSender;
+use helexa_upstream::state::AppState;
+use serde_json::{Value, json};
+use sqlx::Executor;
+use sqlx::Row;
+use sqlx::postgres::PgPool;
+
+const CLIENT_TOKEN: &str = "web-test-operator-token";
+
+async fn spawn_or_skip(test: &str) -> Option<(String, PgPool)> {
+    let Ok(url) = std::env::var("UPSTREAM_TEST_DATABASE_URL") else {
+        eprintln!("skipping {test}: UPSTREAM_TEST_DATABASE_URL not set");
+        return None;
+    };
+    let pool = connect_and_migrate(&url, 16).await.expect("migrate");
+    let mut config = UpstreamConfig {
+        server: Default::default(),
+        db: helexa_upstream::config::DbSettings {
+            url,
+            max_connections: 16,
+        },
+        grant: Default::default(),
+        abuse: Default::default(),
+        client_auth: Default::default(),
+        authz: Default::default(),
+        auth: Default::default(),
+        email: Default::default(), // Log transport
+    };
+    config.client_auth.tokens.push(ClientToken {
+        token: CLIENT_TOKEN.into(),
+        operator_id: "op-web".into(),
+    });
+    let email = EmailSender::from_config(&config.email).unwrap();
+    let state = AppState::new(pool.clone(), config, email);
+    let app = helexa_upstream::build_app(state);
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+    let addr = listener.local_addr().unwrap();
+    tokio::spawn(async move {
+        axum::serve(listener, app).await.unwrap();
+    });
+    Some((format!("http://{addr}"), pool))
+}
+
+fn unique_email() -> String {
+    format!("u-{}@test.local", uuid::Uuid::new_v4())
+}
+
+async fn post(url: String, body: Value, bearer: Option<&str>) -> reqwest::Response {
+    let c = reqwest::Client::new();
+    let mut req = c.post(url).json(&body);
+    if let Some(b) = bearer {
+        req = req.bearer_auth(b);
+    }
+    req.send().await.unwrap()
+}
+
+#[tokio::test]
+async fn verify_endpoint_consumes_token_once() {
+    let Some((base, pool)) = spawn_or_skip("verify_endpoint_consumes_token_once").await else {
+        return;
+    };
+    let email = unique_email();
+    // Register, then mint a verify token directly (the raw token is only in
+    // the email; here we insert a known one to drive the endpoint).
+    assert_eq!(
+        post(
+            format!("{base}/web/v1/register"),
+            json!({"email": email, "password": "password123"}),
+            None
+        )
+        .await
+        .status(),
+        202
+    );
+    let user_id: uuid::Uuid = pool
+        .fetch_one(sqlx::query("SELECT id FROM users WHERE email = $1").bind(&email))
+        .await
+        .unwrap()
+        .get("id");
+    let raw = "verify-raw-token-xyz";
+    pool.execute(
+        sqlx::query(
+            "INSERT INTO email_tokens (token_hash, user_id, kind, expires_at) \
+             VALUES ($1, $2, 'verify', now() + interval '1 hour')",
+        )
+        .bind(sha256(raw))
+        .bind(user_id),
+    )
+    .await
+    .unwrap();
+
+    assert_eq!(
+        post(format!("{base}/web/v1/verify"), json!({"token": raw}), None)
+            .await
+            .status(),
+        200
+    );
+    // Consumed → second attempt fails.
+    assert_eq!(
+        post(format!("{base}/web/v1/verify"), json!({"token": raw}), None)
+            .await
+            .status(),
+        400
+    );
+
+    let verified: bool = pool
+        .fetch_one(sqlx::query("SELECT email_verified FROM users WHERE id = $1").bind(user_id))
+        .await
+        .unwrap()
+        .get("email_verified");
+    assert!(verified);
+}
+
+#[tokio::test]
+async fn account_lifecycle_and_key_resolves_then_archives() {
+    let Some((base, pool)) =
+        spawn_or_skip("account_lifecycle_and_key_resolves_then_archives").await
+    else {
+        return;
+    };
+    let email = unique_email();
+    post(
+        format!("{base}/web/v1/register"),
+        json!({"email": email, "password": "password123"}),
+        None,
+    )
+    .await;
+    // Bypass the email step for the login/key portion.
+    pool.execute(
+        sqlx::query("UPDATE users SET email_verified = true WHERE email = $1").bind(&email),
+    )
+    .await
+    .unwrap();
+
+    // login → session JWT
+    let r = post(
+        format!("{base}/web/v1/login"),
+        json!({"email": email, "password": "password123"}),
+        None,
+    )
+    .await;
+    assert_eq!(r.status(), 200);
+    let token = r.json::<Value>().await.unwrap()["token"]
+        .as_str()
+        .unwrap()
+        .to_string();
+
+    // create key (raw shown once)
+    let r = post(
+        format!("{base}/web/v1/keys"),
+        json!({"label": "laptop"}),
+        Some(&token),
+    )
+    .await;
+    assert_eq!(r.status(), 201);
+    let body: Value = r.json().await.unwrap();
+    let raw_key = body["key"].as_str().unwrap().to_string();
+    let key_id = body["id"].as_str().unwrap().to_string();
+    assert!(raw_key.starts_with("sk-helexa-"));
+
+    // account balance reflects the free grant
+    let r = reqwest::Client::new()
+        .get(format!("{base}/web/v1/account"))
+        .bearer_auth(&token)
+        .send()
+        .await
+        .unwrap();
+    assert_eq!(
+        r.json::<Value>().await.unwrap()["allocation_total"],
+        1_000_000
+    );
+
+    // list keys shows the prefix, never the raw secret
+    let r = reqwest::Client::new()
+        .get(format!("{base}/web/v1/keys"))
+        .bearer_auth(&token)
+        .send()
+        .await
+        .unwrap();
+    let listed = r.json::<Value>().await.unwrap();
+    let k = &listed["keys"][0];
+    assert_eq!(k["id"], key_id);
+    assert!(k.get("key").is_none(), "raw secret never listed");
+    assert!(k["prefix"].as_str().unwrap().starts_with("sk-helexa-"));
+
+    // the key authorizes at the authz surface
+    let r = post(
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw_key}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 200);
+
+    // archive → the key no longer resolves
+    let r = post(
+        format!("{base}/web/v1/keys/{key_id}/archive"),
+        json!({}),
+        Some(&token),
+    )
+    .await;
+    assert_eq!(r.status(), 204);
+    let r = post(
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw_key}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(r.status(), 401);
+}
+
+#[tokio::test]
+async fn fingerprint_abuse_silently_deactivates_all_no_clue() {
+    let Some((base, pool)) =
+        spawn_or_skip("fingerprint_abuse_silently_deactivates_all_no_clue").await
+    else {
+        return;
+    };
+    let fp = format!("fp-{}", uuid::Uuid::new_v4());
+
+    // 5 registrations sharing one fingerprint — every one returns a normal 202.
+    let mut emails = Vec::new();
+    for _ in 0..5 {
+        let email = unique_email();
+        let r = post(
+            format!("{base}/web/v1/register"),
+            json!({"email": email, "password": "password123", "fingerprint": fp}),
+            None,
+        )
+        .await;
+        assert_eq!(r.status(), 202, "registration always looks successful");
+        emails.push(email);
+    }
+
+    // Silent effect: all 5 accounts are deactivated + flagged.
+    let (deactivated, flagged): (i64, i64) = {
+        let row = pool
+            .fetch_one(
+                sqlx::query(
+                    "SELECT \
+                   count(*) FILTER (WHERE a.status = 'deactivated') AS d, \
+                   count(*) FILTER (WHERE a.fingerprint_flagged) AS f \
+                 FROM accounts a JOIN users u ON u.id = a.owner_user_id \
+                 WHERE u.registration_fingerprint = $1",
+                )
+                .bind(&fp),
+            )
+            .await
+            .unwrap();
+        (row.get("d"), row.get("f"))
+    };
+    assert_eq!(deactivated, 5, "all sharing accounts silently deactivated");
+    assert_eq!(flagged, 5);
+
+    // No clue at the authz surface: a key on a deactivated account resolves
+    // as an ordinary 401, indistinguishable from an unknown key.
+    let acct: uuid::Uuid = pool
+        .fetch_one(
+            sqlx::query(
+                "SELECT a.id FROM accounts a JOIN users u ON u.id = a.owner_user_id \
+                 WHERE u.registration_fingerprint = $1 LIMIT 1",
+            )
+            .bind(&fp),
+        )
+        .await
+        .unwrap()
+        .get("id");
+    let raw = "sk-helexa-deactivated-probe";
+    pool.execute(
+        sqlx::query(
+            "INSERT INTO api_keys (account_id, key_hash, key_prefix) VALUES ($1, $2, 'sk-helexa-')",
+        )
+        .bind(acct)
+        .bind(sha256(raw)),
+    )
+    .await
+    .unwrap();
+    let r = post(
+        format!("{base}/authz/v1/resolve"),
+        json!({"api_key": raw}),
+        Some(CLIENT_TOKEN),
+    )
+    .await;
+    assert_eq!(
+        r.status(),
+        401,
+        "deactivated account's key looks like any invalid key"
+    );
+}
+
+#[tokio::test]
+async fn topup_redeem_raises_allocation_single_use() {
+    let Some((base, pool)) = spawn_or_skip("topup_redeem_raises_allocation_single_use").await
+    else {
+        return;
+    };
+    let email = unique_email();
+    post(
+        format!("{base}/web/v1/register"),
+        json!({"email": email, "password": "password123"}),
+        None,
+    )
+    .await;
+    pool.execute(
+        sqlx::query("UPDATE users SET email_verified = true WHERE email = $1").bind(&email),
+    )
+    .await
+    .unwrap();
+    let token = post(
+        format!("{base}/web/v1/login"),
+        json!({"email": email, "password": "password123"}),
+        None,
+    )
+    .await
+    .json::<Value>()
+    .await
+    .unwrap()["token"]
+        .as_str()
+        .unwrap()
+        .to_string();
+
+    // Mint a code worth 500_000 (mint path used by the CLI/faucet).
+    let codes = helexa_upstream::topup::mint(&pool, 500_000, 1, Some("test"))
+        .await
+        .unwrap();
+    let code = &codes[0];
+
+    // Redeem → allocation_total rises from the 1_000_000 free grant.
+    let r = post(
+        format!("{base}/web/v1/redeem"),
+        json!({"code": code}),
+        Some(&token),
+    )
+    .await;
+    assert_eq!(r.status(), 200);
+    assert_eq!(
+        r.json::<Value>().await.unwrap()["allocation_total"],
+        1_500_000
+    );
+
+    // Single-use: a second redemption fails generically (no oracle).
+    let r = post(
+        format!("{base}/web/v1/redeem"),
+        json!({"code": code}),
+        Some(&token),
+    )
+    .await;
+    assert_eq!(r.status(), 400);
+
+    // Unknown code: same generic 400.
+    let r = post(
+        format!("{base}/web/v1/redeem"),
+        json!({"code": "helexa-topup-does-not-exist"}),
+        Some(&token),
+    )
+    .await;
+    assert_eq!(r.status(), 400);
+}
+
+#[tokio::test]
+async fn topup_concurrent_double_redeem_one_winner() {
+    let Some((base, pool)) = spawn_or_skip("topup_concurrent_double_redeem_one_winner").await
+    else {
+        return;
+    };
+    // Two verified accounts.
+    let mut tokens = Vec::new();
+    for _ in 0..2 {
+        let email = unique_email();
+        post(
+            format!("{base}/web/v1/register"),
+            json!({"email": email, "password": "password123"}),
+            None,
+        )
+        .await;
+        pool.execute(
+            sqlx::query("UPDATE users SET email_verified = true WHERE email = $1").bind(&email),
+        )
+        .await
+        .unwrap();
+        let t = post(
+            format!("{base}/web/v1/login"),
+            json!({"email": email, "password": "password123"}),
+            None,
+        )
+        .await
+        .json::<Value>()
+        .await
+        .unwrap()["token"]
+            .as_str()
+            .unwrap()
+            .to_string();
+        tokens.push(t);
+    }
+    let code = helexa_upstream::topup::mint(&pool, 100, 1, None)
+        .await
+        .unwrap()
+        .remove(0);
+
+    // Both accounts race to redeem the same code; exactly one wins.
+    let (a, b) = tokio::join!(
+        post(
+            format!("{base}/web/v1/redeem"),
+            json!({"code": code}),
+            Some(&tokens[0])
+        ),
+        post(
+            format!("{base}/web/v1/redeem"),
+            json!({"code": code}),
+            Some(&tokens[1])
+        ),
+    );
+    let wins = [a.status(), b.status()]
+        .iter()
+        .filter(|s| s.as_u16() == 200)
+        .count();
+    assert_eq!(wins, 1, "exactly one redemption wins the single-use code");
+}

crates/neuron/Cargo.toml

@@ -38,6 +38,7 @@ cudnn = [
 flash-attn = [
     "cuda",
     "candle-transformers/flash-attn",
+    "dep:candle-flash-attn",
 ]
 # Reserved for GPU-only integration tests in later stages.
 cuda-integration = ["cuda"]
@@ -71,6 +72,10 @@ rayon = "1"
 candle-core = "0.10.2"
 candle-nn = "0.10.2"
 candle-transformers = "0.10.2"
+# Direct dependency so `candle_flash_attn::flash_attn` is nameable in
+# the attention core (#95); the candle-transformers feature alone only
+# enables flash inside ITS models.
+candle-flash-attn = { version = "0.10.2", optional = true }
 # Direct dep on cudarc (matching candle's transitive version) so the
 # TP worker pool can call cudarc::nccl::{Comm, Id} directly. Gated on
 # the `cuda` feature; same toolchain requirement as candle's CUDA path.

crates/neuron/src/api.rs

@@ -13,9 +13,10 @@ use axum::response::sse::{Event, KeepAlive, Sse};
 use axum::response::{IntoResponse, Json};
 use axum::routing::{get, post};
 use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
+use cortex_core::entitlements::{HEADER_ACCOUNT_ID, HEADER_KEY_ID};
 use cortex_core::harness::ModelSpec;
 use cortex_core::openai::{ChatCompletionRequest, MessageContent};
-use cortex_core::responses::{ResponsesRequest, ResponsesUsage};
+use cortex_core::responses::{OutputTokensDetails, ResponsesRequest, ResponsesUsage};
 use futures::stream::{self, StreamExt};
 use serde_json::{Value, json};
 use std::convert::Infallible;
@@ -71,6 +72,12 @@ async fn health_handler(State(state): State<Arc<NeuronState>>) -> Json<HealthRes
     // know about activation lifecycle.
     let mut snapshot = state.health_cache.snapshot().await;
     snapshot.activation = state.activation.snapshot().await;
+    // Per-model admission load (#53) — read live from the candle harness so
+    // cortex's load-aware router (#55) can spread traffic and propagate
+    // backpressure. Absent when no candle harness is present.
+    if let Some(candle) = &state.candle {
+        snapshot.models = candle.load_snapshot().await;
+    }
     Json(snapshot)
 }
 
@@ -228,6 +235,17 @@ fn default_enable_thinking(req: &mut ChatCompletionRequest, include_thinking: bo
     }
 }
 
+/// The request's principal for fair-share admission (#54), reconstructed
+/// from the internal headers cortex stamps (#49). cortex strips any
+/// client-supplied copy and asserts the authoritative value, so over the
+/// trusted WireGuard link these are safe to key fair-share on. `None` for an
+/// unauthenticated/direct request — exempt from the per-principal cap.
+fn principal_key(headers: &axum::http::HeaderMap) -> Option<String> {
+    let account = headers.get(HEADER_ACCOUNT_ID)?.to_str().ok()?;
+    let key = headers.get(HEADER_KEY_ID)?.to_str().ok()?;
+    Some(format!("{account}/{key}"))
+}
+
 /// OpenAI-compatible chat completions. Dispatches to streaming SSE when
 /// `stream: true` is set on the request; otherwise returns a single
 /// `ChatCompletionResponse`.
@@ -271,8 +289,14 @@ async fn chat_completions(
     // true`) keep reasoning on.
     default_enable_thinking(&mut req, include_thinking);
 
+    // Fair-share admission principal (#54), from cortex's stamped headers.
+    let principal = principal_key(&headers);
+
     if req.stream.unwrap_or(false) {
-        match candle.chat_completion_stream_with(req, chat_config).await {
+        match candle
+            .chat_completion_stream_with(req, chat_config, principal)
+            .await
+        {
             Ok(rx) => {
                 // Each chunk → one SSE `data: {json}` line. After the
                 // channel closes, append the OpenAI [DONE] terminator.
@@ -289,7 +313,7 @@ async fn chat_completions(
             Err(e) => inference_error_response(e),
         }
     } else {
-        match candle.chat_completion(req).await {
+        match candle.chat_completion(req, principal).await {
             Ok(resp) => Json(resp).into_response(),
             Err(e) => inference_error_response(e),
         }
@@ -302,6 +326,7 @@ async fn chat_completions(
 /// event stream into the Responses event family.
 async fn responses(
     State(state): State<Arc<NeuronState>>,
+    headers: axum::http::HeaderMap,
     Json(req): Json<ResponsesRequest>,
 ) -> impl IntoResponse {
     let Some(candle) = state.candle.as_ref().map(Arc::clone) else {
@@ -336,9 +361,12 @@ async fn responses(
     };
     chat_req.stream = Some(stream_requested);
 
+    // Fair-share admission principal (#54), from cortex's stamped headers.
+    let principal = principal_key(&headers);
+
     if stream_requested {
         match candle
-            .responses_stream(chat_req, response_id, message_item_id)
+            .responses_stream(chat_req, response_id, message_item_id, principal)
             .await
         {
             Ok(rx) => {
@@ -362,7 +390,7 @@ async fn responses(
         // and translate the result. We don't currently re-tokenise
         // to compute usage; the harness returns it via the chat
         // response and we pass it through.
-        match candle.chat_completion(chat_req).await {
+        match candle.chat_completion(chat_req, principal).await {
             Ok(chat_resp) => {
                 // Extract the assistant text (chat completions
                 // always emits one choice on the candle path).
@@ -390,8 +418,14 @@ async fn responses(
                     input_tokens: u.prompt_tokens,
                     output_tokens: u.completion_tokens,
                     total_tokens: u.prompt_tokens + u.completion_tokens,
-                    // Non-streaming reasoning accounting deferred (#64).
+                    // Carry the reasoning sub-count through from the chat
-                    output_tokens_details: None,
+                    // usage — the non-streaming path now splits off the
+                    // `<think>` span and counts it (see `split_off_reasoning`).
+                    output_tokens_details: u.completion_tokens_details.as_ref().map(|d| {
+                        OutputTokensDetails {
+                            reasoning_tokens: d.reasoning_tokens,
+                        }
+                    }),
                     input_tokens_details: None,
                 });
                 let meta = openai_responses::ResponseMeta {
@@ -486,6 +520,24 @@ fn inference_error_response(err: InferenceError) -> axum::response::Response {
             "template_render_failed",
             format!("chat template could not render this request: {detail}"),
         ),
+        // Admission control refused on load (#53): a fast, retryable "busy"
+        // signal. 503 (service busy) + Retry-After; opencode/AI SDK back off.
+        InferenceError::Overloaded { retry_after_secs } => OpenAiError::new(
+            503,
+            "rate_limit_error",
+            "rate_limit_exceeded",
+            "model is busy (admission queue full); retry shortly",
+        )
+        .with_retry_after(retry_after_secs),
+        // Per-principal fair-share cap (#54): 429 rate_limit_exceeded +
+        // Retry-After — the caller is sending too many concurrent requests.
+        InferenceError::PerPrincipalLimit { retry_after_secs } => OpenAiError::new(
+            429,
+            "rate_limit_error",
+            "rate_limit_exceeded",
+            "too many concurrent requests for this key; retry shortly",
+        )
+        .with_retry_after(retry_after_secs),
         InferenceError::Other(e) => OpenAiError::without_code(500, "api_error", format!("{e:#}")),
     };
     envelope_response(env)
@@ -660,6 +712,26 @@ mod error_envelope_tests {
         assert_eq!(error["required_mb"], 8_192);
     }
 
+    #[tokio::test]
+    async fn overloaded_is_503_rate_limited_with_retry_after() {
+        // Admission rejection (#53) → fast, retryable backpressure.
+        let resp = inference_error_response(InferenceError::Overloaded {
+            retry_after_secs: 7,
+        });
+        assert_eq!(resp.status(), StatusCode::SERVICE_UNAVAILABLE);
+        let retry = resp
+            .headers()
+            .get(axum::http::header::RETRY_AFTER)
+            .expect("admission rejection must advertise Retry-After");
+        assert_eq!(retry.to_str().unwrap(), "7");
+
+        let bytes = axum::body::to_bytes(resp.into_body(), usize::MAX)
+            .await
+            .unwrap();
+        let body: Value = serde_json::from_slice(&bytes).unwrap();
+        assert_eq!(body["error"]["code"], "rate_limit_exceeded");
+    }
+
     #[tokio::test]
     async fn insufficient_vram_carries_retry_after() {
         // Transient 503 — VRAM frees as in-flight requests finish, so the

crates/neuron/src/config.rs

@@ -85,6 +85,68 @@ pub struct CandleHarnessConfig {
     /// `/models`, and enforces it. These knobs tune that derivation.
     #[serde(default)]
     pub context_limit: ContextLimitConfig,
+
+    /// Admission control (#53): bounds the per-model wait queue so a busy
+    /// model returns a fast, retryable `429`/`503` instead of stalling new
+    /// requests until their client times out.
+    #[serde(default)]
+    pub admission: AdmissionConfig,
+}
+
+/// `[harness.candle.admission]` settings (#53).
+///
+/// Inference is batch-1, so `max_in_flight` is 1 in practice; the queue
+/// (`max_queue_depth`) absorbs short bursts, and `max_wait_secs` caps how
+/// long a queued request waits before it's refused with backpressure.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct AdmissionConfig {
+    /// Concurrent running requests per model. Batch-1 inference → 1.
+    #[serde(default = "default_admission_max_in_flight")]
+    pub max_in_flight: usize,
+    /// Queued (waiting) requests allowed beyond the in-flight one. The
+    /// `(max_in_flight + max_queue_depth + 1)`-th request is refused
+    /// immediately with `429`/`503` + `Retry-After`.
+    #[serde(default = "default_admission_max_queue_depth")]
+    pub max_queue_depth: usize,
+    /// Maximum seconds a queued request waits for the in-flight slot before
+    /// it is refused (turns the old ~300s client-side hang into a fast,
+    /// honest signal).
+    #[serde(default = "default_admission_max_wait_secs")]
+    pub max_wait_secs: u64,
+    /// Per-principal fair-share cap (#54): max in-flight + queued requests
+    /// for any single principal (resolved from the `x-helexa-*` headers
+    /// cortex stamps), so one client can't monopolize the queue while others
+    /// wait. Over-cap → `429 rate_limit_exceeded` + `Retry-After`. `0`
+    /// disables the cap; anonymous requests are always exempt.
+    #[serde(default = "default_admission_max_per_principal")]
+    pub max_per_principal: usize,
+}
+
+impl Default for AdmissionConfig {
+    fn default() -> Self {
+        Self {
+            max_in_flight: default_admission_max_in_flight(),
+            max_queue_depth: default_admission_max_queue_depth(),
+            max_wait_secs: default_admission_max_wait_secs(),
+            max_per_principal: default_admission_max_per_principal(),
+        }
+    }
+}
+
+fn default_admission_max_in_flight() -> usize {
+    1
+}
+
+fn default_admission_max_queue_depth() -> usize {
+    8
+}
+
+fn default_admission_max_wait_secs() -> u64 {
+    30
+}
+
+fn default_admission_max_per_principal() -> usize {
+    2
 }
 
 /// `[harness.candle.prefix_cache]` settings.

crates/neuron/src/harness/admission.rs

@@ -0,0 +1,373 @@
+//! Per-model admission control (#53).
+//!
+//! Inference against a loaded model is batch-1: one request runs at a time,
+//! serialized by the model's `inference_lock` (single-GPU) / `pool` mutex
+//! (TP). Before this, the wait for that lock was an **unbounded FIFO of
+//! mutex waiters with no timeout** — a busy model made every new request
+//! hang until its client gave up (~300s) with an opaque error.
+//!
+//! [`AdmissionController`] replaces that implicit unbounded wait with an
+//! explicit bounded scheduler: at most `max_in_flight` running (1, batch-1)
+//! plus a bounded queue of `max_queue_depth` waiters, each waiting at most
+//! `max_wait`. When the queue is full or the wait elapses, the request is
+//! rejected *immediately* — an honest, fast, retryable "busy" signal
+//! (`429`/`503` + `Retry-After` per #63) instead of a silent stall.
+//!
+//! The controller is pure async (no CUDA), so the inference paths just call
+//! [`AdmissionController::enter`] before taking the inference lock and hold
+//! the returned [`AdmissionPermit`] for the request's lifetime. Its counters
+//! ([`in_flight`](AdmissionController::in_flight) /
+//! [`queue_depth`](AdmissionController::queue_depth)) are lock-free, so
+//! `/health` can read live load without contending with inference.
+
+use crate::config::AdmissionConfig;
+use std::collections::HashMap;
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+use tokio::sync::{OwnedSemaphorePermit, Semaphore};
+
+/// Why admission was refused. All map to the #63 backpressure envelope
+/// (`rate_limit_exceeded` + `Retry-After`); they differ in cause (and HTTP
+/// status — load → `503`, per-principal → `429`).
+#[derive(Debug, Clone, Copy)]
+pub enum AdmissionRejection {
+    /// The bounded wait queue was already full (server-side load).
+    QueueFull { retry_after_secs: u64 },
+    /// A queue slot was taken but the in-flight slot didn't free within
+    /// `max_wait` (server-side load).
+    Timeout { retry_after_secs: u64 },
+    /// This principal already has `max_per_principal` requests in flight or
+    /// queued (#54 fair-share) — one principal can't monopolize the model.
+    PrincipalCap { retry_after_secs: u64 },
+}
+
+impl AdmissionRejection {
+    pub fn retry_after_secs(&self) -> u64 {
+        match self {
+            AdmissionRejection::QueueFull { retry_after_secs }
+            | AdmissionRejection::Timeout { retry_after_secs }
+            | AdmissionRejection::PrincipalCap { retry_after_secs } => *retry_after_secs,
+        }
+    }
+}
+
+/// Admission accounting, mutated under a brief lock (never held across an
+/// await). `pending` is queued + in-flight overall; `per_principal` is the
+/// same count keyed by principal for fair-share (#54).
+#[derive(Default, Debug)]
+struct AdmissionState {
+    pending: usize,
+    per_principal: HashMap<String, usize>,
+}
+
+/// Bounded batch-1 scheduler for one loaded model, with per-principal
+/// fair-share.
+pub struct AdmissionController {
+    /// In-flight slots — `max_in_flight` permits (1 for batch-1).
+    slots: Arc<Semaphore>,
+    /// Queued + in-flight accounting (overall + per principal).
+    state: Arc<Mutex<AdmissionState>>,
+    /// `max_in_flight + max_queue_depth` — the overall rejection threshold.
+    max_pending: usize,
+    /// Max in-flight + queued for any single principal (#54). `0` disables.
+    max_per_principal: usize,
+    max_in_flight: usize,
+    max_wait: Duration,
+}
+
+impl AdmissionController {
+    pub fn new(cfg: &AdmissionConfig) -> Self {
+        // A controller with zero in-flight slots would deadlock; clamp.
+        let max_in_flight = cfg.max_in_flight.max(1);
+        Self {
+            slots: Arc::new(Semaphore::new(max_in_flight)),
+            state: Arc::new(Mutex::new(AdmissionState::default())),
+            max_pending: max_in_flight + cfg.max_queue_depth,
+            max_per_principal: cfg.max_per_principal,
+            max_in_flight,
+            max_wait: Duration::from_secs(cfg.max_wait_secs),
+        }
+    }
+
+    /// Admit a request for `principal` (`None` = anonymous, exempt from the
+    /// per-principal cap). Reserves a queue slot — fast-rejecting if the
+    /// overall queue is full or the principal is over its fair-share cap —
+    /// then waits up to `max_wait` for an in-flight slot. The returned permit
+    /// must be held for the request's lifetime; dropping it frees the slots.
+    ///
+    /// CANCELLATION SAFETY: the semaphore wait below is where a client
+    /// disconnect lands — axum drops the request future mid-await. The
+    /// reservation therefore lives in a RAII [`PendingReservation`] taken
+    /// BEFORE the await: if this future is dropped while queued, the
+    /// guard's Drop rolls the counts back. (The original version
+    /// incremented raw counters and only decremented on the timeout
+    /// branch — every abandoned wait leaked a `pending` + per-principal
+    /// slot, ratcheting the model into a permanent instant-429 state
+    /// under client retry storms. Observed live 2026-07-02:
+    /// `queue_depth: 1` pinned on an idle model.)
+    pub async fn enter(
+        &self,
+        principal: Option<&str>,
+    ) -> Result<AdmissionPermit, AdmissionRejection> {
+        // Decision + reservation under one brief lock so concurrent callers
+        // can't both slip past the thresholds. No await is held here.
+        let reservation = {
+            let mut st = self.state.lock().expect("admission state poisoned");
+            if st.pending >= self.max_pending {
+                return Err(AdmissionRejection::QueueFull {
+                    retry_after_secs: self.retry_hint(st.pending),
+                });
+            }
+            if let Some(p) = principal
+                && self.max_per_principal > 0
+                && st.per_principal.get(p).copied().unwrap_or(0) >= self.max_per_principal
+            {
+                return Err(AdmissionRejection::PrincipalCap {
+                    retry_after_secs: self.retry_hint(st.pending),
+                });
+            }
+            st.pending += 1;
+            if let Some(p) = principal {
+                *st.per_principal.entry(p.to_string()).or_insert(0) += 1;
+            }
+            PendingReservation {
+                state: Arc::clone(&self.state),
+                principal: principal.map(str::to_string),
+            }
+        };
+
+        match tokio::time::timeout(self.max_wait, Arc::clone(&self.slots).acquire_owned()).await {
+            Ok(Ok(permit)) => Ok(AdmissionPermit {
+                _permit: permit,
+                _reservation: reservation,
+            }),
+            // Semaphore is never closed; treat a closed/elapsed wait the
+            // same. `reservation` drops here, rolling back the counts.
+            Ok(Err(_)) | Err(_) => Err(AdmissionRejection::Timeout {
+                retry_after_secs: self.retry_hint(self.max_pending),
+            }),
+        }
+    }
+
+    /// Requests currently running (holding an in-flight slot).
+    pub fn in_flight(&self) -> usize {
+        self.max_in_flight
+            .saturating_sub(self.slots.available_permits())
+    }
+
+    /// Requests waiting for an in-flight slot.
+    pub fn queue_depth(&self) -> usize {
+        let pending = self.state.lock().expect("admission state poisoned").pending;
+        pending.saturating_sub(self.in_flight())
+    }
+
+    /// Rough `Retry-After`: scale with how backed-up the model is, clamped to
+    /// a sane band. Without per-request timing this is a heuristic, but it
+    /// gives well-behaved clients (opencode/AI SDK) a sensible backoff.
+    fn retry_hint(&self, pending: usize) -> u64 {
+        let queued = pending.saturating_sub(self.max_in_flight) as u64;
+        ((queued + 1) * 2).clamp(1, 120)
+    }
+}
+
+/// Decrement (and prune at zero) a principal's outstanding count.
+fn decrement_principal(map: &mut HashMap<String, usize>, principal: Option<&str>) {
+    if let Some(p) = principal
+        && let Some(count) = map.get_mut(p)
+    {
+        *count -= 1;
+        if *count == 0 {
+            map.remove(p);
+        }
+    }
+}
+
+/// RAII accounting for one reserved slot (queued or in-flight): decrements
+/// `pending` and the principal's fair-share count on drop, whichever way
+/// the reservation ends — admitted-and-finished, wait timeout, or the
+/// caller's future being dropped mid-queue (client disconnect).
+#[derive(Debug)]
+struct PendingReservation {
+    state: Arc<Mutex<AdmissionState>>,
+    principal: Option<String>,
+}
+
+impl Drop for PendingReservation {
+    fn drop(&mut self) {
+        let mut st = self.state.lock().expect("admission state poisoned");
+        st.pending = st.pending.saturating_sub(1);
+        decrement_principal(&mut st.per_principal, self.principal.as_deref());
+    }
+}
+
+/// Held for a request's lifetime; frees the in-flight slot (semaphore
+/// permit) and the queue + fair-share accounting (reservation) on drop.
+#[derive(Debug)]
+pub struct AdmissionPermit {
+    _permit: OwnedSemaphorePermit,
+    _reservation: PendingReservation,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Config with the per-principal cap disabled (0) — most tests exercise
+    /// the overall queue with anonymous (`None`) callers.
+    fn cfg(max_in_flight: usize, max_queue_depth: usize, max_wait_secs: u64) -> AdmissionConfig {
+        AdmissionConfig {
+            max_in_flight,
+            max_queue_depth,
+            max_wait_secs,
+            max_per_principal: 0,
+        }
+    }
+
+    #[tokio::test]
+    async fn admits_up_to_in_flight_and_reports_load() {
+        let ctrl = AdmissionController::new(&cfg(1, 4, 30));
+        assert_eq!(ctrl.in_flight(), 0);
+        let p = ctrl.enter(None).await.expect("first admits");
+        assert_eq!(ctrl.in_flight(), 1);
+        assert_eq!(ctrl.queue_depth(), 0);
+        drop(p);
+        assert_eq!(ctrl.in_flight(), 0);
+    }
+
+    #[tokio::test]
+    async fn rejects_when_queue_full() {
+        // 1 in-flight + 1 queue slot = capacity 2; the 3rd is refused fast.
+        let ctrl = Arc::new(AdmissionController::new(&cfg(1, 1, 30)));
+        let _running = ctrl.enter(None).await.expect("admit running");
+
+        // Fill the single queue slot with a waiter that parks on the semaphore.
+        let ctrl2 = Arc::clone(&ctrl);
+        let waiter = tokio::spawn(async move { ctrl2.enter(None).await.map(|p| drop(p)) });
+        // Give the waiter a moment to occupy the queue slot.
+        tokio::time::sleep(Duration::from_millis(50)).await;
+        assert_eq!(ctrl.queue_depth(), 1);
+
+        // Queue full → immediate QueueFull with a Retry-After hint.
+        match ctrl.enter(None).await {
+            Err(AdmissionRejection::QueueFull { retry_after_secs }) => {
+                assert!(retry_after_secs >= 1)
+            }
+            other => panic!("expected QueueFull, got {other:?}"),
+        }
+
+        // Release the runner so the parked waiter can proceed and finish.
+        drop(_running);
+        waiter.await.unwrap().unwrap();
+    }
+
+    #[tokio::test]
+    async fn rejects_on_wait_timeout() {
+        // Zero queue depth + a runner holding the only slot → a second
+        // request can't even queue, so it's QueueFull, not Timeout. Use a
+        // queue of 1 and a tiny max_wait to exercise the timeout path.
+        let ctrl = Arc::new(AdmissionController::new(&cfg(1, 1, 0)));
+        let _running = ctrl.enter(None).await.expect("admit running");
+        // max_wait 0 → the queued request times out almost immediately.
+        match ctrl.enter(None).await {
+            Err(AdmissionRejection::Timeout { .. }) => {}
+            other => panic!("expected Timeout, got {other:?}"),
+        }
+        // The timed-out request released its queue slot.
+        assert_eq!(ctrl.queue_depth(), 0);
+    }
+
+    #[tokio::test]
+    async fn per_principal_cap_protects_other_principals() {
+        // Generous overall queue, but each principal capped at 1 in-flight+
+        // queued. Principal A holds the running slot; A's second request is
+        // refused (PrincipalCap) rather than occupying the queue, so B's
+        // single request still gets a queue slot and proceeds.
+        let cfg = AdmissionConfig {
+            max_in_flight: 1,
+            max_queue_depth: 8,
+            max_wait_secs: 30,
+            max_per_principal: 1,
+        };
+        let ctrl = Arc::new(AdmissionController::new(&cfg));
+
+        let _a1 = ctrl.enter(Some("acct-a/key-a")).await.expect("A admits");
+
+        // A is over its fair-share cap → fast PrincipalCap, no queue slot taken.
+        match ctrl.enter(Some("acct-a/key-a")).await {
+            Err(AdmissionRejection::PrincipalCap { retry_after_secs }) => {
+                assert!(retry_after_secs >= 1)
+            }
+            other => panic!("expected PrincipalCap, got {other:?}"),
+        }
+
+        // B (a different principal) is admitted to the queue and proceeds
+        // once A releases — it was never stuck behind A's backlog.
+        let ctrl2 = Arc::clone(&ctrl);
+        let b = tokio::spawn(async move { ctrl2.enter(Some("acct-b/key-b")).await.map(drop) });
+        tokio::time::sleep(Duration::from_millis(50)).await;
+        assert_eq!(ctrl.queue_depth(), 1, "B is queued, not rejected");
+        drop(_a1);
+        b.await.unwrap().expect("B is served after A releases");
+    }
+
+    /// Regression for the 2026-07-02 retry-storm incident: a client that
+    /// disconnects while QUEUED drops the `enter()` future mid-await.
+    /// The reservation must roll back — the original implementation
+    /// leaked `pending` + the per-principal count on this path, pinning
+    /// the model in a permanent instant-429 state.
+    #[tokio::test]
+    async fn cancelled_queued_waiter_rolls_back_accounting() {
+        let cfg = AdmissionConfig {
+            max_in_flight: 1,
+            max_queue_depth: 2,
+            max_wait_secs: 30,
+            // Cap 3 lets the runner + both waiters coexist; if the two
+            // cancelled waiters leaked their counts, the principal would
+            // sit at 3 == cap and the post-cancel enter below would hit
+            // PrincipalCap instead of queueing.
+            max_per_principal: 3,
+        };
+        let ctrl = Arc::new(AdmissionController::new(&cfg));
+        let running = ctrl.enter(Some("acct/key")).await.expect("admit running");
+
+        // Two waiters from the same principal park in the queue…
+        let mut waiters = Vec::new();
+        for _ in 0..2 {
+            let c = Arc::clone(&ctrl);
+            waiters.push(tokio::spawn(async move {
+                c.enter(Some("acct/key")).await.map(drop)
+            }));
+        }
+        tokio::time::sleep(Duration::from_millis(50)).await;
+        assert_eq!(ctrl.queue_depth(), 2);
+
+        // …and both clients vanish (abort = the dropped request future).
+        for w in &waiters {
+            w.abort();
+        }
+        for w in waiters {
+            let _ = w.await;
+        }
+        tokio::time::sleep(Duration::from_millis(50)).await;
+
+        assert_eq!(
+            ctrl.queue_depth(),
+            0,
+            "cancelled waiters must not leak queue slots"
+        );
+
+        // The principal's fair-share count must also be clean: with the
+        // runner still holding 1 of its cap of 3, a new request from the
+        // same principal queues instead of hitting PrincipalCap (which a
+        // leak of the two cancelled counts would trigger).
+        let c = Arc::clone(&ctrl);
+        let retry = tokio::spawn(async move { c.enter(Some("acct/key")).await.map(drop) });
+        tokio::time::sleep(Duration::from_millis(50)).await;
+        assert_eq!(ctrl.queue_depth(), 1, "post-cancel request queues normally");
+        drop(running);
+        retry
+            .await
+            .unwrap()
+            .expect("post-cancel request is served — no leaked principal count");
+    }
+}

crates/neuron/src/harness/arch/qwen3_5/decoder.rs

@@ -22,6 +22,7 @@ use super::TextConfig;
 use super::full_attn::Qwen3_5Attention;
 use super::linear_attn::GatedDeltaNet;
 use super::mlp::Qwen3_5MLP;
+use super::moe::Qwen3_5MoeBlock;
 use super::rmsnorm::Qwen3_5RmsNorm;
 use super::rope::RotaryEmbedding;
 use super::snapshot::LayerKvSnapshot;
@@ -35,10 +36,27 @@ enum AttentionKind {
     Linear(GatedDeltaNet),
 }
 
+/// The FFN slot: dense SwiGLU (Qwen3.6) or the high-sparsity MoE block
+/// (qwen3_next 80B-A3B family, #92), selected per layer by
+/// [`TextConfig::layer_uses_moe`].
+enum MlpKind {
+    Dense(Qwen3_5MLP),
+    Moe(Qwen3_5MoeBlock),
+}
+
+impl Module for MlpKind {
+    fn forward(&self, x: &Tensor) -> candle_core::Result<Tensor> {
+        match self {
+            MlpKind::Dense(mlp) => mlp.forward(x),
+            MlpKind::Moe(moe) => moe.forward(x),
+        }
+    }
+}
+
 pub struct Qwen3_5DecoderLayer {
     input_layernorm: Qwen3_5RmsNorm,
     post_attention_layernorm: Qwen3_5RmsNorm,
-    mlp: Qwen3_5MLP,
+    mlp: MlpKind,
     attention: AttentionKind,
 }
 
@@ -73,7 +91,11 @@ impl Qwen3_5DecoderLayer {
             ),
         };
 
-        let mlp = Qwen3_5MLP::load(cfg, &vb.pp("mlp"))?;
+        let mlp = if cfg.layer_uses_moe(layer_idx) {
+            MlpKind::Moe(Qwen3_5MoeBlock::load(cfg, &vb.pp("mlp"))?)
+        } else {
+            MlpKind::Dense(Qwen3_5MLP::load(cfg, &vb.pp("mlp"))?)
+        };
         let input_layernorm =
             Qwen3_5RmsNorm::load(&vb.pp("input_layernorm"), cfg.hidden_size, cfg.rms_norm_eps)?;
         let post_attention_layernorm = Qwen3_5RmsNorm::load(