fix(reasoning): return the model's work to the client on every surface

Three of five response paths generated reasoning, metered it in `usage`,
and then discarded it. A client that never receives reasoning cannot
replay it, so on the next turn the model faced its own prior turn as a
bare assertion with the thinking that produced it deleted — and paid
for the deletion.

  /v1/chat/completions  streaming      already returned it
  /v1/chat/completions  non-streaming  DISCARDED -> now returned
  /v1/responses         streaming      already returned it
  /v1/responses         non-streaming  DISCARDED -> now returned
  /v1/messages          (Anthropic)    DISCARDED -> now returned

**chat, non-streaming.** `collect_chat_completion` had
`InferenceEvent::ReasoningDelta(_) => {}` — inherited from the path it
replaced, with a doc comment rationalising it. Now accumulated onto
`message.reasoning_content`, the field the streaming path already emits
as `choice.delta.reasoning_content`. Still kept out of `content`: it is
not the assistant's reply.

**responses, non-streaming.** `build_response` emitted a Message item
and nothing else, while the streaming projector beside it built a full
reasoning-item lifecycle. Observed live: a response reporting
`output_tokens_details.reasoning_tokens: 4000` with an empty `output` —
4,000 tokens generated, billed, invisible. Now emits a `reasoning` item
ahead of the message, same shape and ordering as streaming, via a new
`ResponsesOutputItem::Reasoning`.

**Anthropic.** `anthropic_to_openai` deliberately accepts
`thinking.budget_tokens` — the only quantitative reasoning control any
of our wire formats offers (#223) — and `openai_to_anthropic` returned
a bare text block. Taking a parameter and silently discarding what it
buys is the defect. Now emits a `thinking` block ahead of the text.

The Anthropic block carries no `signature`. Anthropic signs theirs and a
client round-tripping one may expect the field; we cannot mint a valid
signature and inventing one would be worse than omitting it. If a real
client rejects an unsigned block the fix is a flag, not a return to
deleting the model's work — recorded in the code so the next person
does not have to rediscover the trade.

Six tests, one pair per surface: reasoning survives, and absent
reasoning leaves the previous shape untouched (absent, not empty). The
lack of exactly these tests is why three paths drifted apart while a
fourth was correct.

Also carried: two byte-exactness regression tests written while chasing
a suspected content-corruption bug — a file read back through a tool
result, and one through the whole Responses->chat->prompt chain, both
asserting `<`, `&`, backticks, `${}`, quotes and blank lines survive
verbatim. The corruption turned out to be elsewhere; the tests are worth
keeping since that path had none.

.gitea/workflows/build-prerelease.yml

@@ -262,7 +262,7 @@ jobs:
       # fast instead of triggering a slower uncached rebuild; only a real
       # sccache fault drops the cache. See script/ci-cargo-escalate.sh.
       - name: Clippy (sccache escalation)
-        run: script/ci-cargo-escalate.sh cargo clippy --workspace -- -D warnings
+        run: script/ci-cargo-escalate.sh cargo clippy --workspace --all-targets -- -D warnings
 
   test:
     name: Test

.gitea/workflows/ci.yml

@@ -101,14 +101,34 @@ jobs:
           echo "changed files:"
           printf '%s\n' "${files}" | sed 's/^/  /'
 
-          # Inert for Rust: the SPA, and documentation. Markdown is
-          # matched only at the repo root and under doc/ — a .md inside
-          # crates/ is left to fall through, since `include_str!` could
-          # make one load-bearing without this file knowing.
+          # Inert for Rust: the SPA, documentation, and fleet config.
+          # Markdown is matched only at the repo root and under doc/ — a
+          # .md inside crates/ is left to fall through, since
+          # `include_str!` could make one load-bearing without this file
+          # knowing.
+          #
+          # Fleet config (models.toml, asset/neuron/<host>.toml,
+          # asset/helexa-bench/<host>.toml) is read by a *running*
+          # daemon, never by a build script or an `include_str!`, so it
+          # cannot change a Rust result. Deploying one is a config rsync
+          # plus a service restart — neuron re-reads it and reloads the
+          # model — so spending ~25 minutes on fmt, clippy, test and a
+          # CUDA type-check to prove a one-line TOML edit compiles is
+          # pure latency (#291).
+          #
+          # The `config` job below stays ungated and still validates
+          # these files, so narrowing here removes the meritless work
+          # without removing the check that matters. build-prerelease
+          # already reaches the same conclusion on main: none of its
+          # per-package path regexes match these files, so a config-only
+          # push builds no binaries.
           rust_relevant="$(printf '%s\n' "${files}" \
             | grep -v '^helexa\.ai/' \
             | grep -v '^doc/' \
             | grep -vE '^[^/]+\.md$' \
+            | grep -vE '^asset/neuron/[^/]+\.toml$' \
+            | grep -vE '^asset/helexa-bench/[^/]+\.toml$' \
+            | grep -vE '^models\.toml$' \
             || true)"
           if [ -n "${rust_relevant}" ]; then
             echo "rust=true" >> "$GITHUB_OUTPUT"
@@ -137,6 +157,27 @@ jobs:
       - uses: actions/checkout@v4
       - run: cargo fmt --check --all
 
+  # models.toml and asset/neuron/<host>.toml both describe how a model is
+  # served, and until #283 nothing forced them to agree. #252 is the bill
+  # for that: models.toml omitted `quant`, so a cortex cold-load served
+  # bf16 and could not prefill, while the same model was fine when the
+  # host loaded it from its own config. The sampling override (#283) is a
+  # second field with the same hazard, so the gate lands with it.
+  #
+  # Deliberately ungated by the `changes` classifier: it is a
+  # sub-second Python run with no build, and the whole point is that it
+  # cannot be skipped on the push that introduces the drift. A config
+  # change is neither `rust` nor `web`, so gating it on either would let
+  # a pure-TOML push through unchecked — precisely the push that breaks
+  # this.
+  config:
+    name: Config consistency
+    timeout-minutes: 5
+    runs-on: fedora-43
+    steps:
+      - uses: actions/checkout@v4
+      - run: python3 script/check-config-consistency.py
+
   web:
     name: Web (lint + typecheck + i18n + build)
     timeout-minutes: 20
@@ -179,8 +220,13 @@ jobs:
       # in sync with build-prerelease.yml): a signal death (rustc SIGSEGV
       # / OOM-kill) keeps the cache and fails fast instead of an uncached
       # rebuild; only a real sccache fault drops the cache.
+      # --all-targets so `#[cfg(test)]` modules, tests/, benches/ and
+      # examples/ are linted too (#286). Without it clippy sees only lib
+      # and bin targets, so `-D warnings` never gated a single line of
+      # test code — which is how two tests lost their `#[test]` attribute
+      # and stopped running unnoticed.
       - name: Clippy (sccache escalation)
-        run: script/ci-cargo-escalate.sh cargo clippy --workspace -- -D warnings
+        run: script/ci-cargo-escalate.sh cargo clippy --workspace --all-targets -- -D warnings
 
   test:
     name: Test

.gitea/workflows/deploy.yml

@@ -37,7 +37,31 @@ env:
     ${{ secrets.RSYNC_SSH_KEY }}
 
 jobs:
+  # Gate the whole rollout on the two config files agreeing about every
+  # model they both describe. models.toml (cortex, cold loads) and
+  # asset/neuron/<host>.toml (resident models) each say how a model is
+  # served — quant, and since #283 the sampling override. When they
+  # disagree the model is served one way cold and another way resident,
+  # which is #252.
+  #
+  # Deliberately a *gate*, not a step inside deploy-cortex: an
+  # inconsistent pair must stop the neuron rollout too, and stopping it
+  # before anything restarts is much cheaper than after (a beast restart
+  # costs ~5-6 min of 27B cold-load).
+  config-check:
+    name: Config consistency
+    runs-on: fedora-43
+    if: >-
+      ${{
+        github.event_name == 'workflow_dispatch'
+        || github.event.workflow_run.conclusion == 'success'
+      }}
+    steps:
+      - uses: actions/checkout@v4
+      - run: python3 script/check-config-consistency.py
+
   deploy-cortex:
+    needs: [config-check]
     runs-on: fedora-43
     # Two trigger paths: manual dispatch always runs; workflow_run
     # only runs if the upstream `build-prerelease` actually succeeded.
@@ -47,6 +71,7 @@ jobs:
         || github.event.workflow_run.conclusion == 'success'
       }}
     steps:
+      - uses: actions/checkout@v4
       - name: SSH init
         run: |
           mkdir -p ~/.ssh
@@ -65,6 +90,33 @@ jobs:
           echo "inv=$(ssh gitea_ci@hanzalova.internal \
             'systemctl show -P InvocationID cortex.service' || true)" >> "$GITHUB_OUTPUT"
 
+      # The catalogue is tracked (#283) precisely so it can be rolled out
+      # here rather than hand-copied. Sync it *before* the package step,
+      # so the restart that step performs is the one that picks it up —
+      # and report whether it changed, because a catalogue edit with no
+      # new RPM must still restart cortex or the change never takes
+      # effect.
+      #
+      # Compared by digest rather than rsync's own change detection so the
+      # answer is available to the shell that decides on the restart.
+      - name: Sync gateway catalogue (models.toml)
+        id: catalogue
+        run: |
+          local_sum=$(sha256sum models.toml | cut -d' ' -f1)
+          remote_sum=$(ssh gitea_ci@hanzalova.internal \
+            'sha256sum /etc/cortex/models.toml 2>/dev/null | cut -d" " -f1' || true)
+          if [ "${local_sum}" = "${remote_sum}" ]; then
+            echo "changed=0" >> "$GITHUB_OUTPUT"
+            echo "models.toml unchanged (${local_sum})"
+          else
+            rsync --archive --compress \
+              --chown root:root --chmod 0644 \
+              --rsync-path 'sudo rsync' \
+              models.toml gitea_ci@hanzalova.internal:/etc/cortex/models.toml
+            echo "changed=1" >> "$GITHUB_OUTPUT"
+            echo "models.toml updated: ${remote_sum:-<absent>} -> ${local_sum}"
+          fi
+
       # Gating compares `rpm -q` against the packages.json manifest the
       # publish job maintains — NOT unprivileged `dnf check-update`,
       # which proved unreliable as the gitea_ci user (hung on metadata
@@ -75,6 +127,11 @@ jobs:
           ssh gitea_ci@hanzalova.internal 'bash -s' <<'DEPLOY'
           set -eu
           pkg=cortex
+          # Default to 0 if the sync step somehow produced no output:
+          # `set -u` plus `[ "" -eq 0 ]` would abort the deploy on an
+          # arithmetic error rather than on anything meaningful.
+          config_changed="${{ steps.catalogue.outputs.changed }}"
+          config_changed="${config_changed:-0}"
           # `rpm -q` prints "package X is not installed" on STDOUT and exits
           # 1, so `|| echo not-installed` never fires — it captures the
           # message text instead. Test the exit code.
@@ -90,18 +147,31 @@ jobs:
               p = max(cands, key=lambda p: p.get("buildTime", 0))
               print(p["version"] + "-" + p["release"])
           ' "${pkg}" 2>/dev/null || true)
-          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ]; then
-            echo "${pkg}-${installed} already current — leaving service untouched"
+          # Skip only when BOTH the package and the catalogue are
+          # unchanged — a models.toml edit (a new sampling override, a
+          # corrected quant) must restart cortex even with no new RPM,
+          # exactly as the neuron drop-in does below.
+          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ] \
+             && [ "${config_changed}" -eq 0 ]; then
+            echo "${pkg}-${installed} already current; models.toml unchanged — leaving service untouched"
             exit 0
           fi
-          echo "installed=${installed} published=${latest:-unknown} — deploying"
+          pkg_changed=1
+          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ]; then
+            pkg_changed=0
+          fi
+          echo "installed=${installed} published=${latest:-unknown} pkg_changed=${pkg_changed} config_changed=${config_changed} — deploying"
           if systemctl is-active --quiet cortex.service; then
             sudo /usr/bin/systemctl stop cortex.service
           fi
-          if rpm -q "${pkg}" >/dev/null 2>&1; then
-            sudo /usr/bin/dnf upgrade --refresh --allowerasing -y cortex
-          else
-            sudo /usr/bin/dnf install --refresh --allowerasing -y cortex
+          # A catalogue-only change still restarts, but must not spend a
+          # dnf metadata refresh proving the package is already current.
+          if [ "${pkg_changed}" -eq 1 ]; then
+            if rpm -q "${pkg}" >/dev/null 2>&1; then
+              sudo /usr/bin/dnf upgrade --refresh --allowerasing -y cortex
+            else
+              sudo /usr/bin/dnf install --refresh --allowerasing -y cortex
+            fi
           fi
           sudo /usr/bin/systemctl daemon-reload
           # enable --now: start the service AND enable it for boot so the
@@ -142,7 +212,7 @@ jobs:
           CAPTURE
 
   deploy-neurons:
-    needs: [deploy-cortex]
+    needs: [config-check, deploy-cortex]
     runs-on: fedora-43
     strategy:
       # One neuron failing must not cancel the others. Cortex is up
@@ -177,6 +247,7 @@ jobs:
             load_timeout: 300
             max_prompt_tokens: 16384
     steps:
+      - uses: actions/checkout@v4
       - name: SSH init
         run: |
           mkdir -p ~/.ssh
@@ -192,6 +263,42 @@ jobs:
           echo "inv=$(ssh gitea_ci@${{ matrix.host }} \
             'systemctl show -P InvocationID neuron.service' || true)" >> "$GITHUB_OUTPUT"
 
+      # asset/neuron/<short>.toml is this host's neuron.toml. It carries
+      # the resident-model specs — including the #283 sampling override,
+      # which is inert until the daemon reloads the model, so a change
+      # here has to restart the neuron the same way the context-cap
+      # drop-in does.
+      #
+      # Synced before the package step so the restart that step performs
+      # is the one that picks it up. Digest-compared rather than trusting
+      # rsync's own change detection, because the restart decision needs
+      # the answer and a needless neuron restart costs ~5-6 min of 27B
+      # cold-load (#1).
+      - name: Sync neuron config (${{ matrix.host }})
+        id: neuronconf
+        run: |
+          short="${{ matrix.host }}"; short="${short%%.*}"
+          src="asset/neuron/${short}.toml"
+          if [ ! -f "${src}" ]; then
+            echo "changed=0" >> "$GITHUB_OUTPUT"
+            echo "no ${src} in the repo — leaving this host's config alone"
+            exit 0
+          fi
+          local_sum=$(sha256sum "${src}" | cut -d' ' -f1)
+          remote_sum=$(ssh gitea_ci@${{ matrix.host }} \
+            'sha256sum /etc/neuron/neuron.toml 2>/dev/null | cut -d" " -f1' || true)
+          if [ "${local_sum}" = "${remote_sum}" ]; then
+            echo "changed=0" >> "$GITHUB_OUTPUT"
+            echo "${src} unchanged (${local_sum})"
+          else
+            rsync --archive --compress \
+              --chown root:root --chmod 0644 \
+              --rsync-path 'sudo rsync' \
+              "${src}" gitea_ci@${{ matrix.host }}:/etc/neuron/neuron.toml
+            echo "changed=1" >> "$GITHUB_OUTPUT"
+            echo "${src} updated: ${remote_sum:-<absent>} -> ${local_sum}"
+          fi
+
       # See deploy-cortex for why gating uses the publish manifest and
       # not unprivileged `dnf check-update`.
       - name: Deploy helexa-neuron-${{ matrix.flavour }} (skips when already current)
@@ -200,6 +307,9 @@ jobs:
           set -eu
           pkg=helexa-neuron-${{ matrix.flavour }}
           max_prompt_tokens="${{ matrix.max_prompt_tokens }}"
+          # A neuron.toml change counts as a config change for the
+          # restart decision below, exactly like the drop-in.
+          neuron_toml_changed="${{ steps.neuronconf.outputs.changed }}"
 
           # ── Desired per-model systemd drop-in ─────────────────────────
           # model.conf carries NEURON_MAX_PROMPT_TOKENS so the context cap
@@ -208,7 +318,7 @@ jobs:
           # deploy-managed value wins over any manual local override of the
           # same variable. See doc/context-limits.md.
           conf=/etc/systemd/system/neuron.service.d/model.conf
-          config_changed=0
+          config_changed=${neuron_toml_changed:-0}
           if [ -n "${max_prompt_tokens}" ]; then
             desired=$(printf '%s\n%s\n%s\n%s' \
               "# Managed by .gitea/workflows/deploy.yml - do not edit by hand." \
@@ -242,13 +352,17 @@ jobs:
           # Skip only when BOTH the package and the drop-in are unchanged —
           # a context-cap change must restart the neuron even with no new RPM.
           if [ "${pkg_changed}" -eq 0 ] && [ "${config_changed}" -eq 0 ]; then
-            echo "${pkg}-${installed} current; NEURON_MAX_PROMPT_TOKENS=${max_prompt_tokens:-<unset>} unchanged — leaving service untouched"
+            echo "${pkg}-${installed} current; NEURON_MAX_PROMPT_TOKENS=${max_prompt_tokens:-<unset>} and neuron.toml unchanged — leaving service untouched"
             exit 0
           fi
           echo "installed=${installed} published=${latest:-unknown} pkg_changed=${pkg_changed} config_changed=${config_changed} — deploying"
 
           # Write the drop-in (staged in gitea_ci's dir, installed root-owned).
-          if [ "${config_changed}" -eq 1 ]; then
+          # Guarded on max_prompt_tokens as well as config_changed: since
+          # a neuron.toml change also sets config_changed, an entry with
+          # no max_prompt_tokens would otherwise reach this with `desired`
+          # unset.
+          if [ "${config_changed}" -eq 1 ] && [ -n "${max_prompt_tokens}" ]; then
             printf '%s\n' "${desired}" > /var/lib/gitea_ci/model.conf
             sudo /usr/bin/install -o root -g root -m 0644 -D /var/lib/gitea_ci/model.conf "${conf}"
             rm -f /var/lib/gitea_ci/model.conf
@@ -534,6 +648,7 @@ jobs:
           CAPTURE
 
   deploy-bench:
+    needs: [config-check]
     runs-on: fedora-43
     if: >-
       ${{
@@ -541,6 +656,7 @@ jobs:
         || github.event.workflow_run.conclusion == 'success'
       }}
     steps:
+      - uses: actions/checkout@v4
       - name: SSH init
         run: |
           mkdir -p ~/.ssh
@@ -556,6 +672,28 @@ jobs:
           echo "inv=$(ssh gitea_ci@bob.hanzalova.internal \
             'systemctl show -P InvocationID helexa-bench.service' || true)" >> "$GITHUB_OUTPUT"
 
+      # Same contract as the gateway and neuron configs: the repo is the
+      # only writer, so a config edit must reach the host and restart the
+      # service even when no new RPM landed.
+      - name: Sync bench config
+        id: benchconf
+        run: |
+          src="asset/helexa-bench/bob.toml"
+          local_sum=$(sha256sum "${src}" | cut -d' ' -f1)
+          remote_sum=$(ssh gitea_ci@bob.hanzalova.internal \
+            'sha256sum /etc/helexa-bench/helexa-bench.toml 2>/dev/null | cut -d" " -f1' || true)
+          if [ "${local_sum}" = "${remote_sum}" ]; then
+            echo "changed=0" >> "$GITHUB_OUTPUT"
+            echo "${src} unchanged (${local_sum})"
+          else
+            rsync --archive --compress \
+              --chown root:root --chmod 0644 \
+              --rsync-path 'sudo rsync' \
+              "${src}" gitea_ci@bob.hanzalova.internal:/etc/helexa-bench/helexa-bench.toml
+            echo "changed=1" >> "$GITHUB_OUTPUT"
+            echo "${src} updated: ${remote_sum:-<absent>} -> ${local_sum}"
+          fi
+
       # See deploy-cortex for why gating uses the publish manifest and
       # not unprivileged `dnf check-update`.
       - name: Deploy helexa-bench (skips when already current)
@@ -563,6 +701,8 @@ jobs:
           ssh gitea_ci@bob.hanzalova.internal 'bash -s' <<'DEPLOY'
           set -eu
           pkg=helexa-bench
+          config_changed="${{ steps.benchconf.outputs.changed }}"
+          config_changed="${config_changed:-0}"
           # `rpm -q` prints "package X is not installed" on STDOUT and exits
           # 1, so `|| echo not-installed` never fires — it captures the
           # message text instead. Test the exit code.
@@ -578,18 +718,25 @@ jobs:
               p = max(cands, key=lambda p: p.get("buildTime", 0))
               print(p["version"] + "-" + p["release"])
           ' "${pkg}" 2>/dev/null || true)
-          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ]; then
-            echo "${pkg}-${installed} already current — leaving service untouched"
+          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ] \
+             && [ "${config_changed}" -eq 0 ]; then
+            echo "${pkg}-${installed} already current; config unchanged — leaving service untouched"
             exit 0
           fi
-          echo "installed=${installed} published=${latest:-unknown} — deploying"
+          pkg_changed=1
+          if [ -n "${latest}" ] && [ "${latest}" = "${installed}" ]; then
+            pkg_changed=0
+          fi
+          echo "installed=${installed} published=${latest:-unknown} pkg_changed=${pkg_changed} config_changed=${config_changed} — deploying"
           if systemctl is-active --quiet helexa-bench.service; then
             sudo /usr/bin/systemctl stop helexa-bench.service
           fi
-          if rpm -q "${pkg}" >/dev/null 2>&1; then
-            sudo /usr/bin/dnf upgrade --refresh --allowerasing -y helexa-bench
-          else
-            sudo /usr/bin/dnf install --refresh --allowerasing -y helexa-bench
+          if [ "${pkg_changed}" -eq 1 ]; then
+            if rpm -q "${pkg}" >/dev/null 2>&1; then
+              sudo /usr/bin/dnf upgrade --refresh --allowerasing -y helexa-bench
+            else
+              sudo /usr/bin/dnf install --refresh --allowerasing -y helexa-bench
+            fi
           fi
           sudo /usr/bin/systemctl daemon-reload
           # enable --now: start the service AND enable it for boot so the

.gitignore

@@ -9,7 +9,6 @@ helexa.ai/.env.local
 .idea/
 .vscode/
 cortex.toml
-models.toml
 doc/plan/*
 /target-cuda/
 .claude/

CLAUDE.md

@@ -167,7 +167,7 @@ doc-comment; touch points (the `Job` enum, the dispatch handlers, the
 cargo build --release           # build all crates
 cargo run -p cortex-cli -- serve    # run the gateway
 cargo test                      # run all tests
-cargo clippy --workspace        # lint
+cargo clippy --workspace --all-targets   # lint (test code included)
 ```
 
 ## CI
@@ -177,7 +177,7 @@ pass before merging:
 
 ```sh
 cargo fmt --check --all                    # formatting
-cargo clippy --workspace -- -D warnings   # lint (warnings are errors)
+cargo clippy --workspace --all-targets -- -D warnings   # lint (warnings are errors)
 cargo test --workspace                     # tests
 ```
 
@@ -191,7 +191,8 @@ 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`).
+   `cargo clippy --workspace --all-targets -- -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

asset/helexa-bench/bob.toml

@@ -1,3 +1,10 @@
+# ─────────────────────────────────────────────────────────────────────
+# MANAGED BY CI — .gitea/workflows/deploy.yml is the only writer.
+#
+# Edits made directly to the deployed copy on the host are reverted by
+# the next deploy, and until then the fleet runs something no commit
+# describes. Change it here, commit, let the pipeline ship it.
+# ─────────────────────────────────────────────────────────────────────
 # helexa-bench config for bob.hanzalova.internal.
 #
 # Synced to /etc/helexa-bench/helexa-bench.toml by script/infra-setup.sh
@@ -15,6 +22,20 @@ iteration_pause_secs = 2
 request_timeout_secs = 600
 db_path = "/var/lib/helexa-bench/bench.sqlite"
 
+# Identity bench presents to neurons (#288). Without it bench is
+# anonymous, and since #262 anonymous callers are served only from
+# leftover capacity — capped below max_in_flight and parked at the class
+# gate. That moved beast's concurrency:8 from 168.9 tok/s / 0.53s
+# ttft_p95 to 104.6 / 11.77s the day #262 landed and held there, so every
+# number since measured the yield policy, not serving capacity.
+#
+# A dedicated pair, not a borrowed one: bench then holds its own
+# fair-share allocation (#54) and cannot starve interactive traffic —
+# the failure #262 existed to fix.
+[bench.principal]
+account_id = "helexa-bench"
+key_id = "fleet-benchmark"
+
 [scenarios]
 prompt_sizes = [128, 4096]
 max_tokens = 256
@@ -22,6 +43,10 @@ max_tokens = 256
 # 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]
+# Non-streaming bursts (#285/#288). Kept to the top level: it is where
+# serialization shows most starkly and each level costs a full burst per
+# sample per build.
+concurrency_nonstreaming_levels = [8]
 concurrency_prompt_tokens = 512
 
 # Capability probes (#91) — the reasoning/planning axis the speed

asset/neuron/beast.toml

@@ -1,3 +1,10 @@
+# ─────────────────────────────────────────────────────────────────────
+# MANAGED BY CI — .gitea/workflows/deploy.yml is the only writer.
+#
+# Edits made directly to the deployed copy on the host are reverted by
+# the next deploy, and until then the fleet runs something no commit
+# describes. Change it here, commit, let the pipeline ship it.
+# ─────────────────────────────────────────────────────────────────────
 # neuron.toml for beast.hanzalova.internal
 #
 # 2x RTX 5090 (32 GB each) — TP-2 capable. Pre-warms Qwen3.6-27B with
@@ -22,7 +29,58 @@ name = "candle"
 # 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
+# 2, not 8 (#291). max_in_flight is both the admission bound and the
+# batch engine's slot count, and the KV budget is divided by it at load.
+# At 8 that gave each slot 18,852 tokens out of 150,816 — against a
+# 131,072-token advertised context, which needs 4 GB of KV per card on
+# its own and 32 GB across eight slots, on a 31 GB card already holding
+# ~21 GB of weights. The advertised context and the slot count were
+# incompatible by roughly 7x, and an agentic session ran in a ~19k
+# window while seven idle slots held reservations.
+#
+# Seven days of paired request spans on this host: 89.6% of busy time
+# was ONE request in flight, 93.6% was two or fewer, and utilisation was
+# 4.9%. Most of what little concurrency exists is helexa-bench's own
+# synthetic bursts. Two slots cover 93.6% of observed load and give each
+# ~75k tokens; the batch engine is still worth 1.79x there (5.24x at
+# eight is throughput nobody is asking this host for). Anything above
+# two queues on admission (#53) with Retry-After rather than being
+# refused.
+max_in_flight = 2
+
+# Snapshot budget for the cross-request prefix cache (#292).
+#
+# 3 GiB, not the 1 GiB default. At the measured 32 KiB/token/card, a
+# 1 GiB budget cannot hold a snapshot beyond ~32k tokens of K/V alone —
+# less once the GatedDeltaNet conv/recurrent state is deep-copied on top,
+# which on this arch is 48 of 64 layers in f32. Agentic sessions here run
+# well past that: `cached_tokens` was observed pinned at 2,016 for ten
+# consecutive turns while the prompt grew to 52,869, so every turn
+# re-prefilled ~50k tokens — 37 s each, against ~20 s of actual decode.
+# The cache was not missing; the snapshot would not fit.
+#
+# The snapshot budget is reserved BEFORE the KV budget, not from what is
+# left over: measured on this host, kv_budget_mb = 5737 - budget_mb,
+# exactly 1:1. So this number is a direct trade against how long a
+# single session can get, and both sides have a floor:
+#
+#   budget_mb   kv_budget   tokens/slot (2 slots)   holds a 52.9k snapshot
+#   1024        4713        75,400                  no  ← the #292 symptom
+#   3072        2665        42,600                  yes, but sessions
+#                                                   cannot reach 52.9k
+#   2048        3689        59,000                  yes
+#
+# 2 GiB is the value that satisfies both: a 52,869-token snapshot is
+# ~1,652 MB of K/V plus the GatedDeltaNet state (~1.73 GB total), and
+# per-slot KV still covers sessions comfortably past the deepest
+# observed. 3 GiB was tried first and starved KV — a session could no
+# longer reach the size whose snapshot the cache was being enlarged to
+# hold, which is self-defeating.
+#
+# max_entries stays at the default 8: the byte budget binds first here,
+# so the count cap is not what is forcing snapshots to be small.
+[harness.candle.prefix_cache]
+budget_mb = 2048
 
 [[default_models]]
 # Qwen3.8-27B is the resident flagship (#251). Architecturally identical
@@ -35,3 +93,16 @@ harness = "candle"
 quant = "q6k"
 tensor_parallel = 2
 devices = [0, 1]
+
+# Sampling override (#283) — MUST match the `[models.sampling]` block on
+# this model in models.toml. Two files describe one model; only
+# script/check-config-consistency.py forces them to agree, and #252 is
+# what happens when nothing does.
+#
+# The model's own generation_config.json says temperature = 1.0, which
+# measured a 20% structural-defect rate on ~2k-token code generation
+# against 0/60 at <= 0.6 (Fisher exact one-sided p = 0.0031). Precedence
+# is request > operator > model, so a caller naming a temperature still
+# gets what it asked for.
+[default_models.sampling]
+temperature = 0.6

asset/neuron/benjy.toml

@@ -1,3 +1,10 @@
+# ─────────────────────────────────────────────────────────────────────
+# MANAGED BY CI — .gitea/workflows/deploy.yml is the only writer.
+#
+# Edits made directly to the deployed copy on the host are reverted by
+# the next deploy, and until then the fleet runs something no commit
+# describes. Change it here, commit, let the pipeline ship it.
+# ─────────────────────────────────────────────────────────────────────
 # neuron.toml for benjy.hanzalova.internal
 #
 # 1x RTX 4090 (24 GB) — largest single-GPU host on the fleet. Pre-warms

asset/neuron/quadbrat.toml

@@ -1,3 +1,10 @@
+# ─────────────────────────────────────────────────────────────────────
+# MANAGED BY CI — .gitea/workflows/deploy.yml is the only writer.
+#
+# Edits made directly to the deployed copy on the host are reverted by
+# the next deploy, and until then the fleet runs something no commit
+# describes. Change it here, commit, let the pipeline ship it.
+# ─────────────────────────────────────────────────────────────────────
 # neuron.toml for quadbrat.hanzalova.internal
 #
 # 1x RTX 3060 (12 GB) — small / quantised tier. Pre-warms Qwen3-1.7B

bench/src/api.ts

@@ -1,4 +1,10 @@
-import type { Dimensions, ReportRow, RunRow, SeriesPoint } from "./types";
+import type {
+  Dimensions,
+  MeasurementRegime,
+  ReportRow,
+  RunRow,
+  SeriesPoint,
+} from "./types";
 
 // Empty default → `fetch('/api/...')` hits the dev proxy (vite.config.ts)
 // or the same origin. For a separately-hosted build, set VITE_API_BASE to
@@ -15,6 +21,7 @@ async function getJson<T>(path: string): Promise<T> {
 
 export const getDimensions = () => getJson<Dimensions>("/api/dimensions");
 export const getSummary = () => getJson<ReportRow[]>("/api/summary");
+export const getRegimes = () => getJson<MeasurementRegime[]>("/api/regimes");
 
 // host is resolved server-side (each model maps to one host today), so the
 // public UI selects by model + scenario alone.

bench/src/pages/Trends.tsx

@@ -11,8 +11,8 @@ import {
   XAxis,
   YAxis,
 } from "recharts";
-import { getDimensions, getSeries } from "../api";
-import type { Dimensions, SeriesPoint } from "../types";
+import { getDimensions, getRegimes, getSeries } from "../api";
+import type { Dimensions, MeasurementRegime, SeriesPoint } from "../types";
 import { BASELINE_SOURCE, baselineFor } from "../baseline";
 
 function Picker({
@@ -40,11 +40,95 @@ function Picker({
   );
 }
 
+type SeriesDef = {
+  key: string;
+  name: string;
+  stroke: string;
+  dashed?: boolean;
+};
+
+/** One titled chart over the shared build timeline.
+ *
+ * Every panel draws the same x-axis and the same regime divider, so a
+ * reader can line a change up across metrics — which is the whole point
+ * of having more than two of them. */
+/** A vertical rule at a build where the metric changed meaning. */
+type Rule = { at: string; label: string; detail?: string };
+
+function MetricChart({
+  title,
+  hint,
+  data,
+  lines,
+  rules,
+  unit,
+}: {
+  title: string;
+  hint?: string;
+  data: Record<string, unknown>[];
+  lines: SeriesDef[];
+  rules?: Rule[];
+  unit?: string;
+}) {
+  const hasAny = data.some((d) => lines.some((l) => d[l.key] != null));
+  if (!hasAny) return null;
+  return (
+    <>
+      <h5 className="mt-4">
+        {title}
+        {unit ? <span className="text-muted fw-normal"> ({unit})</span> : null}
+      </h5>
+      {hint && <p className="text-muted small mb-2">{hint}</p>}
+      <ResponsiveContainer width="100%" height={260}>
+        <LineChart data={data} margin={{ top: 8, right: 24, bottom: 8, left: 0 }}>
+          <CartesianGrid strokeDasharray="3 3" />
+          <XAxis dataKey="label" />
+          <YAxis />
+          <Tooltip />
+          <Legend />
+          {(rules ?? [])
+            // Only draw a rule whose build is actually on this x-axis;
+            // recharts would otherwise place it at the origin and imply
+            // the boundary sits before all the data.
+            .filter((r) => data.some((d) => d.label === r.at))
+            .map((r) => (
+              <ReferenceLine
+                key={`${r.at}-${r.label}`}
+                x={r.at}
+                stroke="#bbb"
+                strokeDasharray="3 3"
+                label={{
+                  value: r.label,
+                  position: "top",
+                  fill: "#999",
+                  fontSize: 11,
+                }}
+              />
+            ))}
+          {lines.map((l) => (
+            <Line
+              key={l.key}
+              type="monotone"
+              dataKey={l.key}
+              name={l.name}
+              stroke={l.stroke}
+              strokeDasharray={l.dashed ? "5 5" : undefined}
+              connectNulls
+              dot={false}
+            />
+          ))}
+        </LineChart>
+      </ResponsiveContainer>
+    </>
+  );
+}
+
 export default function Trends() {
   const [dims, setDims] = useState<Dimensions | null>(null);
   const [model, setModel] = useState("");
   const [scenario, setScenario] = useState("");
   const [series, setSeries] = useState<SeriesPoint[]>([]);
+  const [regimes, setRegimes] = useState<MeasurementRegime[]>([]);
   const [err, setErr] = useState<string | null>(null);
 
   useEffect(() => {
@@ -55,6 +139,10 @@ export default function Trends() {
         if (d.scenarios[0]) setScenario(d.scenarios[0]);
       })
       .catch((e) => setErr(String(e)));
+    // A missing/older API just means no rules — not a page failure.
+    getRegimes()
+      .then(setRegimes)
+      .catch(() => setRegimes([]));
   }, []);
 
   useEffect(() => {
@@ -68,10 +156,7 @@ export default function Trends() {
   // Prepend the pre-helexa-bench baseline (dashed, separate keys) so it
   // anchors the timeline without being merged into the live line. Different
   // measurement regime — see baseline.ts / doc/benchmarks.md.
-  const base = useMemo(
-    () => baselineFor(model, scenario),
-    [model, scenario],
-  );
+  const base = useMemo(() => baselineFor(model, scenario), [model, scenario]);
   const data = useMemo(
     () => [
       ...base.map((p) => ({
@@ -85,6 +170,14 @@ export default function Trends() {
         ttft: p.ttft_s_median,
         decode: p.decode_tps_median,
         total: p.total_s_median,
+        ttftP95: p.ttft_p95_s_median,
+        queueWait: p.queue_wait_ms_median,
+        rejected: p.rejected_median,
+        prefillTps: p.prefill_tps_median,
+        reasoning: p.reasoning_tokens_median,
+        cached: p.cached_tokens_median,
+        completion: p.completion_tokens_median,
+        tpot: p.tpot_p95_ms_median,
       })),
     ],
     [series, base],
@@ -95,6 +188,55 @@ export default function Trends() {
   const firstLive = series[0]?.git_sha;
   const showDivider = base.length > 0 && series.length > 0;
 
+  // The build where the measuring identity changed, derived from the
+  // data rather than declared — a constant would go stale the moment
+  // the principal is reconfigured (#288).
+  const identityShift = useMemo(() => {
+    for (let i = 1; i < series.length; i++) {
+      const prev = series[i - 1].principal ?? "anonymous";
+      const cur = series[i].principal ?? "anonymous";
+      if (prev !== cur) {
+        return {
+          at: series[i].git_sha,
+          label: `measured as ${cur === "anonymous" ? "anonymous" : "identified"}`,
+          detail:
+            "The identity bench measured under changed here. Anonymous " +
+            "samples are subject to the #262 yield policy and understate " +
+            "capacity, so points either side are not comparable (#288).",
+        } as Rule;
+      }
+    }
+    return null;
+  }, [series]);
+
+  /** Declared boundaries touching `metric`, plus the regime-independent
+   *  ones that apply to every panel. */
+  const rulesFor = (metric: string): Rule[] => {
+    const out: Rule[] = [];
+    if (showDivider && firstLive) {
+      out.push({ at: firstLive, label: "bench.py → helexa-bench" });
+    }
+    for (const r of regimes) {
+      if (r.affects.includes(metric)) {
+        out.push({ at: r.first_sha, label: r.label, detail: r.detail });
+      }
+    }
+    if (identityShift) out.push(identityShift);
+    return out;
+  };
+
+  // Anonymous and identified samples are not comparable once #262 is in
+  // the build: an anonymous caller is capped below max_in_flight and
+  // parked at the class gate, so it characterises the yield policy
+  // rather than serving capacity. Say so rather than letting someone
+  // read a step change as an engine regression (#288).
+  const identities = useMemo(
+    () => new Set(series.map((p) => p.principal ?? "anonymous")),
+    [series],
+  );
+  const mixedIdentity = identities.size > 1;
+  const anyAnonymous = identities.has("anonymous");
+
   if (err) return <Alert variant="danger">{err}</Alert>;
   if (!dims) return <Spinner animation="border" />;
 
@@ -122,6 +264,25 @@ export default function Trends() {
         </p>
       )}
 
+      {mixedIdentity && (
+        <Alert variant="warning" className="py-2">
+          <strong>Mixed measurement identity.</strong> Some builds were
+          sampled anonymously and some under a principal. Since{" "}
+          <code>#262</code> an anonymous caller is capped below{" "}
+          <code>max_in_flight</code> and yields to identified traffic, so
+          those points measure the admission policy rather than serving
+          capacity. A step change across that boundary is the instrument,
+          not the engine — see <code>#288</code>.
+        </Alert>
+      )}
+      {!mixedIdentity && anyAnonymous && (
+        <Alert variant="secondary" className="py-2 small">
+          Sampled anonymously. Since <code>#262</code> anonymous callers are
+          served from leftover capacity, so these numbers understate what an
+          authenticated caller gets (<code>#288</code>).
+        </Alert>
+      )}
+
       {data.length === 0 ? (
         <Alert variant="info">No data for this selection yet.</Alert>
       ) : (
@@ -133,87 +294,128 @@ export default function Trends() {
               see <code>doc/benchmarks.md</code>.
             </p>
           )}
-          <h5 className="mt-3">decode tok/s (higher is better)</h5>
-          <ResponsiveContainer width="100%" height={280}>
-            <LineChart data={data} margin={{ top: 8, right: 24, bottom: 8, left: 0 }}>
-              <CartesianGrid strokeDasharray="3 3" />
-              <XAxis dataKey="label" />
-              <YAxis />
-              <Tooltip />
-              <Legend />
-              {showDivider && firstLive && (
-                <ReferenceLine
-                  x={firstLive}
-                  stroke="#bbb"
-                  strokeDasharray="3 3"
-                  label={{
-                    value: "bench.py → helexa-bench",
-                    position: "top",
-                    fill: "#999",
-                    fontSize: 11,
-                  }}
-                />
-              )}
-              <Line
-                type="monotone"
-                dataKey="decode"
-                name="decode tok/s"
-                stroke="#0d6efd"
-                connectNulls
-              />
-              {base.length > 0 && (
-                <Line
-                  type="monotone"
-                  dataKey="baseDecode"
-                  name="baseline (bench.py · gateway)"
-                  stroke="#888"
-                  strokeDasharray="5 5"
-                  connectNulls
-                />
-              )}
-            </LineChart>
-          </ResponsiveContainer>
 
-          <h5 className="mt-4">TTFT seconds (lower is better)</h5>
-          <ResponsiveContainer width="100%" height={280}>
-            <LineChart data={data} margin={{ top: 8, right: 24, bottom: 8, left: 0 }}>
-              <CartesianGrid strokeDasharray="3 3" />
-              <XAxis dataKey="label" />
-              <YAxis />
-              <Tooltip />
-              <Legend />
-              {showDivider && firstLive && (
-                <ReferenceLine
-                  x={firstLive}
-                  stroke="#bbb"
-                  strokeDasharray="3 3"
-                  label={{
-                    value: "bench.py → helexa-bench",
-                    position: "top",
-                    fill: "#999",
-                    fontSize: 11,
-                  }}
-                />
-              )}
-              <Line
-                type="monotone"
-                dataKey="ttft"
-                name="TTFT (s)"
-                stroke="#dc3545"
-                connectNulls
-              />
-              {base.length > 0 && (
-                <Line
-                  type="monotone"
-                  dataKey="baseTtft"
-                  name="baseline (bench.py · gateway)"
-                  stroke="#888"
-                  strokeDasharray="5 5"
-                  connectNulls
-                />
-              )}
-            </LineChart>
-          </ResponsiveContainer>
+          <MetricChart
+            title="decode tok/s"
+            unit="higher is better"
+            data={data}
+            rules={rulesFor("decode")}
+            lines={[
+              { key: "decode", name: "decode tok/s", stroke: "#0d6efd" },
+              ...(base.length > 0
+                ? [
+                    {
+                      key: "baseDecode",
+                      name: "baseline (bench.py · gateway)",
+                      stroke: "#888",
+                      dashed: true,
+                    },
+                  ]
+                : []),
+            ]}
+          />
+
+          <MetricChart
+            title="prefill tok/s"
+            unit="higher is better"
+            hint="The other half of serving speed, derived from prefill_tokens / prefill_ms. A prefix-cache hit shortens prefill_ms while the token count stays whole, so a high rate here is itself the cache-hit signal."
+            data={data}
+            rules={rulesFor("prefillTps")}
+            lines={[
+              { key: "prefillTps", name: "prefill tok/s", stroke: "#20c997" },
+            ]}
+          />
+
+          <MetricChart
+            title="TTFT"
+            unit="seconds, lower is better"
+            hint="Median and p95 together on purpose. Under concurrency the median is dominated by whichever streams were admitted immediately; the p95 is the one that moves when a caller is made to wait. Charting only the median is how a 0.53 s → 11.77 s tail went unnoticed for a week."
+            data={data}
+            rules={rulesFor("ttft")}
+            lines={[
+              { key: "ttft", name: "TTFT median (s)", stroke: "#dc3545" },
+              { key: "ttftP95", name: "TTFT p95 (s)", stroke: "#fd7e14" },
+              ...(base.length > 0
+                ? [
+                    {
+                      key: "baseTtft",
+                      name: "baseline (bench.py · gateway)",
+                      stroke: "#888",
+                      dashed: true,
+                    },
+                  ]
+                : []),
+            ]}
+          />
+
+          <MetricChart
+            title="inter-token gap p95"
+            unit="ms, lower is better"
+            hint="Stream smoothness. decode tok/s is a mean over the whole window, so a stream that stalls and then catches up is indistinguishable from one that never stalled — this is the number a user feels."
+            data={data}
+            rules={rulesFor("tpot")}
+            lines={[
+              { key: "tpot", name: "inter-token p95 (ms)", stroke: "#6f42c1" },
+            ]}
+          />
+
+          <MetricChart
+            title="admission"
+            unit="queue wait ms · requests shed"
+            hint="Separates “the server is slow” from “you were queued behind someone”. Queue wait is TTFT minus server-measured prefill; rejected counts honest backpressure rather than silent failures."
+            data={data}
+            rules={rulesFor("queueWait")}
+            lines={[
+              { key: "queueWait", name: "queue wait (ms)", stroke: "#d63384" },
+              { key: "rejected", name: "rejected (count)", stroke: "#adb5bd" },
+            ]}
+          />
+
+          <MetricChart
+            title="tokens per sample"
+            unit="counts"
+            hint="Cost, not speed. Reasoning tokens are the dominant driver on a reasoning model and move independently of every rate above — a template or sampling change can double what the model thinks before answering while the speed charts stay flat. Cached tokens are why prefill timing varies between otherwise identical samples."
+            data={data}
+            rules={rulesFor("completion")}
+            lines={[
+              { key: "completion", name: "completion tokens", stroke: "#0dcaf0" },
+              { key: "reasoning", name: "reasoning tokens", stroke: "#ffc107" },
+              { key: "cached", name: "cached prompt tokens", stroke: "#198754" },
+            ]}
+          />
+
+          {(() => {
+            // Explain every rule actually drawn for this selection. The
+            // chart label is a name; without the reason, a reader still
+            // has to go and rediscover what it meant — which is the cost
+            // this whole feature exists to remove.
+            const shown = [
+              ...regimes
+                .filter((r) => series.some((p) => p.git_sha === r.first_sha))
+                .map((r) => ({ at: r.first_sha, label: r.label, detail: r.detail })),
+              ...(identityShift ? [identityShift] : []),
+            ];
+            if (shown.length === 0) return null;
+            return (
+              <div className="mt-4 pt-3 border-top">
+                <p className="text-muted small mb-2">
+                  <strong>Dashed rules mark measurement-regime changes</strong> —
+                  builds where a number changed meaning rather than value. A step
+                  across one is the instrument, not the engine.
+                </p>
+                <dl className="row small text-muted mb-0">
+                  {shown.map((r) => (
+                    <div key={`${r.at}-${r.label}`} className="mb-2">
+                      <dt className="fw-semibold">
+                        <code>{r.at}</code> — {r.label}
+                      </dt>
+                      <dd className="mb-0">{r.detail}</dd>
+                    </div>
+                  ))}
+                </dl>
+              </div>
+            );
+          })()}
         </>
       )}
     </>

bench/src/types.ts

@@ -41,6 +41,17 @@ export interface SeriesPoint {
   ttft_s_median: number | null;
   decode_tps_median: number | null;
   total_s_median: number | null;
+  ttft_p95_s_median: number | null;
+  queue_wait_ms_median: number | null;
+  rejected_median: number | null;
+  prefill_tps_median: number | null;
+  reasoning_tokens_median: number | null;
+  cached_tokens_median: number | null;
+  completion_tokens_median: number | null;
+  tpot_p95_ms_median: number | null;
+  /** Identity the samples were taken under, or null if anonymous.
+   *  Anonymous and identified points are not comparable (#288). */
+  principal: string | null;
   samples: number;
 }
 
@@ -67,3 +78,13 @@ export interface RunRow {
   ok: boolean;
   error: string | null;
 }
+
+/** A build at which a metric changed meaning rather than value (#288).
+ *  Served from the API so the UI and `report` cite one list. */
+export interface MeasurementRegime {
+  first_sha: string;
+  label: string;
+  detail: string;
+  /** Series keys this affects, as Trends.tsx names them. */
+  affects: string[];
+}

crates/cortex-core/src/catalogue.rs

@@ -1,7 +1,7 @@
 //! Model catalogue — profiles describing how to serve each model.
 
 use crate::discovery::DeviceInfo;
-use crate::harness::{ModelCost, ModelLimit};
+use crate::harness::{ModelCost, ModelLimit, SamplingOverride};
 use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 use std::path::Path;
@@ -57,6 +57,23 @@ pub struct ModelProfile {
     /// on this being explicit per model rather than implicit.
     #[serde(default)]
     pub source: Option<String>,
+    /// Operator sampling override forwarded to neuron on cold load
+    /// (#283). The counterpart to the same field in a host's
+    /// `[[default_models]]` — a model that can be both cold-loaded by
+    /// cortex and resident on a host must carry the same value in both,
+    /// which `script/check-config-consistency.py` enforces in CI.
+    ///
+    /// This is the #252 shape of hazard: two config files describing one
+    /// model, with nothing forcing them to agree.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub sampling: Option<SamplingOverride>,
+    /// The model's own `preserve_thinking` template control, forwarded
+    /// on cold load so a model this gateway loads behaves the same as
+    /// one the host was configured to hold resident. Two config files
+    /// describing one model is how #252 happened; this rides along for
+    /// the same reason `sampling` does.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub preserve_thinking: Option<bool>,
 
     // ── Enrichment (issue #62) ────────────────────────────────
     /// Per-model token budget. When present, advertised in `/v1/models`
@@ -240,6 +257,8 @@ mod tests {
             pinned_on: vec![],
             residency_priority: None,
             source: None,
+            sampling: None,
+            preserve_thinking: None,
             limit: None,
             cost: None,
             capabilities: vec![],

crates/cortex-core/src/harness.rs

@@ -26,7 +26,115 @@ pub struct HarnessHealth {
     pub uptime_secs: Option<u64>,
 }
 
+/// Operator-set sampling defaults for one model (#283).
+///
+/// A model's `generation_config.json` is the model authors' statement of
+/// intent, and honouring it (#272) is right by default. It is not always
+/// right for the workload the operator serves: `Qwen/Qwen3.8-27B`
+/// publishes `temperature = 1.0`, which measured a 20% structural-defect
+/// rate on ~2k-token code generation against 0/60 at `<= 0.6`
+/// (#283, p = 0.0031).
+///
+/// This is the operator's explicit, visible correction to that default —
+/// deliberately *not* a heuristic that inspects the model and guesses.
+/// A temperature guesser is a footgun; a number in a config file the
+/// operator wrote is not.
+///
+/// Precedence is **request > operator > model > built-in fallback**: the
+/// override replaces what the model published, but a caller that names a
+/// value still wins, because an explicit API parameter that is silently
+/// ignored is a contract break.
+///
+/// Every field is optional and overlays independently — setting
+/// `temperature` alone leaves the model's `top_p`/`top_k` in force.
+#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
+pub struct SamplingOverride {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub top_p: Option<f64>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub top_k: Option<usize>,
+    /// Multiplier applied to recently-generated tokens before sampling.
+    /// `1.0` disables it; above that, recently-emitted tokens get less
+    /// likely.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub repeat_penalty: Option<f32>,
+    /// How many recently-generated tokens the penalty considers.
+    ///
+    /// The reason this is operator-settable at all: the built-in default
+    /// is 64, inherited from candle's chat example, and a reasoning
+    /// model's think block runs to tens of thousands of tokens. Anything
+    /// restated more than 64 tokens later is invisible to the penalty,
+    /// which is how a 28,672-token think block can contain the same
+    /// derivation five times over.
+    ///
+    /// Raising it is a real trade, not a free win: the penalty is
+    /// token-level, and source code legitimately repeats tokens
+    /// constantly (`const`, `function`, brace runs). A wide window at a
+    /// meaningful penalty distorts exactly the output this fleet is for.
+    /// Sequence-level repetition wants a sequence-level instrument
+    /// (frequency/presence penalties, or a DRY-style sampler); this knob
+    /// exists so the trade can be *measured* per model rather than
+    /// guessed once and frozen into a constant.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub repeat_last_n: Option<usize>,
+    /// Sequence-wide one-time penalty for already-seen tokens.
+    ///
+    /// This, not `repeat_last_n`, is the remedy Qwen prescribe for the
+    /// endless repetition their models fall into during long reasoning
+    /// (0..2, 1.5 when severe). It is scored over everything generated
+    /// so far, so a passage restated 5,000 tokens later is still
+    /// penalised — the case a trailing window cannot see.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub presence_penalty: Option<f32>,
+    /// Sequence-wide penalty proportional to a token's existing count.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub frequency_penalty: Option<f32>,
+    /// Pin the sampler's RNG for every request to this model.
+    ///
+    /// Exists because parameter search needs a control: comparing two
+    /// `presence_penalty` values across unseeded runs cannot separate
+    /// the parameter's effect from sampling variance, and the clients
+    /// that would otherwise pin it cannot — pi's request builder sends
+    /// no `seed` at all, so a caller-side seed is not available on the
+    /// path that matters.
+    ///
+    /// Note what this does and does not give. It removes sampling noise
+    /// as a confounder; it does not make two runs comparable
+    /// token-by-token, because changing a penalty changes the logits
+    /// and the sequences diverge at the first step where a token choice
+    /// flips.
+    ///
+    /// Leaving this set is a product decision, not just a testing one:
+    /// every identical prompt then yields an identical answer for every
+    /// caller. That is defensible — predictable and cacheable — but it
+    /// removes the variation a user may expect from retrying. A request
+    /// that names its own seed still wins.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub seed: Option<u64>,
+}
+
+impl SamplingOverride {
+    /// True when the operator set nothing — an empty `[sampling]` table
+    /// must behave exactly as no table at all.
+    pub fn is_empty(&self) -> bool {
+        self.temperature.is_none()
+            && self.top_p.is_none()
+            && self.top_k.is_none()
+            && self.repeat_penalty.is_none()
+            && self.repeat_last_n.is_none()
+            && self.presence_penalty.is_none()
+            && self.frequency_penalty.is_none()
+            && self.seed.is_none()
+    }
+}
+
 /// Specification for loading a model through a harness.
+///
+/// Doubles as the `[[default_models]]` entry shape in `neuron.toml`, so
+/// a field added here is available to both the load API and the
+/// operator's per-host config without a second type to keep in step.
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct ModelSpec {
     pub model_id: String,
@@ -34,6 +142,36 @@ pub struct ModelSpec {
     pub quant: Option<String>,
     pub tensor_parallel: Option<u32>,
     pub devices: Option<Vec<u32>>,
+    /// Operator sampling override (#283). Travels with the *model*, not
+    /// the request: neuron is the only place sampling is resolved, so
+    /// the override has to reach it whether the load came from cortex's
+    /// catalogue or from this host's own `[[default_models]]`.
+    ///
+    /// `#[serde(default)]` so a cortex or neuron predating this field
+    /// still round-trips the spec.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub sampling: Option<SamplingOverride>,
+    /// Whether prior assistant turns keep their reasoning when the chat
+    /// template re-renders the conversation.
+    ///
+    /// This is the model's own `preserve_thinking` template control, not
+    /// an invention of ours. `Qwen/Qwen3.8-27B` defaults it to *true* —
+    /// the whole transcript's think blocks are replayed — and `false`
+    /// keeps reasoning only for turns after the last user query, i.e.
+    /// the turn in progress.
+    ///
+    /// `None` leaves the kwarg unset so the template's own default
+    /// applies, which is the only safe default: the model's authors
+    /// chose it, and overriding that fleet-wide on a hunch is how you
+    /// ship a regression you cannot see.
+    ///
+    /// Exposed as operator config so the premise can be A/B'd on a real
+    /// workload — full replay is the larger prompt and pushes against
+    /// the prefix-cache budget and the throughput-derived context
+    /// ceiling, but whether it *helps* the model is a measurement, not
+    /// a belief. A request's own `chat_template_kwargs` still wins.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub preserve_thinking: Option<bool>,
 }
 
 /// Per-model token budget advertised by the catalogue or neuron.
@@ -50,8 +188,77 @@ pub struct ModelLimit {
     /// back to `context − output`.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub input: Option<usize>,
-    /// Maximum number of generation tokens.
+    /// The generation budget a request gets when it names none, and the
+    /// figure `input = context − output` is derived from.
+    ///
+    /// A *reserve*, not a ceiling — a request may ask for more and be
+    /// served it. Named `output` for the shape's history; the number a
+    /// client should plan against is [`ModelLimit::output_ceiling`].
     pub output: usize,
+    /// The largest output a request may name and be served (#278).
+    ///
+    /// Distinct from `output`, which is the default and the KV-planning
+    /// reserve. Publishing the reserve as the ceiling meant a client
+    /// that trusted the advertisement asked for a fraction of what the
+    /// model would happily generate — on a reasoning model, often less
+    /// than its own think block costs.
+    ///
+    /// `#[serde(default)]` so a neuron predating this field still
+    /// deserializes; zero means "not advertised", and consumers fall
+    /// back to `output`.
+    #[serde(default, skip_serializing_if = "is_zero")]
+    pub output_ceiling: usize,
+}
+
+fn is_zero(v: &usize) -> bool {
+    *v == 0
+}
+
+/// What each reasoning effort level costs, in reasoning tokens (#223).
+///
+/// OpenAI-shaped clients can only send `minimal|low|medium|high` — the
+/// ladder is their entire vocabulary, and they cannot express a token
+/// count. So the server picks the numbers, and has to say what they are:
+/// a client choosing `low` with no idea whether that buys 2k tokens or
+/// 20k is guessing, which is the failure #274 exists to end.
+///
+/// Ordered rungs rather than a map, so the ladder reads in the order a
+/// client would climb it.
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+pub struct ReasoningBudgetRung {
+    /// The effort level as the caller spells it on the wire.
+    ///
+    /// Since #290 these are the rungs the **model's own chat template**
+    /// accepts, discovered by rendering it, rather than a ladder the
+    /// operator invented. `Qwen/Qwen3.8-27B` accepts exactly `low`,
+    /// `medium` and `xhigh` and raises on anything else; we previously
+    /// advertised `minimal`/`low`/`medium`/`high`, two of which the
+    /// template rejects and which omitted the model's own default.
+    ///
+    /// Advertise only rungs the template accepts: an invented name is a
+    /// name a caller can send and the model will reject. Note that a
+    /// client need not read this at all — pi-ai takes its levels from a
+    /// static `thinkingLevelMap` in the operator's config — so this is
+    /// a statement of truth, not a control surface we can rely on.
+    pub effort: String,
+    /// Backstop reasoning-token cap for this rung, when the deployment
+    /// sets one.
+    ///
+    /// `None` is the normal case and means "named rung, no cap": effort
+    /// is expressed to the model through its template, and the number of
+    /// tokens it then spends is the model's business. A cap is a safety
+    /// net against a runaway think block (#223), not the mechanism by
+    /// which effort is selected — enforcing effort by truncation is what
+    /// #290 fixed.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub tokens: Option<usize>,
+    /// Whether the model applies this rung when the caller names none.
+    ///
+    /// Taken from the template's own `|default(...)`. Without it a
+    /// caller that says nothing cannot tell what it is getting — on
+    /// Qwen3.8 that silently means `xhigh`.
+    #[serde(default, skip_serializing_if = "std::ops::Not::not")]
+    pub default: bool,
 }
 
 /// Operator-set pricing, **USD per 1,000,000 tokens, as JSON numbers**
@@ -126,6 +333,16 @@ pub struct ModelInfo {
     /// marker tokens (`<think>` / `<\/think>` or similar).
     #[serde(default)]
     pub reasoning: bool,
+    /// The operator's `preserve_thinking` for this loaded model, or
+    /// `None` when unset and the template's own default applies.
+    ///
+    /// Advertised so a run can be *stamped* with the value it ran
+    /// under. This is an A/B knob whose whole purpose is comparison
+    /// between runs, and a comparison whose arms cannot be told apart
+    /// afterwards is not a comparison. Recovering it from deploy
+    /// history is guesswork — a config sync can land mid-session.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub preserve_thinking: Option<bool>,
 
     /// Whether this node can actually serve a request for this model
     /// right now (#245).
@@ -144,6 +361,11 @@ pub struct ModelInfo {
     /// from its own routing table.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub servable: Option<ModelServability>,
+    /// The reasoning-effort ladder this deployment honours (#223).
+    /// Absent when the model does not reason, or the harness has no
+    /// budget to offer.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub reasoning_budget: Vec<ReasoningBudgetRung>,
 }
 
 /// Why a loaded model can or cannot be served right now (#245).

crates/cortex-core/src/node.rs

@@ -44,6 +44,12 @@ pub struct NodeState {
     /// just one whose total topology could. Empty until the first
     /// /health poll.
     pub device_health: Vec<crate::discovery::DeviceHealth>,
+    /// The reasoning-effort ladder this neuron honours (#223), copied
+    /// from its `/models` reply at poll time. Per-node rather than
+    /// per-model because it is host configuration — every model on a
+    /// neuron reports the same rungs — and duplicating it onto each
+    /// model entry would invite the two to disagree.
+    pub reasoning_budget: Vec<crate::harness::ReasoningBudgetRung>,
 }
 
 /// A model registered on a node, with its runtime status.
@@ -187,12 +193,35 @@ pub struct CortexModelEntry {
     /// Served max-seq-len in tokens — mirrors `limit.context`.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_model_len: Option<usize>,
+    /// The same window under the two names generic OpenAI clients look
+    /// for (#278) — a client that probes for one of these rather than
+    /// `max_model_len` would otherwise find no window at all and fall
+    /// back to a number its operator had to invent by hand.
+    ///
+    /// Publishing them is cheap and harmless; do not assume any
+    /// *specific* client consumes them. An earlier version of this
+    /// comment credited pi-ai's provider discovery, which was wrong:
+    /// pi requires `contextWindow` as a static field on its model type
+    /// and reads none of these keys at runtime (checked against 0.82.1
+    /// and 0.84.3). The misattribution mattered — it made an operator's
+    /// hand-set 80k window look like our advertisement failing to land,
+    /// when in fact nothing was reading it and the number had to be
+    /// maintained by hand regardless.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub context_window: Option<usize>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub context_length: 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`.
+    /// The largest output a request may name — mirrors
+    /// `limit.output_ceiling`, falling back to `limit.output` (#278).
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub max_output_tokens: Option<usize>,
+    /// What each effort level buys, in reasoning tokens (#223) — so a
+    /// client picking `low` knows what it asked for.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub reasoning_budget: Vec<crate::harness::ReasoningBudgetRung>,
 }
 
 impl CortexModelEntry {
@@ -204,8 +233,26 @@ impl CortexModelEntry {
     /// dropped the limit.
     pub fn sync_flat_limit(&mut self) {
         self.max_model_len = self.limit.as_ref().map(|l| l.context);
+        self.context_window = self.limit.as_ref().map(|l| l.context);
+        self.context_length = 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);
+        // The ceiling, not the reserve (#278): a client that reads this
+        // field turns it into the cap it sends on each request, so
+        // advertising the reserve handed a reasoning model less budget
+        // than its own think block needs. Falls back to the reserve for
+        // a neuron too old to publish a ceiling.
+        //
+        // Observed with dsh. pi is *not* an example: its `maxTokens` is
+        // a required static field in models.json and it reads nothing
+        // from here — an operator has to copy this number across by
+        // hand, and gets no error if they copy it wrong.
+        self.max_output_tokens = self.limit.as_ref().map(|l| {
+            if l.output_ceiling > 0 {
+                l.output_ceiling
+            } else {
+                l.output
+            }
+        });
     }
 }
 

crates/cortex-core/src/openai.rs

@@ -43,6 +43,20 @@ pub struct ChatCompletionRequest {
     /// why it is a knob rather than a constant (#272).
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub repeat_last_n: Option<usize>,
+    /// OpenAI core, `-2.0..=2.0`. Subtracted once from any token that
+    /// has already appeared, regardless of how often.
+    ///
+    /// Unlike `repetition_penalty` this is scored over the **whole**
+    /// generated sequence, not a trailing window — which is what a
+    /// reasoning model needs, and what Qwen recommend (0..2, 1.5 for
+    /// severe cases) for the endless-repetition failure their own
+    /// models exhibit during long thinking.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub presence_penalty: Option<f32>,
+    /// OpenAI core, `-2.0..=2.0`. Subtracted in proportion to how many
+    /// times a token has already appeared.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub frequency_penalty: Option<f32>,
     /// The deprecated spelling of the output cap. Still what most
     /// clients send; see [`ChatCompletionRequest::effective_max_tokens`].
     #[serde(skip_serializing_if = "Option::is_none")]

crates/cortex-core/src/responses.rs

@@ -70,6 +70,11 @@ pub struct ResponsesRequest {
     /// Window `repetition_penalty` considers (#272).
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub repeat_last_n: Option<usize>,
+    /// OpenAI core; forwarded to the chat path unchanged.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub presence_penalty: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub frequency_penalty: Option<f32>,
     /// Chained-conversation identifier. We don't store responses
     /// server-side yet; if this is `Some`, the handler returns 400.
     #[serde(default, skip_serializing_if = "Option::is_none")]
@@ -145,12 +150,22 @@ pub enum ResponsesInputItem {
     /// 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.
+    /// A reasoning item the model emitted on an earlier turn, replayed
+    /// by the client so the model can continue its own train of thought.
+    ///
+    /// Both spellings are modelled because both are in the wild and one
+    /// of them is ours: OpenAI's o-series carries the text in `content`
+    /// as `reasoning_text` parts, while neuron emits `summary` with a
+    /// `summary_text` part. Clients round-trip the completed item
+    /// verbatim — pi-ai stores our `response.output_item.done` payload
+    /// and replays it unchanged — so whichever field we emit is the
+    /// field we get back, and reading only one of them silently loses
+    /// the turn's reasoning.
     Reasoning {
         #[serde(default)]
         content: Vec<Value>,
+        #[serde(default)]
+        summary: Vec<Value>,
     },
 }
 
@@ -205,11 +220,46 @@ pub struct ResponsesResponse {
     pub status: String,
     pub model: String,
     pub output: Vec<ResponsesOutputItem>,
+    /// Why a `status: "incomplete"` response stopped short. Required in
+    /// practice, not optional decoration: a client cannot tell an
+    /// honest truncation from a protocol fault without it, and at least
+    /// one treats the difference as fatal.
+    ///
+    /// pi-ai's `mapStopReason` maps
+    /// `incomplete` + `reason == "max_output_tokens"` to
+    /// `stopReason: "length"`, and `incomplete` with no reason to
+    /// `stopReason: "error"` ("Response incomplete without a provider
+    /// reason"). Its agent loop then either continues — failing any
+    /// truncated tool calls safely and letting the model recover on the
+    /// next turn — or halts outright. Omitting this field cost a live
+    /// agentic session its whole run.
+    ///
+    /// `None` for `completed` and `in_progress`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub incomplete_details: Option<IncompleteDetails>,
     /// Populated on completion; `None` while streaming.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     pub usage: Option<ResponsesUsage>,
 }
 
+/// Why a response is `incomplete`.
+#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct IncompleteDetails {
+    /// OpenAI spells the output-cap case `"max_output_tokens"`; that is
+    /// the only reason we produce today. `content_filter` is the other
+    /// value the upstream API defines.
+    pub reason: String,
+}
+
+impl IncompleteDetails {
+    /// The response ran into the caller's output budget.
+    pub fn max_output_tokens() -> Self {
+        Self {
+            reason: "max_output_tokens".into(),
+        }
+    }
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ResponsesOutputItem {
@@ -226,6 +276,15 @@ pub enum ResponsesOutputItem {
         #[serde(default = "default_item_status")]
         status: String,
     },
+    /// A think block, as its own output item ahead of the message —
+    /// the same shape the streaming projector emits, so a client sees
+    /// one contract whether or not it streams (#300).
+    Reasoning {
+        id: String,
+        summary: Vec<ResponsesSummaryPart>,
+        #[serde(default = "default_item_status")]
+        status: String,
+    },
     /// Reserved for the day tool-call extraction lands. The wire
     /// shape mirrors `ResponsesInputItem::FunctionCall`.
     FunctionCall {
@@ -238,6 +297,24 @@ pub enum ResponsesOutputItem {
     },
 }
 
+/// One `summary_text` part of a reasoning output item.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ResponsesSummaryPart {
+    /// Always `"summary_text"`.
+    #[serde(rename = "type")]
+    pub part_type: String,
+    pub text: String,
+}
+
+impl ResponsesSummaryPart {
+    pub fn summary_text(text: impl Into<String>) -> Self {
+        Self {
+            part_type: "summary_text".into(),
+            text: text.into(),
+        }
+    }
+}
+
 fn default_item_status() -> String {
     "completed".into()
 }
@@ -317,6 +394,24 @@ pub mod events {
     pub const FUNCTION_CALL_ARGUMENTS_DELTA: &str = "response.function_call_arguments.delta";
     pub const FUNCTION_CALL_ARGUMENTS_DONE: &str = "response.function_call_arguments.done";
     pub const COMPLETED: &str = "response.completed";
+    /// Terminal frame for a response that stopped short. Carries the
+    /// same `response` payload as [`COMPLETED`], with
+    /// `status: "incomplete"` and `incomplete_details.reason`.
+    ///
+    /// A client keyed on the event name — rather than on
+    /// `response.status` — sees no terminal event at all if a truncated
+    /// response is announced as `response.completed`, and waits out its
+    /// idle timeout.
+    pub const INCOMPLETE: &str = "response.incomplete";
+    /// Terminal frame for a response that did not finish because the
+    /// server failed — a poisoned device, an OOM, a producer that died
+    /// mid-stream. Carries `status: "failed"` and an `error` object.
+    ///
+    /// The alternative is worse than it looks: defaulting a missing
+    /// finish to a clean stop reports a crash as a complete answer, and
+    /// the caller believes a truncated or empty reply was the model's
+    /// considered response.
+    pub const FAILED: &str = "response.failed";
 }
 
 #[cfg(test)]
@@ -521,6 +616,7 @@ mod tests {
                 }],
                 status: "completed".into(),
             }],
+            incomplete_details: None,
             usage: Some(ResponsesUsage {
                 input_tokens: 5,
                 output_tokens: 3,
@@ -533,5 +629,36 @@ mod tests {
         let parsed: ResponsesResponse = serde_json::from_str(&json).unwrap();
         assert_eq!(parsed.id, "resp_1");
         assert_eq!(parsed.output.len(), 1);
+        assert!(
+            !json.contains("incomplete_details"),
+            "a completed response must not carry an incomplete reason"
+        );
+    }
+
+    /// The field pi-ai's `mapStopReason` keys on. Absent, it maps
+    /// `incomplete` to `stopReason: "error"` and its agent loop halts;
+    /// present with `max_output_tokens`, it maps to `"length"` and the
+    /// loop continues.
+    #[test]
+    fn an_incomplete_response_serialises_its_reason() {
+        let r = ResponsesResponse {
+            id: "resp_2".into(),
+            object: "response".into(),
+            created_at: 1700,
+            status: "incomplete".into(),
+            model: "m".into(),
+            output: vec![],
+            incomplete_details: Some(IncompleteDetails::max_output_tokens()),
+            usage: None,
+        };
+        let v: serde_json::Value = serde_json::to_value(&r).unwrap();
+        assert_eq!(v["status"], "incomplete");
+        assert_eq!(v["incomplete_details"]["reason"], "max_output_tokens");
+
+        let parsed: ResponsesResponse = serde_json::from_value(v).unwrap();
+        assert_eq!(
+            parsed.incomplete_details,
+            Some(IncompleteDetails::max_output_tokens())
+        );
     }
 }

crates/cortex-core/src/translate.rs

@@ -94,6 +94,28 @@ pub fn anthropic_to_openai(req: MessagesRequest) -> ChatCompletionRequest {
         if let Some(tc) = tool_choice {
             obj.insert("tool_choice".into(), tc);
         }
+        // Anthropic's extended-thinking budget is the one quantitative
+        // reasoning control any of our wire formats offers (#223).
+        // Normalised onto `reasoning.max_tokens`, the shape the harness
+        // resolves, so all three inlets converge on one field instead of
+        // each generation loop learning three spellings.
+        let budget = obj
+            .get("thinking")
+            .filter(|t| t.get("type").and_then(Value::as_str) != Some("disabled"))
+            .and_then(|t| t.get("budget_tokens"))
+            .and_then(Value::as_u64)
+            .filter(|n| *n > 0);
+        if let Some(budget) = budget {
+            let reasoning = obj
+                .entry("reasoning")
+                .or_insert_with(|| Value::Object(Default::default()));
+            if !reasoning.is_object() {
+                *reasoning = Value::Object(Default::default());
+            }
+            if let Some(r) = reasoning.as_object_mut() {
+                r.entry("max_tokens").or_insert_with(|| Value::from(budget));
+            }
+        }
     }
 
     ChatCompletionRequest {
@@ -108,6 +130,8 @@ pub fn anthropic_to_openai(req: MessagesRequest) -> ChatCompletionRequest {
         seed: None,
         repetition_penalty: None,
         repeat_last_n: None,
+        presence_penalty: None,
+        frequency_penalty: None,
         max_tokens: Some(req.max_tokens),
         // Anthropic has one spelling; it maps onto the legacy field.
         max_completion_tokens: None,
@@ -391,17 +415,26 @@ fn anthropic_content_to_text(content: AnthropicContent) -> String {
 pub fn openai_to_anthropic(resp: ChatCompletionResponse) -> MessagesResponse {
     let choice = resp.choices.into_iter().next();
 
-    let (content_text, stop_reason) = match choice {
+    let (content_text, reasoning_text, stop_reason) = match choice {
         Some(c) => {
             let text = match c.message.content {
                 MessageContent::Text(t) => t,
                 MessageContent::Parts(parts) => serde_json::to_string(&parts).unwrap_or_default(),
                 MessageContent::Null => String::new(),
             };
+            // The model's think block, which the harness returns on
+            // `message.reasoning_content` (#300).
+            let reasoning = c
+                .message
+                .extra
+                .get("reasoning_content")
+                .and_then(|v| v.as_str())
+                .map(str::to_string)
+                .filter(|t| !t.is_empty());
             let stop = c.finish_reason.map(|r| map_stop_reason(&r));
-            (text, stop)
+            (text, reasoning, stop)
         }
-        None => (String::new(), None),
+        None => (String::new(), None, None),
     };
 
     let usage = resp.usage.unwrap_or(Usage {
@@ -417,10 +450,32 @@ pub fn openai_to_anthropic(resp: ChatCompletionResponse) -> MessagesResponse {
         id: resp.id,
         response_type: "message".into(),
         role: "assistant".into(),
-        content: vec![ContentBlock {
-            block_type: "text".into(),
-            data: json!({ "text": content_text }),
-        }],
+        // A `thinking` block ahead of the text, which is Anthropic's
+        // own ordering. This path accepted `thinking.budget_tokens` on
+        // the way in (see `anthropic_to_openai`) and returned nothing on
+        // the way out: the caller spent the tokens, could not see them,
+        // and could not replay the model's reasoning next turn (#300).
+        //
+        // Emitted without a `signature`. Anthropic signs its thinking
+        // blocks and a client that round-trips one may expect the field;
+        // we cannot mint a valid signature and inventing one would be
+        // worse than omitting it. If a real client rejects an unsigned
+        // block, the fix is to gate this behind a flag — not to go back
+        // to silently deleting the model's work.
+        content: {
+            let mut blocks = Vec::with_capacity(2);
+            if let Some(thinking) = reasoning_text {
+                blocks.push(ContentBlock {
+                    block_type: "thinking".into(),
+                    data: json!({ "thinking": thinking }),
+                });
+            }
+            blocks.push(ContentBlock {
+                block_type: "text".into(),
+                data: json!({ "text": content_text }),
+            });
+            blocks
+        },
         model: resp.model,
         stop_reason,
         usage: AnthropicUsage {
@@ -1039,3 +1094,66 @@ mod request_tests {
         assert_eq!(tools[0]["function"]["name"], "x");
     }
 }
+
+/// The Anthropic path must return the model's thinking (#300).
+///
+/// It accepted `thinking.budget_tokens` on the way in and returned
+/// a bare text block on the way out: the caller spent the tokens,
+/// never saw them, and could not replay the model's reasoning on
+/// the next turn. Taking a parameter and silently discarding what
+/// it buys is the defect, not the missing field.
+#[test]
+fn anthropic_response_carries_a_thinking_block() {
+    let mut extra = serde_json::Map::new();
+    extra.insert("reasoning_content".into(), json!("weighing the options"));
+    let resp = ChatCompletionResponse {
+        id: "id".into(),
+        object: "chat.completion".into(),
+        created: 0,
+        model: "m".into(),
+        choices: vec![crate::openai::ChatCompletionChoice {
+            index: 0,
+            message: ChatMessage {
+                role: "assistant".into(),
+                content: MessageContent::Text("the answer".into()),
+                extra: Value::Object(extra),
+            },
+            finish_reason: Some("stop".into()),
+            extra: Value::Null,
+        }],
+        usage: None,
+        extra: Value::Null,
+    };
+    let a = openai_to_anthropic(resp);
+    assert_eq!(a.content.len(), 2, "expected thinking + text");
+    assert_eq!(a.content[0].block_type, "thinking", "thinking comes first");
+    assert_eq!(a.content[0].data["thinking"], "weighing the options");
+    assert_eq!(a.content[1].block_type, "text");
+    assert_eq!(a.content[1].data["text"], "the answer");
+}
+
+/// No thinking means the unchanged single text block.
+#[test]
+fn anthropic_response_without_thinking_is_unchanged() {
+    let resp = ChatCompletionResponse {
+        id: "id".into(),
+        object: "chat.completion".into(),
+        created: 0,
+        model: "m".into(),
+        choices: vec![crate::openai::ChatCompletionChoice {
+            index: 0,
+            message: ChatMessage {
+                role: "assistant".into(),
+                content: MessageContent::Text("plain".into()),
+                extra: Value::Object(Default::default()),
+            },
+            finish_reason: Some("stop".into()),
+            extra: Value::Null,
+        }],
+        usage: None,
+        extra: Value::Null,
+    };
+    let a = openai_to_anthropic(resp);
+    assert_eq!(a.content.len(), 1);
+    assert_eq!(a.content[0].block_type, "text");
+}

crates/cortex-gateway/src/handlers.rs

@@ -798,6 +798,10 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                 max_model_len: None,
                 max_input_tokens: None,
                 max_output_tokens: None,
+                context_window: None,
+                context_length: None,
+                // Catalogue-seeded: no neuron has reported rungs yet.
+                reasoning_budget: Vec::new(),
             },
         );
     }
@@ -839,6 +843,13 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                     // so a client never overflows the most-constrained
                     // deployment that might serve it.
                     e.limit = tightest_limit(e.limit.take(), entry.limit.clone());
+                    // The reasoning ladder this neuron honours (#223).
+                    // First non-empty wins: a catalogue-seeded entry
+                    // starts with none, and neurons serving the same
+                    // model are configured alike.
+                    if e.reasoning_budget.is_empty() {
+                        e.reasoning_budget = node.reasoning_budget.clone();
+                    }
                 })
                 .or_insert_with(|| CortexModelEntry {
                     id: model_id.clone(),
@@ -858,6 +869,10 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                     max_model_len: None,
                     max_input_tokens: None,
                     max_output_tokens: None,
+                    context_window: None,
+                    context_length: None,
+                    // What this neuron says each effort level buys (#223).
+                    reasoning_budget: node.reasoning_budget.clone(),
                 });
         }
     }
@@ -919,6 +934,10 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                     max_model_len: None,
                     max_input_tokens: None,
                     max_output_tokens: None,
+                    context_window: None,
+                    context_length: None,
+                    // What this neuron says each effort level buys (#223).
+                    reasoning_budget: node.reasoning_budget.clone(),
                 });
         }
     }
@@ -956,6 +975,10 @@ async fn list_models(State(fleet): State<Arc<CortexState>>) -> Json<Value> {
                 max_model_len: None,
                 max_input_tokens: None,
                 max_output_tokens: None,
+                context_window: None,
+                context_length: None,
+                // Catalogue-seeded: no neuron has reported rungs yet.
+                reasoning_budget: Vec::new(),
             },
         );
     }

crates/cortex-gateway/src/metering.rs

@@ -141,6 +141,15 @@ pub fn usage_sink(
 /// 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`).
+// The `Err` variant is the #63 envelope itself, which is a data-carrying
+// struct (status, type, code, message, param, retry hint, diagnostics) and
+// is passed by value everywhere it is built, matched and rendered. Boxing it
+// at this one boundary would push `Box` conversions through three handlers
+// and both HTTP adapters to save an allocation on a path taken once per
+// rejected request. clippy 1.98 lowered what it considers large enough to
+// flag; the shape is deliberate. Same call as the existing allows on the
+// figment-returning config loaders.
+#[allow(clippy::result_large_err)]
 pub async fn reserve_or_reject(
     provider: Arc<dyn EntitlementProvider>,
     principal: &Principal,
@@ -279,7 +288,7 @@ mod tests {
         // 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}");
+        assert!((1000..1100).contains(&est), "est was {est}");
     }
 
     #[test]

crates/cortex-gateway/src/poller.rs

@@ -130,6 +130,17 @@ async fn poll_neuron(fleet: &CortexState, name: &str, endpoint: &str) {
         Ok(resp) if resp.status().is_success() => {
             match resp.json::<Vec<ModelInfo>>().await {
                 Ok(models) => {
+                    // The reasoning-effort ladder is host configuration:
+                    // every model on a neuron reports the same rungs, so
+                    // it is held once per node rather than copied onto
+                    // each model entry where the two could drift (#223).
+                    // Taken from the first model that offers one — a
+                    // non-reasoning model advertises none.
+                    node.reasoning_budget = models
+                        .iter()
+                        .map(|m| m.reasoning_budget.clone())
+                        .find(|rungs| !rungs.is_empty())
+                        .unwrap_or_default();
                     let mut seen = std::collections::HashSet::new();
                     for upstream in &models {
                         seen.insert(upstream.id.clone());

crates/cortex-gateway/src/router.rs

@@ -535,6 +535,11 @@ async fn profile_to_spec(
         quant: profile.quant.clone(),
         tensor_parallel,
         devices: Some(devices),
+        // The catalogue's operator override rides along with the load
+        // (#283), so a cold-loaded model samples the same way as one
+        // this host was configured to hold resident.
+        sampling: profile.sampling.clone(),
+        preserve_thinking: profile.preserve_thinking,
     }
 }
 
@@ -642,6 +647,8 @@ mod tests {
             pinned_on: vec![],
             residency_priority: None,
             source: source.map(String::from),
+            sampling: None,
+            preserve_thinking: None,
             limit: None,
             cost: None,
             capabilities: vec![],

crates/cortex-gateway/src/state.rs

@@ -66,6 +66,9 @@ impl CortexState {
                     endpoint: nc.endpoint.clone(),
                     healthy: false,
                     models: HashMap::new(),
+                    // Filled from the neuron's /models reply at the
+                    // first poll (#223).
+                    reasoning_budget: Vec::new(),
                     lifecycle_cycles: 0,
                     last_poll: None,
                     discovery: None,

crates/cortex-gateway/tests/budget_enforcement.rs

@@ -1,9 +1,9 @@
 //! 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.
+//! with the #63 code matching the cap-window semantics: `rate_limit_exceeded`
+//! plus `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;

crates/cortex-gateway/tests/model_limits.rs

@@ -69,6 +69,21 @@ capabilities = ["text"]
         let mut nodes = fleet.nodes.write().await;
         let node = nodes.get_mut("mock-node").expect("node exists");
         node.healthy = true;
+        // The rungs the neuron reported at poll time. Since #290 these
+        // come from the model's own chat template rather than operator
+        // configuration, so the fixture uses the shape a real reasoning
+        // model ships: Qwen3.8-27B accepts exactly low/medium/xhigh and
+        // defaults to xhigh.
+        node.reasoning_budget = ["low", "medium", "xhigh"]
+            .into_iter()
+            .map(|effort| cortex_core::harness::ReasoningBudgetRung {
+                default: effort == "xhigh",
+                effort: effort.into(),
+                // No cap: effort is expressed to the model through its
+                // template, not enforced by truncating its thinking.
+                tokens: None,
+            })
+            .collect();
         node.models.insert(
             "test-model".into(),
             ModelEntry {
@@ -85,6 +100,9 @@ capabilities = ["text"]
                     context: 49152,
                     input: Some(40960),
                     output: 8192,
+                    // The ceiling a request may name, distinct from the
+                    // 8192-token reserve above (#278).
+                    output_ceiling: 32768,
                 }),
                 servable: None,
             },
@@ -147,7 +165,53 @@ capabilities = ["text"]
     // vLLM-convention probes (Hermes Agent) auto-detect the window.
     assert_eq!(entry["max_model_len"], 49152);
     assert_eq!(entry["max_input_tokens"], 40960);
-    assert_eq!(entry["max_output_tokens"], 8192);
+    // The ceiling, not the reserve (#278). A client that reads this
+    // field turns it into the cap it sends on each request; advertising
+    // the 8192 reserve told a reasoning model's harness to stop below
+    // the cost of a single think block. Observed with dsh — pi takes
+    // `maxTokens` from static config and reads nothing here.
+    assert_eq!(entry["max_output_tokens"], 32768);
+    // The rungs the model itself accepts (#290). A caller can only send
+    // a rung name, so a name we invent is one the template will reject.
+    // Publishing the true set is correct regardless of who reads it —
+    // pi-ai, for one, takes its levels from static config instead.
+    let rungs = entry["reasoning_budget"]
+        .as_array()
+        .expect("reasoning_budget advertised for a reasoning model");
+    let names: Vec<&str> = rungs
+        .iter()
+        .map(|r| r["effort"].as_str().expect("effort"))
+        .collect();
+    assert_eq!(names, ["low", "medium", "xhigh"]);
+    // Which rung applies when the caller names none. Without this a
+    // silent caller cannot tell what it is getting — on Qwen3.8 that is
+    // xhigh, the most expensive rung on the ladder.
+    let default_rung: Vec<&str> = rungs
+        .iter()
+        .filter(|r| r["default"].as_bool().unwrap_or(false))
+        .map(|r| r["effort"].as_str().expect("effort"))
+        .collect();
+    assert_eq!(
+        default_rung,
+        ["xhigh"],
+        "the model's own default must be visible"
+    );
+    // A named rung carries no token count: effort reaches the model
+    // through its template, and enforcing it by truncation is the defect
+    // #290 fixed.
+    assert!(
+        rungs.iter().all(|r| r.get("tokens").is_none()),
+        "a named rung must not masquerade as a token budget"
+    );
+
+    // The window under the two names a generic client may read, for one
+    // that probes these rather than `max_model_len`. Publish all three;
+    // do not assume a particular client consumes any of them. pi, in
+    // particular, reads none — `contextWindow` is a required static
+    // field on its model type (checked against 0.82.1 and 0.84.3), so
+    // its window is hand-maintained whatever we advertise.
+    assert_eq!(entry["context_window"], 49152);
+    assert_eq!(entry["context_length"], 49152);
 
     // No limit → flat fields omitted entirely, never 0 or a guess.
     let unknown = body["data"]
@@ -157,7 +221,13 @@ capabilities = ["text"]
         .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"] {
+    for key in [
+        "max_model_len",
+        "max_input_tokens",
+        "max_output_tokens",
+        "context_window",
+        "context_length",
+    ] {
         assert!(
             unknown.get(key).is_none(),
             "{key} must be omitted when the window is unknown"

crates/cortex-gateway/tests/prompt_prevalidation.rs

@@ -96,6 +96,7 @@ async fn spawn_gateway(neuron: &str, context: usize) -> String {
                     context,
                     input: None,
                     output: 16,
+                    output_ceiling: 0,
                 }),
                 servable: None,
             },

crates/helexa-bench/src/api.rs

@@ -34,6 +34,7 @@ pub fn api_routes(state: ApiState) -> Router {
     Router::new()
         .route("/api/health", get(health))
         .route("/api/dimensions", get(dimensions))
+        .route("/api/regimes", get(regimes))
         .route("/api/summary", get(summary))
         .route("/api/scaling", get(scaling))
         .route("/api/concurrency", get(concurrency))
@@ -65,6 +66,14 @@ async fn health(State(s): State<ApiState>) -> Result<Json<serde_json::Value>, Ap
     Ok(Json(json!({ "status": "ok", "run_count": count })))
 }
 
+/// Declared measurement-regime boundaries (#288).
+///
+/// Static — no store access — but served from the API rather than
+/// hardcoded in the frontend so `report` and the UI cite one list.
+async fn regimes() -> Json<&'static [crate::regime::MeasurementRegime]> {
+    Json(crate::regime::REGIMES)
+}
+
 async fn dimensions(State(s): State<ApiState>) -> Result<Json<crate::store::Dimensions>, ApiError> {
     let store = s.lock().await;
     store.dimensions().map(Json).map_err(err500)

crates/helexa-bench/src/client.rs

@@ -6,6 +6,7 @@ use crate::config::{TargetConfig, TargetKind};
 use anyhow::{Context, Result, anyhow};
 use cortex_core::build_info::BuildInfo;
 use cortex_core::discovery::{DiscoveryResponse, HealthResponse};
+use cortex_core::entitlements::{HEADER_ACCOUNT_ID, HEADER_KEY_ID};
 use cortex_core::harness::{ModelInfo, ModelSpec};
 use cortex_core::openai::ModelsResponse;
 use std::time::Duration;
@@ -19,10 +20,49 @@ pub struct TargetClient {
 
 impl TargetClient {
     pub fn new(request_timeout: Duration) -> Result<Self> {
-        let http = reqwest::Client::builder()
-            .timeout(request_timeout)
-            .build()
-            .context("building HTTP client")?;
+        Self::with_principal(request_timeout, None)
+    }
+
+    /// Build a client that stamps `principal`'s identity headers on every
+    /// request (#288).
+    ///
+    /// Set on the client rather than at each call site so no scenario can
+    /// be added later that silently measures as anonymous — which is the
+    /// failure this fixes, and it was invisible for a week.
+    ///
+    /// `None` keeps the historical anonymous behaviour, which is correct
+    /// for `openai` targets (the headers mean nothing to a foreign
+    /// engine) and for a fresh install with no principal configured.
+    pub fn with_principal(
+        request_timeout: Duration,
+        principal: Option<&crate::config::PrincipalSettings>,
+    ) -> Result<Self> {
+        let mut builder = reqwest::Client::builder().timeout(request_timeout);
+        if let Some(p) = principal {
+            let mut headers = reqwest::header::HeaderMap::new();
+            headers.insert(
+                HEADER_ACCOUNT_ID,
+                reqwest::header::HeaderValue::from_str(&p.account_id)
+                    .context("principal.account_id is not a valid header value")?,
+            );
+            headers.insert(
+                HEADER_KEY_ID,
+                reqwest::header::HeaderValue::from_str(&p.key_id)
+                    .context("principal.key_id is not a valid header value")?,
+            );
+            tracing::info!(
+                account_id = %p.account_id,
+                key_id = %p.key_id,
+                "bench authenticates: measurements reflect identified-caller capacity (#288)"
+            );
+            builder = builder.default_headers(headers);
+        } else {
+            tracing::warn!(
+                "bench has no [bench.principal]: requests are anonymous, so since #262 \
+                 they measure the anonymous-yield policy rather than serving capacity (#288)"
+            );
+        }
+        let http = builder.build().context("building HTTP client")?;
         Ok(TargetClient { http })
     }
 
@@ -169,6 +209,15 @@ impl TargetClient {
             quant: None,
             tensor_parallel: (info.devices.len() > 1).then_some(info.devices.len() as u32),
             devices: Some(info.devices.clone()),
+            // `GET /models` does not report the operator sampling
+            // override (#283), so a spec reconstructed here cannot carry
+            // it — same limitation as `quant` directly above. Benign
+            // while this is only used to re-load a model the host
+            // already has a `[[default_models]]` entry for, since that
+            // entry supplies the override; it would matter if this were
+            // ever used to load a model the host has no config for.
+            sampling: None,
+            preserve_thinking: None,
         })
     }
 
@@ -233,9 +282,11 @@ impl TargetClient {
                         cost: None,
                         tool_call: false,
                         reasoning: false,
+                        preserve_thinking: None,
                         // A generic OpenAI endpoint says nothing about
                         // serviceability, and absent means "no opinion".
                         servable: None,
+                        reasoning_budget: Vec::new(),
                     })
                     .collect())
             }

crates/helexa-bench/src/config.rs

@@ -67,6 +67,44 @@ pub struct BenchSettings {
     /// SQLite system-of-record path.
     #[serde(default = "default_db_path")]
     pub db_path: String,
+    /// Identity bench presents to a neuron (#288).
+    ///
+    /// Without this bench is *anonymous*, and since #262 anonymous
+    /// callers are served only from capacity left over once identified
+    /// traffic is satisfied — they are capped below `max_in_flight` and
+    /// park at the class gate ahead of every other gate. Measured on
+    /// beast, that moved `concurrency:8` from 168.9 tok/s / 0.53 s
+    /// ttft_p95 to 104.6 / 11.77 s the day #262 landed, and held there.
+    ///
+    /// So an unauthenticated bench does not measure serving capacity; it
+    /// measures the yield policy. Real traffic arrives through cortex
+    /// with a principal and is subject to none of it.
+    ///
+    /// Deliberately not an admission *exemption*: bench should be
+    /// subject to the same rules as any caller, or it measures something
+    /// no user experiences — the same class of error in the other
+    /// direction. Give it a principal and let it queue like everyone
+    /// else.
+    #[serde(default)]
+    pub principal: Option<PrincipalSettings>,
+}
+
+/// The account/key pair bench stamps on inference requests (#288).
+///
+/// These are the headers cortex asserts after a bearer resolves (#49).
+/// Bench talks to neurons **directly** over the WireGuard mesh, so it
+/// stamps them itself — the same trust model the link already relies on,
+/// since a neuron accepts them from cortex on exactly that basis.
+///
+/// Give bench its *own* pair rather than borrowing a real caller's: it
+/// then holds its own fair-share allocation (#54) and cannot starve
+/// interactive traffic, which is the failure #262 was fixing.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct PrincipalSettings {
+    /// Stamped as `x-helexa-account-id`.
+    pub account_id: String,
+    /// Stamped as `x-helexa-key-id`.
+    pub key_id: String,
 }
 
 impl Default for BenchSettings {
@@ -77,6 +115,7 @@ impl Default for BenchSettings {
             iteration_pause_secs: default_iter_pause(),
             request_timeout_secs: default_timeout(),
             db_path: default_db_path(),
+            principal: None,
         }
     }
 }
@@ -110,6 +149,18 @@ pub struct ScenarioConfig {
     /// deliberately, e.g. `concurrency_levels = [2, 4, 8]`.
     #[serde(default)]
     pub concurrency_levels: Vec<u32>,
+    /// Concurrency levels to run **non-streaming** (#285). Separate from
+    /// `concurrency_levels` because the two shapes are not
+    /// interchangeable: streaming multiplexes through the batch engine,
+    /// non-streaming did not until #285 and serialized completely — a
+    /// measured 1.00x aggregate throughput at every level against 3.98x
+    /// streamed.
+    ///
+    /// Defaults to the top level only. The signal is starkest there and
+    /// each extra level costs a full burst per sample per build; an
+    /// operator who wants the whole curve can list it.
+    #[serde(default = "default_concurrency_nonstreaming_levels")]
+    pub concurrency_nonstreaming_levels: Vec<u32>,
     /// Square image sizes (px per side) — one `image:<px>` scenario per
     /// entry (#203), run only against models advertising the `image`
     /// capability. Defaults to `[1024]`.
@@ -148,6 +199,7 @@ impl Default for ScenarioConfig {
             prompt_sizes: default_prompt_sizes(),
             max_tokens: default_max_tokens(),
             concurrency_levels: Vec::new(),
+            concurrency_nonstreaming_levels: Vec::new(),
             concurrency_prompt_tokens: default_concurrency_prompt_tokens(),
             image_sizes: default_image_sizes(),
             capability_probes: Vec::new(),
@@ -215,6 +267,10 @@ fn default_timeout() -> u64 {
 fn default_db_path() -> String {
     "/var/lib/helexa-bench/bench.sqlite".to_string()
 }
+fn default_concurrency_nonstreaming_levels() -> Vec<u32> {
+    vec![8]
+}
+
 fn default_api_enabled() -> bool {
     true
 }
@@ -243,6 +299,44 @@ fn default_capability_max_tokens() -> u64 {
 // figment's, not ours, so suppress the lint here.
 #[allow(clippy::result_large_err)]
 mod tests {
+
+    /// #288: a config with no `[bench.principal]` must still parse — the
+    /// anonymous behaviour is the historical one and correct for
+    /// `openai` targets — but a configured one must round-trip, because
+    /// a silently-dropped principal reverts the measurement to the
+    /// anonymous-yield path without changing a visible number.
+    #[test]
+    fn principal_is_optional_and_round_trips() {
+        let parse = |t: &str| -> BenchConfig {
+            figment::Figment::new()
+                .merge(figment::providers::Toml::string(t))
+                .extract()
+                .expect("parse")
+        };
+        let none = parse("[bench]\nsamples_per_version = 3\n");
+        assert!(none.bench.principal.is_none());
+
+        let some = parse(
+            "[bench]\nsamples_per_version = 3\n\n\
+             [bench.principal]\naccount_id = \"acct\"\nkey_id = \"k1\"\n",
+        );
+        let p = some.bench.principal.expect("principal");
+        assert_eq!((p.account_id.as_str(), p.key_id.as_str()), ("acct", "k1"));
+    }
+
+    /// The non-streaming burst defaults on (#285). If it defaulted off,
+    /// the path that serialized completely would stay unmeasured on every
+    /// fleet that had not opted in — which is how it went unnoticed.
+    #[test]
+    fn nonstreaming_concurrency_defaults_to_the_top_level() {
+        let cfg: BenchConfig = figment::Figment::new()
+            .merge(figment::providers::Toml::string(
+                "[scenarios]\nprompt_sizes = [128]\n",
+            ))
+            .extract()
+            .expect("parse");
+        assert_eq!(cfg.scenarios.concurrency_nonstreaming_levels, vec![8]);
+    }
     use super::*;
     use figment::Jail;
 

crates/helexa-bench/src/lib.rs

@@ -7,6 +7,7 @@
 pub mod api;
 pub mod client;
 pub mod config;
+pub mod regime;
 pub mod report;
 pub mod scenario;
 pub mod store;

crates/helexa-bench/src/regime.rs

@@ -0,0 +1,148 @@
+//! Measurement-regime boundaries — builds where a number changed
+//! meaning rather than value.
+//!
+//! A trend line silently spans them, and every reader re-derives the
+//! same thing: that a cliff was the instrument, not the engine. Three
+//! have already cost real investigation time:
+//!
+//! - beast's `concurrency:8` fell 168.9 → 104.6 tok/s with `ttft_p95`
+//!   rising 0.53 → 11.77 s, and it read as a 38% throughput regression
+//!   until the one commit in the window turned out to be an admission
+//!   *policy* change that bench, being anonymous, was subject to (#288).
+//! - quadbrat's `capability:debug-reason` TTFT fell 88.8 → 0.1 s, which
+//!   read as a latency win and was a redefinition of when the clock
+//!   stops.
+//! - the same redefinition put decode rates at ~250 tok/s on a 3060
+//!   beforehand, by dividing reasoning-inclusive token counts by a
+//!   visible-content-only window.
+//!
+//! So the boundaries are declared, not remembered. Declared here rather
+//! than in the frontend because `report` renders the same series into
+//! `benchmarks.md` and needs the same caveats.
+//!
+//! ## What belongs here
+//!
+//! Only changes to **what a metric means**. A change that made the
+//! server genuinely faster, slower, or differently behaved does *not*
+//! belong — that is the signal these lines exist to keep readable. #280
+//! (fetch the model's own chat template) moved quadbrat's
+//! `completion_tokens` from 4096 to 1053 by ending a truncation loop;
+//! that is a real behavioural change and stays unmarked.
+//!
+//! The identified/anonymous split is deliberately absent too: it is
+//! recorded per-row in `runs.principal`, so the UI derives it from the
+//! data instead of trusting a constant to stay true.
+
+use serde::Serialize;
+
+/// A build at which one or more metrics changed meaning.
+#[derive(Debug, Clone, Serialize)]
+pub struct MeasurementRegime {
+    /// Short SHA of the first build measured under the new semantics.
+    /// Matched against `SeriesPoint::git_sha`, so a boundary whose build
+    /// never appears in a selection simply does not draw.
+    pub first_sha: &'static str,
+    /// Terse label for the chart rule.
+    pub label: &'static str,
+    /// What changed, and why the step is not a performance result.
+    pub detail: &'static str,
+    /// Series keys the change affects, as the API/UI name them. A
+    /// boundary only draws on panels carrying one of these — marking
+    /// every panel would train people to ignore the rules.
+    pub affects: &'static [&'static str],
+}
+
+/// The declared boundaries, oldest first.
+pub const REGIMES: &[MeasurementRegime] = &[
+    MeasurementRegime {
+        first_sha: "92dddc0",
+        label: "#262 anonymous callers yield",
+        detail: "Anonymous traffic became servable only from capacity left over \
+                 once identified traffic is satisfied — capped below max_in_flight \
+                 and parked at the class gate. An anonymous bench therefore \
+                 characterises the yield policy rather than serving capacity. \
+                 Steps here are admission policy, not the engine (#288).",
+        affects: &["decode", "ttft", "ttftP95", "queueWait", "rejected"],
+    },
+    MeasurementRegime {
+        first_sha: "0fe7aa3",
+        label: "#117 reasoning counts as liveness",
+        detail: "TTFT previously waited for the first *visible content* chunk, so \
+                 on a thinking model it measured the whole reasoning span (88.8 s \
+                 on quadbrat). It now stops at the first delta of any kind. The \
+                 same change fixed decode rates computed by dividing \
+                 reasoning-inclusive token counts by a content-only window, which \
+                 had produced ~250 tok/s on a 3060.",
+        affects: &["decode", "ttft"],
+    },
+];
+
+/// Boundaries affecting `metric`, in declaration order.
+pub fn regimes_for(metric: &str) -> Vec<&'static MeasurementRegime> {
+    REGIMES
+        .iter()
+        .filter(|r| r.affects.contains(&metric))
+        .collect()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// A boundary that names no metric would draw nowhere and mislead by
+    /// omission; one naming an unknown key would draw nowhere and look
+    /// fine. Both are typos this catches.
+    #[test]
+    fn every_regime_affects_a_known_series_key() {
+        // The keys `Trends.tsx` maps `SeriesPoint` onto.
+        const KNOWN: &[&str] = &[
+            "decode",
+            "ttft",
+            "ttftP95",
+            "queueWait",
+            "rejected",
+            "prefillTps",
+            "reasoning",
+            "cached",
+            "completion",
+            "tpot",
+        ];
+        for r in REGIMES {
+            assert!(
+                !r.affects.is_empty(),
+                "{} affects nothing, so it would never draw",
+                r.first_sha
+            );
+            for key in r.affects {
+                assert!(
+                    KNOWN.contains(key),
+                    "{} names unknown series key {key:?}",
+                    r.first_sha
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn regimes_for_selects_by_metric() {
+        assert!(regimes_for("ttft").iter().any(|r| r.first_sha == "0fe7aa3"));
+        assert!(
+            regimes_for("completion").is_empty(),
+            "#280 changed behaviour, not meaning — it must not be marked"
+        );
+    }
+
+    /// Short SHAs are matched against `SeriesPoint::git_sha`, which the
+    /// store records at the width `/version` reports.
+    #[test]
+    fn declared_shas_are_short_form() {
+        for r in REGIMES {
+            assert_eq!(
+                r.first_sha.len(),
+                7,
+                "{} is not a 7-char short sha; it would never match a series point",
+                r.first_sha
+            );
+        }
+    }
+}

crates/helexa-bench/src/scenario.rs

@@ -9,6 +9,7 @@
 use crate::config::ScenarioConfig;
 use anyhow::{Context, Result, anyhow};
 use async_trait::async_trait;
+use cortex_core::entitlements::HEADER_KEY_ID;
 use cortex_core::harness::ModelInfo;
 use cortex_core::openai::ChatCompletionChunk;
 use eventsource_stream::Eventsource;
@@ -46,6 +47,17 @@ pub struct RunCtx<'a> {
     pub model_id: String,
     pub max_tokens: u64,
     pub timeout: Duration,
+    /// Base `x-helexa-key-id` bench presents (#288), when configured.
+    ///
+    /// The client already sends this as a default header, so every
+    /// scenario is identified without having to remember. A concurrency
+    /// burst overrides it per stream — see
+    /// [`ConcurrencyScenario`] — because fair-share (#54) bounds one
+    /// principal to `max_per_principal` in-flight-or-queued requests,
+    /// which defaults to 2. Eight streams under one identity would be
+    /// capped at two and the rest rejected, measuring the fair-share cap
+    /// instead of the server.
+    pub principal_key_id: Option<String>,
 }
 
 /// Operator-felt metrics for a single measured request.
@@ -72,6 +84,34 @@ pub struct ScenarioMetrics {
     pub decode_ms: Option<u64>,
     /// Tokens submitted to prefill — the denominator for prefill tok/s.
     pub prefill_tokens: Option<u64>,
+    /// Tokens spent inside the reasoning span, from
+    /// `usage.completion_tokens_details.reasoning_tokens` (#223). A
+    /// sub-count of `completion_tokens`.
+    ///
+    /// Worth a trend of its own: on a reasoning model this is the
+    /// dominant cost driver and it moves independently of speed. A
+    /// template change, a sampling change (#283) or a reasoning-budget
+    /// change can double what the model thinks before answering while
+    /// every tok/s number stays flat — the bill doubles and no chart
+    /// moves.
+    pub reasoning_tokens: Option<u64>,
+    /// p95 inter-token arrival gap in milliseconds — the tail of the
+    /// stream's smoothness, client-observed.
+    ///
+    /// `decode_tps` is a mean over the whole decode window, so a stream
+    /// that stalls for a second and then catches up is indistinguishable
+    /// from one that never stalled. This is the number a user actually
+    /// feels, and the one a batching stall or a mid-stream rebatch shows
+    /// up in. `None` for non-streaming, which has no inter-token gaps by
+    /// construction, and for streams too short to have a tail.
+    pub tpot_p95_ms: Option<f64>,
+    /// Prompt tokens served from neuron's prefix KV cache (#269), from
+    /// `usage.prompt_tokens_details.cached_tokens`.
+    ///
+    /// The reason prefill timing varies between otherwise identical
+    /// samples. #269 exists because the saving was invisible and every
+    /// client reported a 0% hit rate; bench then threw the number away.
+    pub cached_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
@@ -203,6 +243,9 @@ impl Scenario for ImageLatencyScenario {
                 .as_u64()
                 .map(|d| d + timing["decode_ms"].as_u64().unwrap_or(0)),
             prefill_tokens: None,
+            reasoning_tokens: None,
+            cached_tokens: None,
+            tpot_p95_ms: None,
             concurrency: None,
             ttft_p95_s: None,
             queue_wait_ms_median: None,
@@ -233,6 +276,18 @@ pub fn build_scenarios(cfg: &ScenarioConfig) -> Vec<Box<dyn Scenario>> {
             id: format!("concurrency:{n}"),
             concurrency: n,
             approx_prompt_tokens: cfg.concurrency_prompt_tokens,
+            streaming: true,
+        }) as Box<dyn Scenario>);
+    }
+    // Non-streaming bursts (#285). A separate cell id so the two shapes
+    // never average together — they measured 1.00x against 3.98x before
+    // #285, and a combined number would have hidden both.
+    for &n in &cfg.concurrency_nonstreaming_levels {
+        scenarios.push(Box::new(ConcurrencyScenario {
+            id: format!("concurrency:{n}:nostream"),
+            concurrency: n,
+            approx_prompt_tokens: cfg.concurrency_prompt_tokens,
+            streaming: false,
         }) as Box<dyn Scenario>);
     }
     for &side in &cfg.image_sizes {
@@ -257,7 +312,7 @@ pub fn build_scenarios(cfg: &ScenarioConfig) -> Vec<Box<dyn Scenario>> {
 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))
+    tokio::time::timeout(ctx.timeout, stream_and_measure(ctx, &payload, None))
         .await
         .map_err(|_| anyhow!("cold probe timed out after {:?}", ctx.timeout))?
 }
@@ -295,7 +350,7 @@ impl Scenario for ChatLatencyScenario {
     async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics> {
         let prompt = build_prompt(self.approx_prompt_tokens);
         let payload = chat_payload(ctx, &prompt);
-        let fut = stream_and_measure(ctx, &payload);
+        let fut = stream_and_measure(ctx, &payload, None);
         tokio::time::timeout(ctx.timeout, fut)
             .await
             .map_err(|_| anyhow!("request timed out after {:?}", ctx.timeout))?
@@ -313,6 +368,16 @@ pub struct ConcurrencyScenario {
     id: String,
     concurrency: u32,
     approx_prompt_tokens: u32,
+    /// Whether the burst streams (#285).
+    ///
+    /// Both shapes matter and they are not interchangeable. Streaming
+    /// multiplexes through the batch engine; non-streaming did not until
+    /// #285, and serialized completely — 1.00x aggregate throughput at
+    /// every concurrency level against 3.98x streamed. A bench that only
+    /// streams cannot see that, which is why it went unmeasured: the
+    /// neuron reports `in_flight: 8`, admission accepts, `/health` looks
+    /// healthy, and throughput is single-stream.
+    streaming: bool,
 }
 
 #[async_trait]
@@ -327,13 +392,26 @@ impl Scenario for ConcurrencyScenario {
 
     async fn run(&self, ctx: &RunCtx) -> Result<ScenarioMetrics> {
         let prompt = build_prompt(self.approx_prompt_tokens);
-        let payload = chat_payload(ctx, &prompt);
+        let payload = if self.streaming {
+            chat_payload(ctx, &prompt)
+        } else {
+            nonstreaming_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.
+        // Fire all requests at once; each is independently timed and capped
+        // by the per-request timeout so one hung request 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 streaming = self.streaming;
+        let payload = &payload;
+        let futs = (0..self.concurrency).map(move |i| async move {
+            // Each stream presents its own identity — see
+            // `with_stream_identity`.
+            if streaming {
+                tokio::time::timeout(ctx.timeout, stream_and_measure(ctx, payload, Some(i))).await
+            } else {
+                tokio::time::timeout(ctx.timeout, request_and_measure(ctx, payload, Some(i))).await
+            }
         });
         let results = futures::future::join_all(futs).await;
         let burst_window = burst_start.elapsed().as_secs_f64();
@@ -351,7 +429,7 @@ impl Scenario for ConcurrencyScenario {
         }
         if streams.is_empty() {
             return Err(anyhow!(
-                "all {} concurrent streams failed ({rejected} shed by admission)",
+                "all {} concurrent requests failed ({rejected} shed by admission)",
                 self.concurrency
             ));
         }
@@ -380,9 +458,50 @@ impl Scenario for ConcurrencyScenario {
             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,
+            // Median server-side phase timing across the burst. These
+            // were `None`, which left the prefill-tok/s panel blank for
+            // exactly the cells where queue behaviour matters most, and
+            // made `queue_wait_ms` (TTFT minus server prefill, #85)
+            // uncomputable for a burst — the one number that separates
+            // "the server is slow" from "you were queued".
+            //
+            // Median rather than sum: every stream prefills its own copy
+            // of the same prompt, so the representative per-stream cost
+            // is the meaningful figure and it composes with `ttft_s`,
+            // which is also a median.
+            prefill_ms: median(
+                &streams
+                    .iter()
+                    .filter_map(|m| m.prefill_ms.map(|v| v as f64))
+                    .collect::<Vec<_>>(),
+            )
+            .map(|v| v as u64),
+            decode_ms: median(
+                &streams
+                    .iter()
+                    .filter_map(|m| m.decode_ms.map(|v| v as f64))
+                    .collect::<Vec<_>>(),
+            )
+            .map(|v| v as u64),
+            prefill_tokens: median(
+                &streams
+                    .iter()
+                    .filter_map(|m| m.prefill_tokens.map(|v| v as f64))
+                    .collect::<Vec<_>>(),
+            )
+            .map(|v| v as u64),
+            // Summed across the burst, like `completion_tokens`: the cell
+            // describes the whole burst, so a per-stream figure here
+            // would not compose with it.
+            reasoning_tokens: sum_opt(&streams, |m| m.reasoning_tokens),
+            cached_tokens: sum_opt(&streams, |m| m.cached_tokens),
+            // The worst stream in the burst, not a median of medians:
+            // under load the question is how bad it got for somebody,
+            // and averaging tails is how a stall stays invisible.
+            tpot_p95_ms: streams
+                .iter()
+                .filter_map(|m| m.tpot_p95_ms)
+                .max_by(|a, b| a.total_cmp(b)),
             concurrency: Some(self.concurrency),
             ttft_p95_s: percentile(&ttfts, 95.0),
             queue_wait_ms_median: median(&queue_waits),
@@ -393,6 +512,146 @@ impl Scenario for ConcurrencyScenario {
     }
 }
 
+/// Sum an optional per-stream count across a burst, `None` when no
+/// stream reported it — so "nobody measured it" stays distinguishable
+/// from "measured as zero", the same distinction #269 exists to
+/// preserve.
+fn sum_opt(
+    streams: &[ScenarioMetrics],
+    f: impl Fn(&ScenarioMetrics) -> Option<u64>,
+) -> Option<u64> {
+    let vals: Vec<u64> = streams.iter().filter_map(f).collect();
+    (!vals.is_empty()).then(|| vals.iter().sum())
+}
+
+/// Per-stream identity for a concurrency burst (#288/#54).
+///
+/// Fair-share bounds *one principal* to `max_per_principal`
+/// in-flight-or-queued requests, defaulting to 2. A burst of 8 under a
+/// single identity would therefore have 6 rejected with `PrincipalCap`,
+/// and the scenario would measure the fair-share cap rather than the
+/// server.
+///
+/// Giving each stream its own key is also the more faithful model: a
+/// concurrency-8 burst is meant to stand in for eight simultaneous
+/// *callers*, not one caller issuing eight requests. Fair-share is still
+/// exercised — each identity is subject to it — just not self-inflicted.
+fn with_stream_identity(
+    rb: reqwest::RequestBuilder,
+    ctx: &RunCtx<'_>,
+    stream: Option<u32>,
+) -> reqwest::RequestBuilder {
+    match (ctx.principal_key_id.as_deref(), stream) {
+        (Some(base), Some(i)) => rb.header(HEADER_KEY_ID, format!("{base}-{i}")),
+        // No principal configured, or a single-request scenario: the
+        // client's default headers already carry the base identity.
+        _ => rb,
+    }
+}
+
+/// The non-streaming counterpart to [`chat_payload`] (#285).
+///
+/// `stream_options` is deliberately absent: usage is part of the body on
+/// this shape, and sending the streaming-only option to a strict server
+/// is a needless compatibility risk.
+fn nonstreaming_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": false,
+    })
+}
+
+/// Time one non-streaming chat completion (#285).
+///
+/// There is no first-chunk moment on this shape — the whole body arrives
+/// at once — so `ttft_s` is the full request wall-clock rather than a
+/// separate measurement. That is the honest reading: for a non-streaming
+/// caller, time-to-first-token *is* time-to-everything, and it is exactly
+/// what makes the serialization this scenario exists to detect so
+/// expensive.
+///
+/// `decode_tps` is per-request; the burst aggregate is computed by the
+/// caller over the whole window, the same way the streaming path does it.
+async fn request_and_measure(
+    ctx: &RunCtx<'_>,
+    payload: &serde_json::Value,
+    stream: Option<u32>,
+) -> Result<ScenarioMetrics> {
+    let start = Instant::now();
+    let resp = with_stream_identity(ctx.client.post(&ctx.chat_url), ctx, stream)
+        .json(payload)
+        .send()
+        .await
+        .context("sending non-streaming chat request")?;
+    if !resp.status().is_success() {
+        let status = resp.status();
+        let body = resp.text().await.unwrap_or_default();
+        return Err(anyhow!("upstream returned {status}: {}", body.trim()));
+    }
+    let v: serde_json::Value = resp
+        .json()
+        .await
+        .context("non-streaming chat response was not JSON")?;
+    let total_s = start.elapsed().as_secs_f64();
+
+    let usage = v.get("usage");
+    let completion_tokens = usage
+        .and_then(|u| u.get("completion_tokens"))
+        .and_then(|t| t.as_u64())
+        .unwrap_or(0);
+    let prompt_tokens = usage
+        .and_then(|u| u.get("prompt_tokens"))
+        .and_then(|t| t.as_u64());
+    let reasoning_tokens = usage
+        .and_then(|u| u.get("completion_tokens_details"))
+        .and_then(|d| d.get("reasoning_tokens"))
+        .and_then(|t| t.as_u64());
+    let cached_tokens = usage
+        .and_then(|u| u.get("prompt_tokens_details"))
+        .and_then(|d| d.get("cached_tokens"))
+        .and_then(|t| t.as_u64());
+    let timing = usage.and_then(|u| u.get("helexa_timing"));
+    let field = |name: &str| timing.and_then(|t| t.get(name)).and_then(|x| x.as_u64());
+    let (prefill_ms, decode_ms, prefill_tokens) = (
+        field("prefill_ms"),
+        field("decode_ms"),
+        field("prefill_tokens"),
+    );
+
+    // Prefer the server's own decode window when it reports one (#85);
+    // it excludes prefill, so it is comparable with the streaming path's
+    // decode-window rate rather than being diluted by it.
+    let decode_tps = match decode_ms {
+        Some(ms) if ms > 200 => Some(completion_tokens as f64 / (ms as f64 / 1000.0)),
+        _ if total_s > 0.2 => Some(completion_tokens as f64 / total_s),
+        _ => None,
+    };
+
+    Ok(ScenarioMetrics {
+        ttft_s: total_s,
+        decode_tps,
+        total_s,
+        prompt_tokens,
+        completion_tokens,
+        prefill_ms,
+        decode_ms,
+        prefill_tokens,
+        reasoning_tokens,
+        cached_tokens,
+        // Non-streaming has no inter-token gaps: the body arrives whole.
+        tpot_p95_ms: None,
+        concurrency: None,
+        ttft_p95_s: None,
+        queue_wait_ms_median: None,
+        rejected: None,
+        artifact: None,
+        image_units: 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
@@ -425,7 +684,7 @@ impl Scenario for CapabilityScenario {
             "stream": true,
             "stream_options": {"include_usage": true},
         });
-        let fut = stream_and_measure_inner(ctx, &payload, true);
+        let fut = stream_and_measure_inner(ctx, &payload, true, None);
         tokio::time::timeout(ctx.timeout, fut)
             .await
             .map_err(|_| anyhow!("capability probe timed out after {:?}", ctx.timeout))?
@@ -468,8 +727,9 @@ fn percentile(values: &[f64], p: f64) -> Option<f64> {
 async fn stream_and_measure(
     ctx: &RunCtx<'_>,
     payload: &serde_json::Value,
+    stream: Option<u32>,
 ) -> Result<ScenarioMetrics> {
-    stream_and_measure_inner(ctx, payload, false).await
+    stream_and_measure_inner(ctx, payload, false, stream).await
 }
 
 /// As [`stream_and_measure`] but accumulates the full visible text when
@@ -479,11 +739,10 @@ async fn stream_and_measure_inner(
     ctx: &RunCtx<'_>,
     payload: &serde_json::Value,
     capture_text: bool,
+    stream_ix: Option<u32>,
 ) -> Result<ScenarioMetrics> {
     let start = Instant::now();
-    let resp = ctx
-        .client
-        .post(&ctx.chat_url)
+    let resp = with_stream_identity(ctx.client.post(&ctx.chat_url), ctx, stream_ix)
         .json(payload)
         .send()
         .await
@@ -503,6 +762,10 @@ async fn stream_and_measure_inner(
     let mut prefill_ms: Option<u64> = None;
     let mut decode_ms: Option<u64> = None;
     let mut prefill_tokens: Option<u64> = None;
+    let mut reasoning_tokens: Option<u64> = None;
+    let mut cached_tokens: Option<u64> = None;
+    // Inter-token arrival gaps, for the p95 tail.
+    let mut gaps_ms: Vec<f64> = Vec::new();
     let mut captured = String::new();
 
     while let Some(event) = stream.next().await {
@@ -540,6 +803,11 @@ async fn stream_and_measure_inner(
             if content.is_some() || reasoning.is_some() {
                 if first.is_none() {
                     first = Some(now);
+                } else if let Some(prev) = last {
+                    // Gaps only *between* generated deltas: the wait for
+                    // the first one is TTFT and belongs to prefill, not
+                    // to stream smoothness.
+                    gaps_ms.push(now.duration_since(prev).as_secs_f64() * 1000.0);
                 }
                 last = Some(now);
                 chunk_count += 1;
@@ -551,6 +819,14 @@ async fn stream_and_measure_inner(
         if let Some(usage) = chunk.usage {
             prompt_tokens = Some(usage.prompt_tokens);
             completion_tokens = Some(usage.completion_tokens);
+            reasoning_tokens = usage
+                .completion_tokens_details
+                .as_ref()
+                .map(|d| d.reasoning_tokens);
+            cached_tokens = usage
+                .prompt_tokens_details
+                .as_ref()
+                .map(|d| d.cached_tokens);
             if let Some(t) = usage.helexa_timing {
                 prefill_ms = Some(t.prefill_ms);
                 decode_ms = Some(t.decode_ms);
@@ -589,6 +865,9 @@ async fn stream_and_measure_inner(
         prefill_ms,
         decode_ms,
         prefill_tokens,
+        reasoning_tokens,
+        cached_tokens,
+        tpot_p95_ms: percentile(&gaps_ms, 95.0),
         // Concurrency fields unset on the single-request path; the
         // concurrency scenario builds its own aggregate (#89).
         concurrency: None,
@@ -604,6 +883,113 @@ async fn stream_and_measure_inner(
 mod tests {
     use super::*;
 
+    /// A burst must report server-side phase timing, or `queue_wait_ms`
+    /// (TTFT minus server prefill, #85) cannot be computed for the very
+    /// cells it was designed for, and the prefill-tok/s panel stays
+    /// blank under concurrency.
+    #[test]
+    fn a_burst_reports_median_phase_timing_not_none() {
+        let s = |prefill_ms, decode_ms, prefill_tokens| ScenarioMetrics {
+            ttft_s: 0.1,
+            decode_tps: Some(10.0),
+            total_s: 1.0,
+            prompt_tokens: Some(100),
+            completion_tokens: 10,
+            prefill_ms,
+            decode_ms,
+            prefill_tokens,
+            reasoning_tokens: None,
+            cached_tokens: None,
+            tpot_p95_ms: Some(5.0),
+            concurrency: None,
+            ttft_p95_s: None,
+            queue_wait_ms_median: None,
+            rejected: None,
+            artifact: None,
+            image_units: None,
+        };
+        let streams = [
+            s(Some(10), Some(100), Some(512)),
+            s(Some(20), Some(200), Some(512)),
+            s(Some(30), Some(300), Some(512)),
+        ];
+        let med = |f: fn(&ScenarioMetrics) -> Option<u64>| {
+            median(
+                &streams
+                    .iter()
+                    .filter_map(|m| f(m).map(|v| v as f64))
+                    .collect::<Vec<_>>(),
+            )
+            .map(|v| v as u64)
+        };
+        assert_eq!(med(|m| m.prefill_ms), Some(20));
+        assert_eq!(med(|m| m.decode_ms), Some(200));
+        // Every stream prefills the same prompt, so the median is the
+        // prompt size rather than a sum across the burst.
+        assert_eq!(med(|m| m.prefill_tokens), Some(512));
+    }
+
+    /// #288/#54: a burst must present one identity *per stream*.
+    ///
+    /// Fair-share bounds a single principal to `max_per_principal`
+    /// in-flight-or-queued requests, which defaults to 2. Eight streams
+    /// under one key would have six rejected with `PrincipalCap`, and
+    /// `concurrency:8` would silently measure the fair-share cap instead
+    /// of the server — a worse failure than the anonymity it replaced,
+    /// because it looks like backpressure.
+    #[test]
+    fn a_burst_presents_one_identity_per_stream() {
+        let client = reqwest::Client::new();
+        let ctx = RunCtx {
+            client: &client,
+            chat_url: "http://x/v1/chat/completions".into(),
+            model_id: "m".into(),
+            max_tokens: 16,
+            timeout: Duration::from_secs(1),
+            principal_key_id: Some("fleet-benchmark".into()),
+        };
+        let key_of = |stream: Option<u32>| -> Option<String> {
+            with_stream_identity(client.post(&ctx.chat_url), &ctx, stream)
+                .build()
+                .expect("build")
+                .headers()
+                .get(HEADER_KEY_ID)
+                .map(|v| v.to_str().expect("utf8").to_string())
+        };
+        let keys: Vec<Option<String>> = (0..3).map(|i| key_of(Some(i))).collect();
+        assert_eq!(
+            keys,
+            vec![
+                Some("fleet-benchmark-0".to_string()),
+                Some("fleet-benchmark-1".to_string()),
+                Some("fleet-benchmark-2".to_string()),
+            ]
+        );
+        // Distinct, or the cap bites anyway.
+        let uniq: std::collections::HashSet<_> = keys.iter().collect();
+        assert_eq!(uniq.len(), 3);
+    }
+
+    /// A single-request scenario adds no override: the client's default
+    /// headers already carry the base identity, and a per-request copy
+    /// would be a second place for it to drift.
+    #[test]
+    fn a_single_request_scenario_keeps_the_client_identity() {
+        let client = reqwest::Client::new();
+        let ctx = RunCtx {
+            client: &client,
+            chat_url: "http://x/v1/chat/completions".into(),
+            model_id: "m".into(),
+            max_tokens: 16,
+            timeout: Duration::from_secs(1),
+            principal_key_id: Some("fleet-benchmark".into()),
+        };
+        let req = with_stream_identity(client.post(&ctx.chat_url), &ctx, None)
+            .build()
+            .expect("build");
+        assert!(req.headers().get(HEADER_KEY_ID).is_none());
+    }
+
     #[test]
     fn prompt_grows_with_token_target() {
         let small = build_prompt(128);
@@ -646,6 +1032,7 @@ mod tests {
             prompt_sizes: vec![128],
             max_tokens: 64,
             concurrency_levels: vec![2, 8],
+            concurrency_nonstreaming_levels: vec![8],
             concurrency_prompt_tokens: 512,
             capability_probes: vec![CapabilityProbe {
                 name: "plan".into(),

crates/helexa-bench/src/store.rs

@@ -14,6 +14,22 @@ use std::path::Path;
 #[derive(Debug, Clone)]
 pub struct RunRecord {
     pub ts: String, // RFC3339
+    /// Identity this sample was taken under (#288), `account/key`, or
+    /// `None` for an anonymous run. Anonymous samples measure the
+    /// admission yield policy rather than serving capacity once #262 is
+    /// in the build, so this is the column that says whether two rows
+    /// are comparable at all.
+    pub principal: Option<String>,
+    /// Tokens spent reasoning (#223) — the dominant cost driver on a
+    /// reasoning model, and one that moves independently of every tok/s
+    /// number on the dashboard.
+    pub reasoning_tokens: Option<u64>,
+    /// Prompt tokens served from the prefix cache (#269) — the reason
+    /// prefill timing varies between otherwise identical samples.
+    pub cached_tokens: Option<u64>,
+    /// p95 inter-token arrival gap (ms) — stream smoothness, which a
+    /// mean tok/s cannot express.
+    pub tpot_p95_ms: Option<f64>,
     // target
     pub target_name: String,
     pub target_kind: String,
@@ -204,6 +220,18 @@ impl Store {
                 ("artifact", "TEXT"),
                 ("quality_score", "REAL"),
                 ("scorer", "TEXT"),
+                // #288: which identity the sample was taken under. Rows
+                // predating this are NULL, which is honest — they were
+                // anonymous, and since #262 that measures a different
+                // thing. Without the column the 2026-08-18 discontinuity
+                // is folklore rather than data.
+                ("principal", "TEXT"),
+                // #223 / #269: both are on the wire and were being discarded.
+                ("reasoning_tokens", "INTEGER"),
+                ("cached_tokens", "INTEGER"),
+                // #285: inter-token tail. `decode_tps` is a mean and hides
+                // a stall that later catches up.
+                ("tpot_p95_ms", "REAL"),
             ],
         )?;
         Ok(())
@@ -264,6 +292,7 @@ impl Store {
                 swap_unload_ms, swap_load_ms,
                 artifact, quality_score, scorer,
                 image_units,
+                principal, reasoning_tokens, cached_tokens, tpot_p95_ms,
                 ok, error
             ) VALUES (
                 ?1, ?2, ?3, ?4,
@@ -280,7 +309,8 @@ impl Store {
                 ?42, ?43,
                 ?44, ?45, ?46,
                 ?47,
-                ?48, ?49
+                ?48, ?49, ?50, ?51,
+                ?52, ?53
             )",
             params![
                 r.ts,
@@ -330,6 +360,10 @@ impl Store {
                 r.quality_score,
                 r.scorer,
                 r.image_units,
+                r.principal,
+                r.reasoning_tokens,
+                r.cached_tokens,
+                r.tpot_p95_ms,
                 r.ok as i64,
                 r.error,
             ],
@@ -703,7 +737,15 @@ impl Store {
             }
         };
         let mut stmt = self.conn.prepare(
-            "SELECT git_sha, build_timestamp, package_version, ttft_s, decode_tps, total_s, ts
+            "SELECT git_sha, build_timestamp, package_version, ttft_s, decode_tps, total_s,
+                    ttft_p95_s, queue_wait_ms, rejected,
+                    -- Prefill rate derived here rather than stored: the
+                    -- inputs are already recorded and a stored rate
+                    -- would be a third thing to keep consistent.
+                    CASE WHEN prefill_ms > 0 AND prefill_tokens IS NOT NULL
+                         THEN (prefill_tokens * 1000.0) / prefill_ms END AS prefill_tps,
+                    reasoning_tokens, cached_tokens, completion_tokens,
+                    tpot_p95_ms, principal, ts
              FROM runs
              WHERE ok=1 AND target_name=?1 AND model_id=?2 AND scenario_id=?3
              ORDER BY id",
@@ -717,7 +759,16 @@ impl Store {
                     ttft_s: r.get(3)?,
                     decode_tps: r.get(4)?,
                     total_s: r.get(5)?,
-                    ts: r.get(6)?,
+                    ttft_p95_s: r.get(6)?,
+                    queue_wait_ms: r.get(7)?,
+                    rejected: r.get(8)?,
+                    prefill_tps: r.get(9)?,
+                    reasoning_tokens: r.get(10)?,
+                    cached_tokens: r.get(11)?,
+                    completion_tokens: r.get(12)?,
+                    tpot_p95_ms: r.get(13)?,
+                    principal: r.get(14)?,
+                    ts: r.get(15)?,
                 })
             })?
             .collect::<rusqlite::Result<_>>()?;
@@ -838,6 +889,33 @@ pub struct SeriesPoint {
     pub ttft_s_median: Option<f64>,
     pub decode_tps_median: Option<f64>,
     pub total_s_median: Option<f64>,
+    // ── Added because the two above cannot show what changed ──────
+    /// Tail TTFT. The median above stayed flat at ~0.3 s across the
+    /// build where p95 went 0.53 s → 11.77 s (#288); charting only the
+    /// median is why that went unnoticed for a week.
+    pub ttft_p95_s_median: Option<f64>,
+    /// Admission queue wait — separates "the server is slow" from
+    /// "you were queued behind someone".
+    pub queue_wait_ms_median: Option<f64>,
+    /// Requests shed by admission during the burst.
+    pub rejected_median: Option<f64>,
+    /// Prefill throughput, derived from `prefill_tokens / prefill_ms`.
+    /// The other half of serving speed; only decode was ever charted.
+    pub prefill_tps_median: Option<f64>,
+    /// Tokens spent reasoning (#223) — the dominant cost driver on a
+    /// reasoning model, and independent of every speed metric here.
+    pub reasoning_tokens_median: Option<f64>,
+    /// Prompt tokens served from the prefix cache (#269).
+    pub cached_tokens_median: Option<f64>,
+    /// Total generated tokens — catches a model that starts truncating
+    /// or rambling, which no rate metric shows.
+    pub completion_tokens_median: Option<f64>,
+    /// p95 inter-token gap (ms) — stream smoothness.
+    pub tpot_p95_ms_median: Option<f64>,
+    /// Identity the samples were taken under (#288), or `None` if
+    /// anonymous. Anonymous and identified points are **not
+    /// comparable** once #262 is in the build.
+    pub principal: Option<String>,
     pub samples: usize,
 }
 
@@ -848,6 +926,15 @@ struct SeriesRaw {
     ttft_s: Option<f64>,
     decode_tps: Option<f64>,
     total_s: Option<f64>,
+    ttft_p95_s: Option<f64>,
+    queue_wait_ms: Option<f64>,
+    rejected: Option<f64>,
+    prefill_tps: Option<f64>,
+    reasoning_tokens: Option<f64>,
+    cached_tokens: Option<f64>,
+    completion_tokens: Option<f64>,
+    tpot_p95_ms: Option<f64>,
+    principal: Option<String>,
     ts: String,
 }
 
@@ -880,6 +967,19 @@ fn aggregate_series(raws: Vec<SeriesRaw>) -> Vec<SeriesPoint> {
                 ttft_s_median: median(rows.iter().filter_map(|r| r.ttft_s)),
                 decode_tps_median: median(rows.iter().filter_map(|r| r.decode_tps)),
                 total_s_median: median(rows.iter().filter_map(|r| r.total_s)),
+                ttft_p95_s_median: median(rows.iter().filter_map(|r| r.ttft_p95_s)),
+                queue_wait_ms_median: median(rows.iter().filter_map(|r| r.queue_wait_ms)),
+                rejected_median: median(rows.iter().filter_map(|r| r.rejected)),
+                prefill_tps_median: median(rows.iter().filter_map(|r| r.prefill_tps)),
+                reasoning_tokens_median: median(rows.iter().filter_map(|r| r.reasoning_tokens)),
+                cached_tokens_median: median(rows.iter().filter_map(|r| r.cached_tokens)),
+                completion_tokens_median: median(rows.iter().filter_map(|r| r.completion_tokens)),
+                tpot_p95_ms_median: median(rows.iter().filter_map(|r| r.tpot_p95_ms)),
+                // Any row's principal: a build's samples are all taken
+                // under one identity, and a mixture would mean the
+                // config changed mid-build, which the chart should show
+                // as a discontinuity rather than average away.
+                principal: rows.iter().find_map(|r| r.principal.clone()),
                 samples: rows.len(),
             };
             (sort_key, point)
@@ -1281,6 +1381,10 @@ mod tests {
     fn rec(target: &str, sha: &str, model: &str, scenario: &str, ok: bool) -> RunRecord {
         RunRecord {
             ts: "2026-06-13T00:00:00Z".into(),
+            principal: None,
+            reasoning_tokens: None,
+            cached_tokens: None,
+            tpot_p95_ms: None,
             target_name: target.into(),
             target_kind: "neuron".into(),
             endpoint: "http://x:13131".into(),

crates/helexa-bench/src/sweep.rs

@@ -78,7 +78,10 @@ pub struct Sweeper {
 
 impl Sweeper {
     pub fn new(cfg: BenchConfig) -> Result<Self> {
-        let client = TargetClient::new(cfg.bench.request_timeout())?;
+        let client = TargetClient::with_principal(
+            cfg.bench.request_timeout(),
+            cfg.bench.principal.as_ref(),
+        )?;
         let store = Store::open(&cfg.bench.db_path)?;
         Ok(Sweeper { cfg, client, store })
     }
@@ -169,6 +172,7 @@ impl Sweeper {
             model_id: model.id.clone(),
             max_tokens: self.cfg.scenarios.max_tokens,
             timeout: self.cfg.bench.request_timeout(),
+            principal_key_id: self.cfg.bench.principal.as_ref().map(|p| p.key_id.clone()),
         };
         let cold = crate::scenario::cold_probe(&ctx).await;
         let swap = SwapTiming { unload_ms, load_ms };
@@ -252,6 +256,7 @@ impl Sweeper {
                     model_id: model.id.clone(),
                     max_tokens: self.cfg.scenarios.max_tokens,
                     timeout: self.cfg.bench.request_timeout(),
+                    principal_key_id: self.cfg.bench.principal.as_ref().map(|p| p.key_id.clone()),
                 };
 
                 // One unmeasured warmup when the cell is empty (matches
@@ -346,6 +351,15 @@ impl Sweeper {
 
         RunRecord {
             ts: chrono::Utc::now().to_rfc3339(),
+            // #288: stamp the identity this sample was taken under, so a
+            // later reader can tell an anonymous row from an identified
+            // one instead of inferring it from the date.
+            principal: self
+                .cfg
+                .bench
+                .principal
+                .as_ref()
+                .map(|p| format!("{}/{}", p.account_id, p.key_id)),
             target_name: target.name.clone(),
             target_kind: kind_str(target.kind).to_string(),
             endpoint: target.endpoint.clone(),
@@ -383,6 +397,9 @@ impl Sweeper {
             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),
+            reasoning_tokens: m.and_then(|x| x.reasoning_tokens),
+            cached_tokens: m.and_then(|x| x.cached_tokens),
+            tpot_p95_ms: m.and_then(|x| x.tpot_p95_ms),
             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),

crates/helexa-bench/tests/api.rs

@@ -15,6 +15,10 @@ fn rec(
     ok: bool,
 ) -> RunRecord {
     RunRecord {
+        principal: None,
+        reasoning_tokens: None,
+        cached_tokens: None,
+        tpot_p95_ms: None,
         ts: "2026-06-13T00:00:00Z".into(),
         target_name: host.into(),
         target_kind: "neuron".into(),

crates/helexa-bench/tests/sweep_integration.rs

@@ -80,6 +80,7 @@ async fn spawn_mock(sha: &str) -> (String, Arc<Mutex<String>>) {
 fn config_for(endpoint: String, db_path: String) -> BenchConfig {
     BenchConfig {
         bench: BenchSettings {
+            principal: None,
             sweep_interval_secs: 1,
             samples_per_version: 2,
             iteration_pause_secs: 0,
@@ -91,6 +92,7 @@ fn config_for(endpoint: String, db_path: String) -> BenchConfig {
             prompt_sizes: vec![128], // single scenario keeps assertions simple
             max_tokens: 16,
             concurrency_levels: Vec::new(),
+            concurrency_nonstreaming_levels: vec![],
             concurrency_prompt_tokens: 512,
             capability_probes: Vec::new(),
         },

crates/helexa-router/src/catalogue.rs

@@ -80,10 +80,19 @@ fn router_entry(cortex: &str, e: &CortexModelEntry) -> CortexModelEntry {
         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,
+        // Every flat ecosystem field is derived from `limit` by the
+        // final sync pass in aggregate_models, so they are cleared here
+        // rather than copied. Spread from `e` keeps a field added to the
+        // entry later from being silently dropped at this hop.
+        ..CortexModelEntry {
+            max_model_len: None,
+            max_input_tokens: None,
+            max_output_tokens: None,
+            context_window: None,
+            context_length: None,
+            reasoning_budget: Vec::new(),
+            ..e.clone()
+        }
     }
 }
 
@@ -164,6 +173,9 @@ mod tests {
             max_model_len: None,
             max_input_tokens: None,
             max_output_tokens: None,
+            context_window: None,
+            context_length: None,
+            reasoning_budget: Vec::new(),
         }
     }
 
@@ -224,6 +236,7 @@ mod tests {
             context: 32_768,
             input: None,
             output: 4096,
+            output_ceiling: 0,
         });
         a.cost = Some(ModelCost {
             input: 0.50,
@@ -236,6 +249,7 @@ mod tests {
             context: 16_384, // tighter
             input: None,
             output: 4096,
+            output_ceiling: 0,
         });
         b.cost = Some(ModelCost {
             input: 0.20, // cheaper

crates/helexa-router/tests/dispatch.rs

@@ -42,6 +42,9 @@ fn model_entry(loaded: bool, feasible: bool) -> CortexModelEntry {
         max_model_len: None,
         max_input_tokens: None,
         max_output_tokens: None,
+        context_window: None,
+        context_length: None,
+        reasoning_budget: Vec::new(),
     }
 }
 

crates/helexa-tools/src/config.rs

@@ -58,6 +58,11 @@ mod tests {
         assert!(cfg.max_redirects > 0);
     }
 
+    // The closure's error type is `figment::Error`, which is 208 bytes and
+    // not ours to shrink; `Jail::expect_with` dictates the signature. The
+    // usual remedy (box the Err) is not available through a third-party
+    // callback contract, and this is a test closure run once.
+    #[allow(clippy::result_large_err)]
     #[test]
     fn loads_toml_with_env_override() {
         figment::Jail::expect_with(|jail| {

crates/neuron/src/api.rs

@@ -486,9 +486,9 @@ async fn responses(
                     input_tokens: u.prompt_tokens,
                     output_tokens: u.completion_tokens,
                     total_tokens: u.prompt_tokens + u.completion_tokens,
-                    // Carry the reasoning sub-count through from the chat
-                    // usage — the non-streaming path now splits off the
-                    // `<think>` span and counts it (see `split_off_reasoning`).
+                    // Carry the reasoning sub-count through from the
+                    // chat usage: the collector counts the `<think>`
+                    // span and returns its text separately.
                     output_tokens_details: u.completion_tokens_details.as_ref().map(|d| {
                         OutputTokensDetails {
                             reasoning_tokens: d.reasoning_tokens,
@@ -502,8 +502,18 @@ async fn responses(
                     model_id,
                     message_item_id: mint_message_item_id(),
                 };
-                let _ = chat_resp; // make the borrow-checker happy if `text` consumed it
-                let resp = openai_responses::build_response(&meta, text, finish, usage);
+                // The model's think block, which the chat collector now
+                // returns on `message.reasoning_content` (#300). Without
+                // this the caller is billed for `reasoning_tokens` it
+                // never sees, and cannot replay the model's own
+                // reasoning on the next turn.
+                let reasoning = chat_resp
+                    .choices
+                    .first()
+                    .and_then(|c| c.message.extra.get("reasoning_content"))
+                    .and_then(|v| v.as_str())
+                    .map(str::to_string);
+                let resp = openai_responses::build_response(&meta, text, finish, usage, reasoning);
                 Json(resp).into_response()
             }
             Err(e) => inference_error_response(e),

crates/neuron/src/config.rs

@@ -86,6 +86,16 @@ pub struct CandleHarnessConfig {
     #[serde(default)]
     pub context_limit: ContextLimitConfig,
 
+    /// What each OpenAI effort level costs, in reasoning tokens (#223).
+    ///
+    /// OpenAI-shaped clients cannot send a token count — their whole
+    /// vocabulary is the `minimal|low|medium|high` ladder — so the
+    /// numbers live here and are advertised on `/models`. Operator
+    /// configurable because the right values depend on the model and the
+    /// hardware.
+    #[serde(default)]
+    pub reasoning_budget: crate::harness::reasoning_budget::ReasoningBudgetLadder,
+
     /// 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.
@@ -333,8 +343,24 @@ pub struct ContextLimitConfig {
     /// Generation reserve (tokens) left below the context wall:
     /// `input = context − output_reserve_tokens`. Defaults to neuron's
     /// default `max_tokens`.
+    ///
+    /// This is the budget a request gets when it names none, and the
+    /// figure the input/output split is derived from. It is **not** the
+    /// most a request may ask for — see `max_output_tokens` (#278).
     #[serde(default = "default_output_reserve_tokens")]
     pub output_reserve_tokens: usize,
+
+    /// The largest output a request may name, advertised as
+    /// `max_output_tokens` on `/v1/models` (#278).
+    ///
+    /// `None` derives it as `min(max_position_embeddings,
+    /// DEFAULT_OUTPUT_CEILING)`. Separate from the reserve because the
+    /// two answer different questions and were previously the same
+    /// number: publishing the reserve told every client that read the
+    /// advertisement to cap itself at the default, which on a reasoning
+    /// model can be less than one think block.
+    #[serde(default)]
+    pub max_output_tokens: Option<usize>,
 }
 
 impl Default for ContextLimitConfig {
@@ -346,6 +372,7 @@ impl Default for ContextLimitConfig {
             activation_headroom_mb: default_activation_headroom_mb(),
             min_free_floor_mb: default_context_min_free_floor_mb(),
             output_reserve_tokens: default_output_reserve_tokens(),
+            max_output_tokens: None,
         }
     }
 }

crates/neuron/src/harness/admission.rs

@@ -1051,7 +1051,7 @@ mod tests {
 
         // 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)) });
+        let waiter = tokio::spawn(async move { ctrl2.enter(None).await.map(drop) });
         // Give the waiter a moment to occupy the queue slot.
         tokio::time::sleep(Duration::from_millis(50)).await;
         assert_eq!(ctrl.queue_depth(), 1);

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

@@ -380,7 +380,23 @@ impl GatedDeltaNet {
             self.head_v_dim,
         )?;
         // Stash the updated recurrent state for the next call.
-        self.state.recurrent_state = Some(new_state.to_dtype(dtype)?);
+        //
+        // The delta rule ran in f32 and `state_init` upcasts back to f32
+        // on entry, so storing at the model dtype is a round trip that
+        // buys only memory (#284). Decode takes it once per token.
+        //
+        // Storing f32 keeps `new_state` as-is, which on the CUDA decode
+        // path is the same buffer the kernel updated in place — so the
+        // stored state aliases it, where the `to_dtype` branch always
+        // allocated afresh. That is sound: the pre-update value is dead
+        // by this point, and prefix snapshots deep-copy
+        // (`snapshot_state`) precisely so they never share storage with
+        // a buffer a later forward will overwrite.
+        self.state.recurrent_state = Some(if gdn_state_f32() {
+            new_state
+        } else {
+            new_state.to_dtype(dtype)?
+        });
 
         // core_attn_out: (B, H, L, D_v) → (B, L, H, D_v) → (B*L*H, D_v).
         let core_attn_out = core_attn_out.transpose(1, 2)?.contiguous()?; // (B, L, H, D_v)
@@ -464,6 +480,33 @@ fn chunked_prefill_enabled() -> bool {
     })
 }
 
+/// `NEURON_GDN_STATE_F32=1` keeps the recurrent state in f32 between
+/// forwards instead of storing it at the model's activation dtype (#284).
+///
+/// The delta rule computes in f32 throughout, so storing the result at
+/// BF16 and upcasting it again next call is a pure round-trip loss.
+/// Measured on the chunked path: bit-exact carrying f32 at every chunk
+/// size, against 1.3e-3 relative error after 16 prefill chunks and
+/// **1.0e-2 after 4096 decode steps** — decode round-trips once per
+/// token, across 48 of this arch's 64 layers.
+///
+/// Off by default on purpose. Every #283 measurement was taken against
+/// the bf16 baseline, and flipping this silently would move the ground
+/// underneath the sampling A/B and the decode-rate work (#282, #285)
+/// at the same time. Turn it on deliberately, measure, then decide the
+/// default.
+///
+/// Costs ~2× the recurrent slab: +37.7 MB per GPU per sequence on the
+/// deployed 27B under TP=2, so +302 MB/GPU at `max_in_flight = 8`.
+fn gdn_state_f32() -> bool {
+    static ENABLED: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+    *ENABLED.get_or_init(|| {
+        std::env::var("NEURON_GDN_STATE_F32")
+            .map(|v| v == "1" || v.eq_ignore_ascii_case("true"))
+            .unwrap_or(false)
+    })
+}
+
 /// Chunk-parallel gated delta rule — a faithful port of the HF
 /// reference `torch_chunk_gated_delta_rule` (chunk_size = 64) in
 /// `transformers/models/qwen3_5/modeling_qwen3_5.py`, minus the steps
@@ -1097,6 +1140,94 @@ mod tests {
         assert!(d_state < 2e-4, "final state diverged: {d_state}");
     }
 
+    /// #284: the state carried *between* forwards must not lose
+    /// precision. Every other parity test here is single-call and
+    /// all-f32, so none of them crosses the boundary where
+    /// `GatedDeltaNet::forward` stores the state at the model's
+    /// activation dtype — which is why the defect lived.
+    ///
+    /// Two properties, one test:
+    ///
+    /// 1. carrying f32, an N-chunk run equals a single call **exactly**,
+    ///    so the chunk-boundary math itself is sound (this kills the
+    ///    "chunked prefill corrupts state" hypothesis, #283 H1);
+    /// 2. carrying the model dtype loses a measurable, chunk-count-
+    ///    dependent amount — the defect the `NEURON_GDN_STATE_F32` flag
+    ///    exists to switch off.
+    ///
+    /// Asserting (2) is deliberate: if a future change makes the bf16
+    /// carry lossless this test fails loudly and the flag can go, rather
+    /// than the flag quietly outliving its reason.
+    #[test]
+    fn state_carried_across_chunks_is_exact_in_f32_and_lossy_in_bf16() {
+        let (b, h, dk, dv) = (1, 2, 32, 32);
+        let dev = Device::Cpu;
+        let l = 1024;
+        // Long-memory decay: g near zero keeps the state alive across
+        // every boundary, which is both the realistic regime for
+        // long-context retrieval and the one that accumulates most.
+        let scale = 1.0 / (dk as f64).sqrt();
+        let q = Tensor::randn(0f32, 1.0, (b, h, l, dk), &dev).unwrap();
+        let q = (l2norm(&q, 1e-6).unwrap() * scale).unwrap();
+        let k = Tensor::randn(0f32, 1.0, (b, h, l, dk), &dev).unwrap();
+        let k = l2norm(&k, 1e-6).unwrap();
+        let v = (Tensor::randn(0f32, 1.0, (b, h, l, dv), &dev).unwrap() * 0.5).unwrap();
+        let g = (Tensor::rand(0f32, 1f32, (b, h, l), &dev).unwrap() * -0.02).unwrap();
+        let beta = Tensor::rand(0.05f32, 0.95f32, (b, h, l), &dev).unwrap();
+        let zeros = || Tensor::zeros((b, h, dk, dv), DType::F32, &dev).unwrap();
+
+        // Ground truth: the whole sequence in one call, f32 throughout.
+        let (out_ref, _) = run_chunk_gated_delta_rule(&q, &k, &v, &g, &beta, zeros()).unwrap();
+
+        // Replay it in `chunk`-sized pieces, carrying the state the way
+        // `forward` does.
+        let replay = |chunk: usize, carry: Option<DType>| {
+            let mut state = zeros();
+            let mut outs = Vec::new();
+            let mut pos = 0;
+            while pos < l {
+                let n = chunk.min(l - pos);
+                let take = |t: &Tensor| t.narrow(2, pos, n).unwrap().contiguous().unwrap();
+                let (o, s) = run_chunk_gated_delta_rule(
+                    &take(&q),
+                    &take(&k),
+                    &take(&v),
+                    &take(&g),
+                    &take(&beta),
+                    state,
+                )
+                .unwrap();
+                state = match carry {
+                    Some(dt) => s.to_dtype(dt).unwrap().to_dtype(DType::F32).unwrap(),
+                    None => s,
+                };
+                outs.push(o);
+                pos += n;
+            }
+            Tensor::cat(&outs, 2).unwrap()
+        };
+
+        for chunk in [128usize, 256, 512] {
+            let exact = max_abs_diff(&out_ref, &replay(chunk, None));
+            assert_eq!(
+                exact, 0.0,
+                "chunk={chunk}: an f32 state carry must reproduce the \
+                 single-call result bit for bit, got {exact:e}"
+            );
+        }
+
+        // And the dtype round trip is what costs us. Compared on the
+        // final chunk, where the accumulated state has done the most work.
+        let tail = |t: &Tensor| t.narrow(2, l - 128, 128).unwrap().contiguous().unwrap();
+        let lossy = max_abs_diff(&tail(&out_ref), &tail(&replay(128, Some(DType::BF16))));
+        assert!(
+            lossy > 1e-5,
+            "storing the state at BF16 between chunks is expected to lose \
+             precision (#284); if this now passes losslessly the flag can \
+             be retired, got {lossy:e}"
+        );
+    }
+
     /// Adversarially correlated inputs: near-identical keys with
     /// beta ≈ 1 and negligible decay make the UT-transform matrix T
     /// maximally coherent — raw powers of T grow combinatorially

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

@@ -562,9 +562,9 @@ mod tests {
                 .forward_batch_decode(&input, &positions, mask.as_ref())
                 .expect("batched step");
             assert_eq!(h.dims()[0], 3);
-            for row in 0..3 {
+            for (row, exp) in expected.iter().enumerate() {
                 let got: Vec<f32> = h.i((row, 0, ..)).unwrap().to_vec1().unwrap();
-                let diff = max_abs_diff(&expected[row][t], &got);
+                let diff = max_abs_diff(&exp[t], &got);
                 assert!(diff < 1e-4, "row {row} step {t} diverged: {diff}");
             }
         }
@@ -610,9 +610,9 @@ mod tests {
         let h = model
             .forward_batch_decode(&input, &[3, 3], None)
             .expect("step");
-        for row in 0..2 {
+        for (row, exp) in expected.iter().enumerate() {
             let got: Vec<f32> = h.i((row, 0, ..)).unwrap().to_vec1().unwrap();
-            let diff = max_abs_diff(&expected[row], &got);
+            let diff = max_abs_diff(exp, &got);
             assert!(diff < 1e-4, "row {row} diverged: {diff}");
         }
     }
@@ -674,9 +674,9 @@ mod tests {
             let h = model
                 .forward_batch_decode(&input, &positions, mask.as_ref())
                 .expect("batched step");
-            for row in 0..2 {
+            for (row, exp) in expected.iter().enumerate() {
                 let got: Vec<f32> = h.i((row, 0, ..)).unwrap().to_vec1().unwrap();
-                let diff = max_abs_diff(&expected[row][t], &got);
+                let diff = max_abs_diff(&exp[t], &got);
                 assert!(diff < 1e-4, "batched row {row} step {t}: {diff}");
             }
         }

crates/neuron/src/harness/chat_template.rs

@@ -233,7 +233,24 @@ pub fn render_chat_template(
                 MessageContent::Null => Value::Null,
             };
             let mut obj = serde_json::Map::new();
-            obj.insert("role".into(), Value::String(m.role.clone()));
+            // `developer` is OpenAI's successor to `system` and is what
+            // a client sends once a model is declared reasoning-capable
+            // (pi-ai: `model.reasoning && supportsDeveloperRole`). Most
+            // chat templates — every one on this fleet — only know the
+            // older spelling and `raise_exception` on anything else, so
+            // an unmodified client got a 422 for a system prompt it sent
+            // correctly.
+            //
+            // Rewritten only when the template does not handle the role
+            // itself: a template that knows `developer` may treat it as
+            // distinct from `system`, and substituting there would be us
+            // deciding something the model's authors already decided.
+            let role = if m.role == "developer" && !template.contains("developer") {
+                "system"
+            } else {
+                m.role.as_str()
+            };
+            obj.insert("role".into(), Value::String(role.to_string()));
             obj.insert("content".into(), content_value);
             // Forward extras (e.g. tool_calls on assistant turns,
             // tool_call_id on tool result turns). HF templates that
@@ -421,6 +438,139 @@ mod tests {
         assert!(out.contains("<|vision_start|>") && out.contains("<|vision_end|>"));
     }
 
+    /// A file read back through a tool result must reach the model
+    /// byte-for-byte.
+    ///
+    /// The reported symptom: the model writes a file correctly, reads it
+    /// back, and sees errors that are not in the file — consistently,
+    /// after every write. Anything that mangles `<`, `&`, `"` or
+    /// newlines on this path would produce exactly that, and would make
+    /// the model burn its budget reconciling a file against a corrupted
+    /// view of itself.
+    #[test]
+    fn a_file_read_back_through_a_tool_result_is_byte_exact() {
+        let template = include_str!("../../tests/data/qwen3.8-27b.chat_template.jinja");
+        // The characters most likely to be mangled, in the shapes a real
+        // web project contains them.
+        let file = "<!DOCTYPE html>\n<html lang=\"en\">\n<head>\n  <meta charset=\"utf-8\">\n</head>\n<body>\n  <script>\n    if (a && b || c < d) { g.fillStyle = '#b8905a'; }\n    const s = `rgba(${r},${g},0.5)`;\n\n    // a comment & an ampersand\n  </script>\n</body>\n</html>\n";
+
+        let mut tool_extras = serde_json::Map::new();
+        tool_extras.insert("tool_call_id".into(), json!("call_0"));
+        let mut assistant_extras = serde_json::Map::new();
+        assistant_extras.insert(
+            "tool_calls".into(),
+            json!([{"id":"call_0","type":"function",
+                    "function":{"name":"read","arguments":"{\"path\":\"/tmp/index.html\"}"}}]),
+        );
+
+        let messages = vec![
+            user_msg("read the file back"),
+            ChatMessage {
+                role: "assistant".into(),
+                content: MessageContent::Null,
+                extra: Value::Object(assistant_extras),
+            },
+            ChatMessage {
+                role: "tool".into(),
+                content: MessageContent::Text(file.to_string()),
+                extra: Value::Object(tool_extras),
+            },
+            user_msg("is it correct?"),
+        ];
+        let out =
+            render_chat_template(template, &messages, &Value::Null, &Value::Null).expect("render");
+
+        for probe in [
+            "<!DOCTYPE html>",
+            "<html lang=\"en\">",
+            "if (a && b || c < d)",
+            "g.fillStyle = '#b8905a';",
+            "`rgba(${r},${g},0.5)`",
+            "// a comment & an ampersand",
+        ] {
+            assert!(
+                out.contains(probe),
+                "the model would not see {probe:?} verbatim — rendered prompt mangles it"
+            );
+        }
+        // And nothing HTML-escaped anywhere.
+        for bad in ["&lt;", "&gt;", "&amp;", "&quot;", "&#x27;", "&#39;"] {
+            assert!(
+                !out.contains(bad),
+                "found {bad:?} in the rendered prompt: content is being HTML-escaped"
+            );
+        }
+    }
+
+    /// Blank lines inside a replayed tool-call argument must survive
+    /// rendering intact.
+    ///
+    /// A model that writes a source file and then re-reads it from its
+    /// own replayed tool call is comparing what it remembers writing
+    /// against what the prompt now shows. Any whitespace the pipeline
+    /// eats makes those disagree, and the model has no way to tell a
+    /// rendering artifact from a mistake of its own — which is a
+    /// plausible route into the announce-then-stop stall (#296).
+    ///
+    /// The JSON string→object normalisation above is the step that
+    /// could plausibly lose them, since it round-trips the arguments
+    /// through a parser.
+    #[test]
+    fn a_replayed_tool_call_keeps_the_blank_lines_in_its_arguments() {
+        let template = include_str!("../../tests/data/qwen3.8-27b.chat_template.jinja");
+        // Blank lines between sections, exactly as a written source file
+        // carries them.
+        let file_body = "'use strict';\n\n// ---- constants ----\nconst A = 1;\n\n\nfunction f() {\n  return A;\n}\n";
+        let args = json!({ "path": "/tmp/game.js", "content": file_body });
+        let mut assistant_extras = serde_json::Map::new();
+        assistant_extras.insert(
+            "tool_calls".into(),
+            json!([{
+                "id": "call_0", "type": "function",
+                "function": {
+                    "name": "write",
+                    // The OpenAI wire form: arguments as a JSON *string*.
+                    "arguments": serde_json::to_string(&args).unwrap()
+                }
+            }]),
+        );
+        let mut tool_extras = serde_json::Map::new();
+        tool_extras.insert("tool_call_id".into(), json!("call_0"));
+
+        let messages = vec![
+            user_msg("write the file"),
+            ChatMessage {
+                role: "assistant".into(),
+                content: MessageContent::Null,
+                extra: Value::Object(assistant_extras),
+            },
+            ChatMessage {
+                role: "tool".into(),
+                content: MessageContent::Text("Successfully wrote 84 bytes".into()),
+                extra: Value::Object(tool_extras),
+            },
+            user_msg("now check it parses"),
+        ];
+        let out =
+            render_chat_template(template, &messages, &Value::Null, &Value::Null).expect("render");
+
+        assert!(
+            out.contains(file_body),
+            "the file body must appear verbatim in the rendered prompt"
+        );
+        let blanks_in = file_body.lines().filter(|l| l.is_empty()).count();
+        assert_eq!(blanks_in, 3, "fixture sanity");
+        // And specifically the double newline the model would notice.
+        assert!(
+            out.contains("'use strict';\n\n// ---- constants ----"),
+            "blank line after the prologue was eaten"
+        );
+        assert!(
+            out.contains("const A = 1;\n\n\nfunction f()"),
+            "consecutive blank lines were collapsed"
+        );
+    }
+
     /// The second turn of an OpenAI-native agentic loop: the client
     /// replays its own assistant turn, which carries `tool_calls` and
     /// `"content": null`. That must render as the tool-call block —
@@ -498,6 +648,78 @@ mod tests {
     /// Minimal Qwen3-style template — enough surface to confirm
     /// our renderer threads role + content correctly without
     /// loading a real model's tokenizer_config.json.
+    /// A reasoning-capable model makes OpenAI clients send the system
+    /// prompt as `developer` — and every chat template on this fleet
+    /// raises on an unknown role, so the caller got a 422 for a prompt
+    /// it had sent correctly. Observed live: pi with `reasoning: true`,
+    /// `template_render_failed: Unexpected message role`.
+    #[test]
+    fn the_developer_role_reaches_a_template_that_only_knows_system() {
+        const SYSTEM_ONLY: &str = "{%- for message in messages -%}\
+{%- if message.role == 'system' -%}SYS:{{ message.content }}\
+{%- elif message.role == 'user' -%}USR:{{ message.content }}\
+{%- else -%}{{ raise_exception('Unexpected message role.') }}\
+{%- endif -%}\
+{%- endfor -%}";
+        let messages = vec![
+            ChatMessage {
+                role: "developer".into(),
+                content: MessageContent::Text("be terse".into()),
+                extra: Value::Object(Default::default()),
+            },
+            user_msg("hi"),
+        ];
+        let out = render_chat_template(SYSTEM_ONLY, &messages, &Value::Null, &Value::Null)
+            .expect("a developer role must not blow up a system-only template");
+        assert!(out.contains("SYS:be terse"), "got {out:?}");
+    }
+
+    /// A template that models `developer` itself keeps it: the role may
+    /// mean something distinct there, and substituting would override a
+    /// decision the model's authors already made.
+    #[test]
+    fn a_template_that_knows_developer_keeps_it() {
+        const KNOWS_DEVELOPER: &str = "{%- for message in messages -%}\
+{%- if message.role == 'developer' -%}DEV:{{ message.content }}\
+{%- else -%}{{ message.role }}:{{ message.content }}\
+{%- endif -%}\
+{%- endfor -%}";
+        let messages = vec![ChatMessage {
+            role: "developer".into(),
+            content: MessageContent::Text("be terse".into()),
+            extra: Value::Object(Default::default()),
+        }];
+        let out = render_chat_template(KNOWS_DEVELOPER, &messages, &Value::Null, &Value::Null)
+            .expect("render");
+        assert!(out.contains("DEV:be terse"), "got {out:?}");
+    }
+
+    /// The fetcher and the reader must agree on the filenames, or the
+    /// reader looks for something nobody downloaded and falls back to a
+    /// built-in prompt format without anything failing (#280 — observed
+    /// live: every model neuron downloaded itself was served with the
+    /// fallback because no sidecar was ever fetched).
+    #[test]
+    fn every_template_source_this_reads_is_one_the_loader_fetches() {
+        for name in crate::harness::candle::CandleHarness::SIDECAR_FILES {
+            assert!(
+                !name.is_empty(),
+                "sidecar names are the contract between fetch and read"
+            );
+        }
+        // The three the reader consults, in its own priority order.
+        for source in [
+            "chat_template.jinja",
+            "chat_template.json",
+            "tokenizer_config.json",
+        ] {
+            assert!(
+                crate::harness::candle::CandleHarness::SIDECAR_FILES.contains(&source),
+                "{source} is read by load_chat_template_alongside but never fetched"
+            );
+        }
+    }
+
     const QWEN3_LIKE: &str = "{%- for message in messages -%}\
 <|im_start|>{{ message.role }}\n{{ message.content }}<|im_end|>\n\
 {%- endfor -%}\

crates/neuron/src/harness/context_limit.rs

@@ -234,10 +234,23 @@ fn round_down(tokens: usize, granularity: usize) -> usize {
 
 const CONTEXT_GRANULARITY: usize = 1024;
 
+/// Default ceiling on what a request may name as its output budget.
+///
+/// Not derived from anything — a deliberate figure, and configurable
+/// per host (`[context_limit] max_output_tokens`). The evidence behind
+/// it: a reasoning model asked to build a small application front-loads
+/// one long think block before it acts, measured at 27,504 tokens for a
+/// frontier model on a task Qwen3.8-27B was cut off attempting at
+/// 24,433 (#271, #278). A ceiling below that turns a capable model into
+/// one that never reaches its first tool call, so the default sits
+/// above the observed cost of that class of turn.
+const DEFAULT_OUTPUT_CEILING: usize = 32_768;
+
 /// Derive `limit{context,input,output}` for a loaded model.
 ///
 /// ```text
-/// output  = output_reserve_tokens
+/// output         = output_reserve_tokens          (default + KV planning)
+/// output_ceiling = min(max_output_tokens, max_position_embeddings)
 /// vram_ceiling       = (free_tightest − activation_headroom − min_free_floor) / kv_bytes_per_token_per_card
 /// throughput_ceiling = target_prefill_latency_secs × prefill_tok_per_sec
 /// context = min(max_position_embeddings, vram_ceiling, throughput_ceiling) [clamped by `hard_ceiling` if set]
@@ -318,10 +331,26 @@ pub fn derive_limit(
     );
 
     let input = context.saturating_sub(output);
+    // The ceiling a request may name, distinct from the reserve above
+    // (#278). Bounded by the model's own position limit rather than by
+    // the live context, because the advertisement is read once and
+    // cached by clients: a ceiling that shrank with fleet VRAM would be
+    // stale in whichever direction hurt — either rejecting what we said
+    // we would serve, or leaving budget unclaimed. A request that will
+    // not actually fit is refused at request time, with the numbers,
+    // instead of being under-promised here.
+    let output_ceiling = round_down(
+        cfg.max_output_tokens
+            .unwrap_or(DEFAULT_OUTPUT_CEILING)
+            .min(profile.max_position_embeddings),
+        CONTEXT_GRANULARITY,
+    )
+    .max(output);
     ModelLimit {
         context,
         input: Some(input),
         output,
+        output_ceiling,
     }
 }
 
@@ -329,6 +358,82 @@ pub fn derive_limit(
 mod tests {
     use super::*;
 
+    /// The ceiling a client is told to plan against must not be the
+    /// reserve. Advertising the reserve is what told a reasoning model's
+    /// harness to cap itself below the cost of one think block (#278).
+    #[test]
+    fn the_advertised_ceiling_is_not_the_reserve() {
+        let cfg = ContextLimitConfig::default();
+        let profile = ContextProfile {
+            max_position_embeddings: 131_072,
+            kv_bytes_per_token_per_card: 65_536,
+            world_size: 2,
+        };
+        let limit = derive_limit(&profile, 0, 0.0, None, &cfg);
+        assert_eq!(limit.output, cfg.output_reserve_tokens, "reserve unchanged");
+        assert_eq!(limit.output_ceiling, DEFAULT_OUTPUT_CEILING);
+        assert!(limit.output_ceiling > limit.output);
+    }
+
+    /// A model whose own position limit is below the default ceiling
+    /// cannot generate past it, so the advertisement must not claim it.
+    #[test]
+    fn the_ceiling_never_exceeds_the_models_own_position_limit() {
+        let cfg = ContextLimitConfig::default();
+        let profile = ContextProfile {
+            max_position_embeddings: 16_384,
+            kv_bytes_per_token_per_card: 65_536,
+            world_size: 1,
+        };
+        let limit = derive_limit(&profile, 0, 0.0, None, &cfg);
+        assert_eq!(limit.output_ceiling, 16_384);
+    }
+
+    /// The operator's number wins, which is what makes the default a
+    /// starting point rather than a guess baked into the binary.
+    #[test]
+    fn a_configured_ceiling_overrides_the_default() {
+        let cfg = ContextLimitConfig {
+            max_output_tokens: Some(65_536),
+            ..ContextLimitConfig::default()
+        };
+        let profile = ContextProfile {
+            max_position_embeddings: 131_072,
+            kv_bytes_per_token_per_card: 65_536,
+            world_size: 2,
+        };
+        assert_eq!(
+            derive_limit(&profile, 0, 0.0, None, &cfg).output_ceiling,
+            65_536
+        );
+    }
+
+    /// The ceiling is deliberately *not* a function of live VRAM: it is
+    /// read once and cached by clients, so it has to mean the same thing
+    /// tomorrow. The live context still moves — that is what a request
+    /// is judged against — but the contract does not.
+    #[test]
+    fn the_ceiling_holds_still_while_the_context_moves() {
+        let cfg = ContextLimitConfig::default();
+        let profile = ContextProfile {
+            max_position_embeddings: 131_072,
+            kv_bytes_per_token_per_card: 65_536,
+            world_size: 2,
+        };
+        let roomy = derive_limit(&profile, 60_000, 5_000.0, None, &cfg);
+        // Tight enough that the VRAM term binds — the same shape as the
+        // fleet advertising 109568 one hour and 131072 the next.
+        let tight = derive_limit(&profile, 6_000, 5_000.0, None, &cfg);
+        assert!(
+            tight.context < roomy.context,
+            "the live window should shrink under VRAM pressure"
+        );
+        assert_eq!(
+            tight.output_ceiling, roomy.output_ceiling,
+            "the advertised ceiling must not move with fleet load"
+        );
+    }
+
     /// beast Qwen3.6-27B: 16 full-attn layers, 4 kv heads, head_dim 256,
     /// f16 (2 B), TP=2 → 64 KiB/token total, 32 KiB/token/card.
     fn beast_profile() -> ContextProfile {

crates/neuron/src/harness/engine.rs

@@ -410,6 +410,10 @@ pub struct EngineRequest {
     /// → fallback (#272), rather than loose scalars each caller could
     /// default differently.
     pub sampling: crate::harness::sampling::SamplingParams,
+    /// Reasoning-token budget for this request (#223). `None` is
+    /// unbounded — the default, and the behaviour before it existed.
+    /// Per-request because slots in one batch can differ.
+    pub reasoning_budget: Option<usize>,
     pub eos_id: Option<u32>,
     pub tool_schemas: ToolSchemas,
     pub tx: mpsc::Sender<InferenceEvent>,
@@ -467,6 +471,12 @@ struct Slot {
     max_new: usize,
     eos_id: Option<u32>,
     lp: LogitsProcessor,
+    /// The slot's resolved sampling (#272). Held per-slot because the
+    /// repetition penalty and its window are per-request knobs, and
+    /// slots in one batch can carry different ones.
+    sampling: crate::harness::sampling::SamplingParams,
+    /// Bounds this slot's think block (#223); `None` = unbounded.
+    governor: Option<crate::harness::reasoning_budget::ReasoningGovernor>,
     router: mpsc::Sender<RouterMsg>,
     /// Set by the router when the consumer hangs up; the engine stops
     /// feeding the slot and compacts it out.
@@ -574,7 +584,13 @@ async fn run_engine(cfg: EngineConfig, mut rx: mpsc::Receiver<EngineRequest>) {
                     break;
                 }
             };
-            let nt = match sample_with_penalty(&logits, &slot.generated, &mut slot.lp) {
+            let nt = match sample_with_penalty(
+                &logits,
+                &slot.generated,
+                &mut slot.lp,
+                &slot.sampling,
+                slot.governor.as_mut(),
+            ) {
                 Ok(t) => t,
                 Err(e) => {
                     let health = logits_health_slice(&logits_vec);
@@ -779,6 +795,7 @@ async fn prefill_join(
         prompt_tokens,
         max_new,
         sampling,
+        reasoning_budget,
         eos_id,
         tool_schemas,
         tx,
@@ -787,6 +804,15 @@ async fn prefill_join(
     } = req;
 
     let mut lp = LogitsProcessor::from_sampling(sampling.seed, sampling.to_sampling());
+    // Built before prefill's first sample so a prompt that opens the
+    // think block is governed from its very first token (#223).
+    let mut governor = crate::harness::reasoning_budget::ReasoningGovernor::new(
+        reasoning_budget,
+        cfg.reasoning_tokens
+            .as_ref()
+            .map(|p| (p.open_id, p.close_id)),
+        prompt_opens_reasoning(&prompt_tokens, cfg.reasoning_tokens.as_ref()),
+    );
 
     let prompt_len = prompt_tokens.len();
     let prefill_start = std::time::Instant::now();
@@ -796,17 +822,18 @@ async fn prefill_join(
     // First token from the prefill logits.
     let generated: Vec<u32> = Vec::new();
     let logits = Tensor::new(logits_vec.as_slice(), &Device::Cpu)?;
-    let first = match sample_with_penalty(&logits, &generated, &mut lp) {
-        Ok(t) => t,
-        Err(e) => {
-            let health = logits_health_slice(&logits_vec);
-            tracing::warn!(
-                ?health,
-                "batch engine: prefill sample failed; logits unhealthy"
-            );
-            return Err(e);
-        }
-    };
+    let first =
+        match sample_with_penalty(&logits, &generated, &mut lp, &sampling, governor.as_mut()) {
+            Ok(t) => t,
+            Err(e) => {
+                let health = logits_health_slice(&logits_vec);
+                tracing::warn!(
+                    ?health,
+                    "batch engine: prefill sample failed; logits unhealthy"
+                );
+                return Err(e);
+            }
+        };
 
     // Router task for this slot.
     let hangup = Arc::new(AtomicBool::new(false));
@@ -835,6 +862,8 @@ async fn prefill_join(
         max_new,
         eos_id,
         lp,
+        sampling,
+        governor,
         router: router_tx,
         hangup,
         finished: None,
@@ -1186,6 +1215,9 @@ mod tests {
                     &crate::harness::sampling::ModelGenerationDefaults::default(),
                     0,
                 ),
+                // Unbounded: this fixture asserts exact token sequences,
+                // so nothing may substitute a close marker (#223).
+                reasoning_budget: None,
                 eos_id: None,
                 tool_schemas: ToolSchemas::new(),
                 tx,

crates/neuron/src/harness/mod.rs

@@ -12,6 +12,8 @@ pub mod image;
 pub mod prefix_cache;
 pub mod preflight;
 pub mod preprocess;
+pub mod reasoning_budget;
+pub mod reasoning_effort;
 pub mod sampling;
 pub mod speculative;
 pub mod tp;

crates/neuron/src/harness/preflight.rs

@@ -401,6 +401,8 @@ mod tests {
             quant: quant.map(String::from),
             tensor_parallel: tp,
             devices: None,
+            sampling: None,
+            preserve_thinking: None,
         }
     }
 

crates/neuron/src/harness/reasoning_budget.rs

@@ -0,0 +1,305 @@
+//! Bounding how long a model may think before it must answer (#223).
+//!
+//! A reasoning model with no stopping criterion will fill whatever
+//! output budget it is given. Measured on Qwen3.8-27B: given 16,384
+//! tokens it used 16,384; given 32,768 it used 32,768; both times it was
+//! still opening new topics when the cap bit, and both times the caller
+//! received no answer at all.
+//!
+//! The fix is not to stop the model thinking — the run that did produce
+//! files designed for 27,157 tokens first, and that design was good. It
+//! is to bound the thinking so the rest of the output budget survives
+//! for the answer.
+//!
+//! Two halves live here:
+//!
+//! - [`ReasoningBudget`] — how many reasoning tokens this request may
+//!   spend, resolved from whichever control the caller's wire format
+//!   gave them.
+//! - [`ReasoningGovernor`] — the per-request state that enforces it, by
+//!   substituting the model's own `</think>` token when the budget runs
+//!   out so generation transitions to answering instead of being cut off.
+
+use serde::{Deserialize, Serialize};
+
+/// Per-model mapping from an OpenAI effort level to a token budget.
+///
+/// OpenAI-shaped clients cannot send a token count — pi-ai's vocabulary
+/// is the ladder `off | minimal | low | medium | high | xhigh | max` and
+/// nothing else — so the numbers have to live on the server. Operator
+/// configurable, because the right values depend on the model and the
+/// hardware; a derived default would be a guess baked into the binary.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ReasoningBudgetLadder {
+    pub minimal: usize,
+    pub low: usize,
+    pub medium: usize,
+    pub high: usize,
+    /// `xhigh` and `max` are only offered by a client when the model
+    /// declares them; absent here means "not offered", and a request
+    /// naming one falls back to `high` rather than being rejected.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub xhigh: Option<usize>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub max: Option<usize>,
+    /// Tokens of the caller's **own** output budget held back for the
+    /// answer, when the caller names no explicit `reasoning.max_tokens`.
+    ///
+    /// This is what an OpenAI-shaped client gets instead of the rungs
+    /// above, and it is deliberately not derived from effort. Effort is
+    /// a prompt instruction the model is trained to follow (#290);
+    /// making it *also* set a guillotine is what told Qwen3.8-27B to
+    /// reason at `xhigh` and then cut it off at `medium`'s 12,288.
+    ///
+    /// A reserve has no such double meaning. The model thinks as long
+    /// as it likes within the budget the caller declared, minus enough
+    /// to answer — so the caller always gets an answer, and the size of
+    /// the think block stays the model's decision rather than ours.
+    ///
+    /// Set to `0` to disable, restoring fully unbounded reasoning.
+    #[serde(default = "default_answer_reserve")]
+    pub answer_reserve_tokens: usize,
+}
+
+/// Enough for a substantive answer with tool calls, not so much that it
+/// meaningfully shortens a think block: at the 32,768 budget pi sends,
+/// this leaves ~28.7 k for reasoning — above the ~27.2 k a full agentic
+/// design turn was measured to use.
+fn default_answer_reserve() -> usize {
+    4_096
+}
+
+impl Default for ReasoningBudgetLadder {
+    /// Rungs sized against observed behaviour rather than round numbers:
+    /// a long agentic design turn on this class of model runs to ~27k
+    /// tokens, so `high` sits above that (bounded but rarely binding),
+    /// `medium` at roughly half (enough to plan a whole implementation),
+    /// and `low` at the scale of a focused single-file change.
+    fn default() -> Self {
+        Self {
+            minimal: 1_024,
+            low: 4_096,
+            medium: 12_288,
+            high: 32_768,
+            xhigh: None,
+            max: None,
+            answer_reserve_tokens: default_answer_reserve(),
+        }
+    }
+}
+
+impl ReasoningBudgetLadder {
+    /// Resolve an effort level to a token budget.
+    ///
+    /// `none`/`off` are handled before this — they disable thinking
+    /// outright via the chat template — so they have no budget here.
+    /// An unrecognised level resolves to `None` (unbounded) rather than
+    /// to a guess: silently substituting a budget nobody asked for is
+    /// the failure mode this whole area has been full of.
+    pub fn for_effort(&self, effort: &str) -> Option<usize> {
+        match effort.trim().to_ascii_lowercase().as_str() {
+            "minimal" => Some(self.minimal),
+            "low" => Some(self.low),
+            "medium" => Some(self.medium),
+            "high" => Some(self.high),
+            "xhigh" => Some(self.xhigh.unwrap_or(self.high)),
+            "max" => Some(self.max.or(self.xhigh).unwrap_or(self.high)),
+            _ => None,
+        }
+    }
+
+    /// The mapping as a client should see it on `/v1/models` — the rungs
+    /// this deployment actually offers, so a caller picking `low` knows
+    /// what it bought instead of guessing (#274).
+    pub fn advertised(&self) -> Vec<(String, usize)> {
+        let mut out = vec![
+            ("minimal".to_string(), self.minimal),
+            ("low".to_string(), self.low),
+            ("medium".to_string(), self.medium),
+            ("high".to_string(), self.high),
+        ];
+        if let Some(v) = self.xhigh {
+            out.push(("xhigh".to_string(), v));
+        }
+        if let Some(v) = self.max {
+            out.push(("max".to_string(), v));
+        }
+        out
+    }
+}
+
+/// Enforces a resolved budget over one request's token stream.
+///
+/// Fed every sampled token; returns the token that should actually be
+/// emitted. While the model is inside a think block and over budget, it
+/// returns the close marker instead, exactly once — after which the
+/// model is answering and the governor is inert.
+#[derive(Debug, Clone)]
+pub struct ReasoningGovernor {
+    limit: usize,
+    open_id: u32,
+    close_id: u32,
+    inside: bool,
+    used: usize,
+    forced: bool,
+}
+
+impl ReasoningGovernor {
+    /// `None` when the request named no budget, or the model has no
+    /// reasoning markers to enforce one with — in both cases generation
+    /// proceeds exactly as it did before this existed.
+    pub fn new(
+        limit: Option<usize>,
+        markers: Option<(u32, u32)>,
+        prompt_opened_reasoning: bool,
+    ) -> Option<Self> {
+        let limit = limit?;
+        let (open_id, close_id) = markers?;
+        Some(Self {
+            limit,
+            open_id,
+            close_id,
+            inside: prompt_opened_reasoning,
+            used: 0,
+            forced: false,
+        })
+    }
+
+    /// Whether the forced close has already been spent.
+    pub fn forced(&self) -> bool {
+        self.forced
+    }
+
+    /// Reasoning tokens counted so far.
+    pub fn used(&self) -> usize {
+        self.used
+    }
+
+    /// Govern one sampled token, returning the token to emit.
+    pub fn govern(&mut self, sampled: u32) -> u32 {
+        if sampled == self.open_id {
+            self.inside = true;
+            return sampled;
+        }
+        if sampled == self.close_id {
+            self.inside = false;
+            return sampled;
+        }
+        if !self.inside {
+            return sampled;
+        }
+        if self.used >= self.limit {
+            // Budget spent: hand the model its own close marker so it
+            // transitions to answering. Cutting the stream here instead
+            // would reproduce the very failure this exists to prevent.
+            self.inside = false;
+            self.forced = true;
+            return self.close_id;
+        }
+        self.used += 1;
+        sampled
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    const OPEN: u32 = 100;
+    const CLOSE: u32 = 101;
+
+    fn governor(limit: usize, opened: bool) -> ReasoningGovernor {
+        ReasoningGovernor::new(Some(limit), Some((OPEN, CLOSE)), opened).expect("governor")
+    }
+
+    #[test]
+    fn a_request_with_no_budget_is_ungoverned() {
+        assert!(ReasoningGovernor::new(None, Some((OPEN, CLOSE)), true).is_none());
+    }
+
+    /// A model with no detected markers cannot be governed — better
+    /// unbounded than truncated at an arbitrary token.
+    #[test]
+    fn a_model_without_markers_is_ungoverned() {
+        assert!(ReasoningGovernor::new(Some(10), None, true).is_none());
+    }
+
+    /// The case that motivated this: the chat template left the think
+    /// block open, so the model is reasoning from the first token.
+    #[test]
+    fn a_prompt_opened_block_is_bounded_and_force_closed() {
+        let mut g = governor(3, true);
+        assert_eq!(g.govern(7), 7);
+        assert_eq!(g.govern(8), 8);
+        assert_eq!(g.govern(9), 9);
+        assert_eq!(g.govern(10), CLOSE, "the fourth token must close the block");
+        assert!(g.forced());
+        assert_eq!(g.used(), 3);
+        // Answering now — the governor is inert.
+        assert_eq!(g.govern(11), 11);
+        assert_eq!(g.govern(12), 12);
+    }
+
+    /// Content before the block opens is never counted against the
+    /// reasoning budget.
+    #[test]
+    fn only_tokens_inside_the_block_are_counted() {
+        let mut g = governor(2, false);
+        assert_eq!(g.govern(5), 5);
+        assert_eq!(g.govern(6), 6);
+        assert_eq!(g.used(), 0);
+        assert_eq!(g.govern(OPEN), OPEN);
+        assert_eq!(g.govern(7), 7);
+        assert_eq!(g.govern(8), 8);
+        assert_eq!(g.govern(9), CLOSE);
+        assert_eq!(g.used(), 2);
+    }
+
+    /// A model that finishes thinking on its own is untouched — the
+    /// budget is a ceiling, not a target.
+    #[test]
+    fn a_model_that_stops_on_its_own_is_not_forced() {
+        let mut g = governor(100, true);
+        for t in 20..30 {
+            assert_eq!(g.govern(t), t);
+        }
+        assert_eq!(g.govern(CLOSE), CLOSE);
+        assert!(!g.forced());
+        for t in 40..50 {
+            assert_eq!(g.govern(t), t);
+        }
+        assert!(!g.forced(), "content after the block must not be governed");
+    }
+
+    #[test]
+    fn the_ladder_resolves_the_levels_clients_can_send() {
+        let l = ReasoningBudgetLadder::default();
+        assert_eq!(l.for_effort("minimal"), Some(1_024));
+        assert_eq!(l.for_effort("LOW"), Some(4_096));
+        assert_eq!(l.for_effort(" medium "), Some(12_288));
+        assert_eq!(l.for_effort("high"), Some(32_768));
+        // Not offered by default → falls back to high rather than
+        // rejecting a level the client was told it could send.
+        assert_eq!(l.for_effort("xhigh"), Some(32_768));
+        assert_eq!(l.for_effort("max"), Some(32_768));
+        // Unknown → unbounded, never a substituted guess.
+        assert_eq!(l.for_effort("enthusiastic"), None);
+        assert_eq!(l.for_effort("none"), None);
+    }
+
+    #[test]
+    fn the_advertised_ladder_lists_only_what_is_offered() {
+        let l = ReasoningBudgetLadder::default();
+        let names: Vec<_> = l.advertised().into_iter().map(|(k, _)| k).collect();
+        assert_eq!(names, ["minimal", "low", "medium", "high"]);
+        let l2 = ReasoningBudgetLadder {
+            xhigh: Some(49_152),
+            ..ReasoningBudgetLadder::default()
+        };
+        assert!(
+            l2.advertised()
+                .iter()
+                .any(|(k, v)| k == "xhigh" && *v == 49_152)
+        );
+    }
+}

crates/neuron/src/harness/reasoning_effort.rs

@@ -0,0 +1,302 @@
+//! Which reasoning-effort rungs a model actually supports, discovered
+//! from the model itself (#290).
+//!
+//! ## Why this is discovered rather than configured
+//!
+//! `Qwen/Qwen3.8-27B` accepts exactly three efforts and hard-fails on
+//! anything else:
+//!
+//! ```jinja
+//! {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
+//! {%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %}
+//!     {{- raise_exception('Unexpected reasoning effort ...') }}
+//! ```
+//!
+//! We previously advertised an invented ladder — `minimal`, `low`,
+//! `medium`, `high` — of which two values the template rejects outright
+//! and which omits `xhigh`, the model's own default and strongest rung.
+//!
+//! A caution about *why* that mattered, because the obvious story is
+//! wrong. It is tempting to say clients read our advertisement and were
+//! misled by it. pi-ai does not: its `getSupportedThinkingLevels` reads
+//! `model.thinkingLevelMap`, a static field in the operator's
+//! `models.json`, and treats `xhigh`/`max` as unavailable unless that
+//! map names them explicitly. So a pi user saw exactly
+//! `minimal, low, medium, high` — the invented ladder, to the value —
+//! for a reason that had nothing to do with what we published.
+//!
+//! The coincidence is the trap: our wrong advertisement and pi's
+//! default produced the same symptom, which made the advertisement look
+//! causal. Publishing rungs the model actually accepts is still correct
+//! — it is the truth, other clients may read it, and it is what makes
+//! `apply_effort_kwarg` right — but do not assume a given client
+//! consumes it. For pi the operator must also set `thinkingLevelMap`,
+//! by hand, and nothing validates that they match.
+//!
+//! A hardcoded per-model table would drift the moment a model ships a
+//! new template, which is exactly the class of defect #280 fixed by
+//! fetching the template instead of assuming it. So the rungs come from
+//! the artifact.
+//!
+//! ## How the probe works
+//!
+//! `raise_exception` is already wired to a render error, so the template
+//! answers the question directly: render it once per candidate effort
+//! and see which ones survive.
+//!
+//! Two ambiguities the probe has to resolve, or it would report
+//! confident nonsense:
+//!
+//! 1. **A template that ignores `reasoning_effort` entirely** renders
+//!    successfully for every candidate. Supporting "all six" would be a
+//!    lie. So a candidate only counts when the rendered prompt actually
+//!    *differs* across efforts — if every render is identical the model
+//!    has no effort control and we advertise none.
+//! 2. **The default** is whichever supported candidate renders
+//!    identically to a render with no effort supplied. Qwen3.8 defaults
+//!    to `xhigh`; without detecting that, a caller who says nothing gets
+//!    the strongest rung while we believe it got nothing.
+
+use anyhow::Result;
+use cortex_core::openai::{ChatMessage, MessageContent};
+use serde_json::{Value, json};
+
+/// Effort names in ascending order of thinking, as clients spell them.
+///
+/// Deliberately a superset of what any one model supports: this is the
+/// *candidate* set the probe tries, and the conventional vocabulary
+/// OpenAI-compatible clients use (pi-ai's `EXTENDED_THINKING_LEVELS`
+/// minus `off`, which is not an effort but a separate switch handled by
+/// `enable_thinking`).
+pub const CANDIDATE_EFFORTS: &[&str] = &["minimal", "low", "medium", "high", "xhigh", "max"];
+
+/// What one model's template accepts.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct SupportedEfforts {
+    /// Accepted efforts, ascending, as the template spells them.
+    pub levels: Vec<String>,
+    /// The effort applied when the caller names none — the template's
+    /// own `|default(...)`. `None` when it cannot be determined.
+    pub default: Option<String>,
+}
+
+impl SupportedEfforts {
+    pub fn is_empty(&self) -> bool {
+        self.levels.is_empty()
+    }
+
+    pub fn supports(&self, effort: &str) -> bool {
+        self.levels.iter().any(|l| l == effort)
+    }
+
+    /// Map a caller's effort onto the nearest rung this model has.
+    ///
+    /// Exact match wins. Otherwise walk **down** from the requested
+    /// level first, then up — a caller asking for more thinking than the
+    /// model offers should get the most it has, and a caller asking for
+    /// less should not be silently upgraded into a slower, costlier
+    /// response than they asked for.
+    ///
+    /// Descending-first is the opposite of pi-ai's `clampThinkingLevel`,
+    /// which searches upward first. That is deliberate: pi is choosing
+    /// among levels it believes exist, whereas we are the ones who know
+    /// what exists, and over-delivering thinking is the failure that
+    /// costs the caller money and latency.
+    ///
+    /// `None` when the model has no rungs at all.
+    pub fn nearest(&self, requested: &str) -> Option<&str> {
+        if self.levels.is_empty() {
+            return None;
+        }
+        if let Some(exact) = self.levels.iter().find(|l| *l == requested) {
+            return Some(exact.as_str());
+        }
+        let idx = CANDIDATE_EFFORTS.iter().position(|c| *c == requested)?;
+        for c in CANDIDATE_EFFORTS[..idx].iter().rev() {
+            if let Some(hit) = self.levels.iter().find(|l| l.as_str() == *c) {
+                return Some(hit.as_str());
+            }
+        }
+        for c in &CANDIDATE_EFFORTS[idx + 1..] {
+            if let Some(hit) = self.levels.iter().find(|l| l.as_str() == *c) {
+                return Some(hit.as_str());
+            }
+        }
+        None
+    }
+}
+
+/// Render a probe prompt at `effort`, or `None` if the template rejects
+/// it.
+fn probe_render(template: &str, effort: Option<&str>) -> Option<String> {
+    let messages = [ChatMessage {
+        role: "user".into(),
+        content: MessageContent::Text("probe".into()),
+        extra: Value::Null,
+    }];
+    let kwargs = match effort {
+        Some(e) => json!({ "reasoning_effort": e }),
+        None => Value::Null,
+    };
+    super::chat_template::render_chat_template(template, &messages, &Value::Null, &kwargs).ok()
+}
+
+/// Discover which efforts `template` accepts (#290).
+///
+/// Errors are not propagated: a template that cannot render even the
+/// baseline probe yields no rungs, and the caller behaves as it did
+/// before this existed.
+pub fn probe(template: &str, model_id: &str) -> Result<SupportedEfforts> {
+    let Some(baseline) = probe_render(template, None) else {
+        tracing::debug!(
+            model = %model_id,
+            "reasoning effort: template does not render a baseline probe; advertising no rungs"
+        );
+        return Ok(SupportedEfforts::default());
+    };
+
+    let mut rendered: Vec<(&str, String)> = Vec::new();
+    for candidate in CANDIDATE_EFFORTS {
+        if let Some(out) = probe_render(template, Some(candidate)) {
+            rendered.push((candidate, out));
+        }
+    }
+
+    // A template that never reads `reasoning_effort` renders the same
+    // bytes whatever we pass. Advertising every candidate in that case
+    // would be the invented-ladder mistake again, in a new place.
+    let effort_aware = rendered.windows(2).any(|w| w[0].1 != w[1].1)
+        || rendered.iter().any(|(_, out)| *out != baseline);
+    if !effort_aware {
+        tracing::debug!(
+            model = %model_id,
+            "reasoning effort: template ignores reasoning_effort; advertising no rungs"
+        );
+        return Ok(SupportedEfforts::default());
+    }
+
+    let default = rendered
+        .iter()
+        .find(|(_, out)| *out == baseline)
+        .map(|(name, _)| (*name).to_string());
+    let levels: Vec<String> = rendered
+        .iter()
+        .map(|(name, _)| (*name).to_string())
+        .collect();
+
+    tracing::info!(
+        model = %model_id,
+        levels = %levels.join(","),
+        default = ?default,
+        "reasoning effort: rungs discovered from the model's chat template (#290)"
+    );
+    Ok(SupportedEfforts { levels, default })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The probe against the **real** deployed Qwen3.8-27B template,
+    /// not a hand-written approximation of it. A synthetic template
+    /// proves the probe logic; only the shipped artifact proves the
+    /// answer, and the artifact is what #290 is about.
+    #[test]
+    fn the_real_qwen38_template_yields_low_medium_xhigh_defaulting_to_xhigh() {
+        let path = concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/tests/data/qwen3.8-27b.chat_template.jinja"
+        );
+        let Ok(template) = std::fs::read_to_string(path) else {
+            // The fixture is checked in; if it is ever removed this test
+            // should fail loudly rather than silently pass.
+            panic!("missing fixture {path}");
+        };
+        let s = probe(&template, "Qwen/Qwen3.8-27B").expect("probe");
+        assert_eq!(
+            s.levels,
+            vec!["low", "medium", "xhigh"],
+            "the shipped template accepts exactly these"
+        );
+        assert_eq!(s.default.as_deref(), Some("xhigh"));
+        // The two rungs we used to advertise are rejected outright.
+        assert!(!s.supports("minimal"));
+        assert!(!s.supports("high"));
+    }
+
+    /// The shape Qwen3.8-27B actually ships: three rungs, `xhigh`
+    /// default, `raise_exception` on anything else.
+    const QWEN38: &str = r#"
+{%- set resolved = reasoning_effort|default('xhigh') %}
+{%- if resolved not in ('xhigh', 'medium', 'low') %}
+{{- raise_exception('Unexpected reasoning effort ' ~ resolved) }}
+{%- endif %}
+{%- if resolved == 'xhigh' %}EFFORT-XHIGH{% elif resolved == 'low' %}EFFORT-LOW{% else %}EFFORT-MEDIUM{% endif %}
+{%- for m in messages %}<|im_start|>{{ m.role }}
+{{ m.content }}<|im_end|>
+{%- endfor %}
+"#;
+
+    /// A template with no effort control at all.
+    const PLAIN: &str = r#"
+{%- for m in messages %}<|im_start|>{{ m.role }}
+{{ m.content }}<|im_end|>
+{%- endfor %}
+"#;
+
+    #[test]
+    fn discovers_exactly_the_rungs_the_template_accepts() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.levels, vec!["low", "medium", "xhigh"]);
+        assert!(!s.supports("minimal"), "template raises on minimal");
+        assert!(!s.supports("high"), "template raises on high");
+    }
+
+    /// Without this, a caller who names no effort silently receives the
+    /// strongest rung while we believe it received nothing — which is
+    /// how the xhigh instruction ended up on every request in #290.
+    #[test]
+    fn detects_the_templates_own_default() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.default.as_deref(), Some("xhigh"));
+    }
+
+    /// A template that ignores the kwarg must advertise nothing rather
+    /// than appearing to support all six.
+    #[test]
+    fn a_template_without_effort_control_advertises_no_rungs() {
+        let s = probe(PLAIN, "test").expect("probe");
+        assert!(s.is_empty(), "got {:?}", s.levels);
+        assert_eq!(s.nearest("medium"), None);
+    }
+
+    #[test]
+    fn an_exact_match_is_returned_unchanged() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.nearest("medium"), Some("medium"));
+        assert_eq!(s.nearest("xhigh"), Some("xhigh"));
+    }
+
+    /// `high` does not exist on this model. Descending first means it
+    /// resolves to `medium`, not `xhigh`: a caller is never given more
+    /// thinking (slower, dearer) than it asked for.
+    #[test]
+    fn an_unsupported_level_resolves_downward_not_upward() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.nearest("high"), Some("medium"));
+    }
+
+    /// `minimal` is below every rung the model has, so there is nothing
+    /// to descend to and the nearest is the weakest available.
+    #[test]
+    fn a_level_below_every_rung_resolves_up_to_the_weakest() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.nearest("minimal"), Some("low"));
+    }
+
+    #[test]
+    fn an_unknown_level_name_resolves_to_nothing_rather_than_a_guess() {
+        let s = probe(QWEN38, "test").expect("probe");
+        assert_eq!(s.nearest("turbo"), None);
+    }
+}

crates/neuron/src/harness/sampling.rs

@@ -1,10 +1,29 @@
-//! Sampling parameter resolution (#272).
+//! Sampling parameter resolution (#272, #283).
 //!
-//! Three sources, in priority order:
+//! Four sources, in priority order:
 //!
 //! 1. what the **request** asked for,
-//! 2. what the **model** published in `generation_config.json`,
-//! 3. a built-in fallback.
+//! 2. what the **operator** set for this model (#283),
+//! 3. what the **model** published in `generation_config.json`,
+//! 4. a built-in fallback.
+//!
+//! Tier (2) exists because honouring the model's own file is right by
+//! default and still wrong sometimes: `Qwen/Qwen3.8-27B` publishes
+//! `temperature = 1.0`, which measured a 20% structural-defect rate on
+//! ~2k-token code generation against 0/60 at `<= 0.6` (#283,
+//! p = 0.0031). The operator's correction is explicit and lives in a
+//! config file — never a heuristic that inspects the model and guesses,
+//! which would be a footgun.
+//!
+//! The operator sits *below* the request on purpose. An explicit API
+//! parameter that is silently ignored is a contract break; the observed
+//! failure came from callers that named nothing, so a default-tier
+//! override is sufficient to fix it.
+//!
+//! Mechanically, tier (2) is folded into tier (3) once at load time by
+//! [`ModelGenerationDefaults::with_operator_override`], so the
+//! per-request resolve path stays a two-tier lookup and every call site
+//! gets the override for free — there is no site that can forget it.
 //!
 //! Before this, (2) did not exist and (3) was a hardcoded
 //! `temperature = 0.7` with `top_p = None` — and `None` selects
@@ -56,6 +75,15 @@ pub struct ModelGenerationDefaults {
     pub temperature: Option<f64>,
     pub top_p: Option<f64>,
     pub top_k: Option<usize>,
+    /// Not published in `generation_config.json` — this struct is the
+    /// resolved per-model default layer, and the operator override
+    /// (#283) lands here so the repetition knobs sit in the same
+    /// precedence chain as the rest.
+    pub repeat_penalty: Option<f32>,
+    pub repeat_last_n: Option<usize>,
+    pub presence_penalty: Option<f32>,
+    pub frequency_penalty: Option<f32>,
+    pub seed: Option<u64>,
 }
 
 impl ModelGenerationDefaults {
@@ -98,6 +126,76 @@ impl ModelGenerationDefaults {
             }
         }
     }
+
+    /// Overlay the operator's per-model override (#283) onto what the
+    /// model published, field by field.
+    ///
+    /// Applied **once at load**, not per request, so the override cannot
+    /// be forgotten at one of the resolve sites — the same reasoning
+    /// that made #272 resolve sampling once rather than at eight
+    /// `LogitsProcessor` construction sites.
+    ///
+    /// Each field overlays independently: setting only `temperature`
+    /// leaves the model's `top_p`/`top_k` in force, because an operator
+    /// correcting one number should not silently blank the other two
+    /// back to "no truncation".
+    ///
+    /// Logged at `info` whenever it changes something — an operator
+    /// override that is in force but invisible is how a fleet ends up
+    /// with two configs that disagree and nobody noticing (#252).
+    pub fn with_operator_override(
+        mut self,
+        model_id: &str,
+        operator: Option<&cortex_core::harness::SamplingOverride>,
+    ) -> Self {
+        let Some(op) = operator else {
+            return self;
+        };
+        if op.is_empty() {
+            return self;
+        }
+        let before = self.clone();
+        if let Some(t) = op.temperature {
+            self.temperature = Some(t);
+        }
+        if let Some(p) = op.top_p {
+            self.top_p = Some(p);
+        }
+        if let Some(k) = op.top_k {
+            self.top_k = Some(k);
+        }
+        if let Some(r) = op.repeat_penalty {
+            self.repeat_penalty = Some(r);
+        }
+        if let Some(n) = op.repeat_last_n {
+            self.repeat_last_n = Some(n);
+        }
+        if let Some(v) = op.presence_penalty {
+            self.presence_penalty = Some(v);
+        }
+        if let Some(v) = op.frequency_penalty {
+            self.frequency_penalty = Some(v);
+        }
+        if let Some(v) = op.seed {
+            self.seed = Some(v);
+        }
+        tracing::info!(
+            model = %model_id,
+            model_temperature = ?before.temperature,
+            model_top_p = ?before.top_p,
+            model_top_k = ?before.top_k,
+            temperature = ?self.temperature,
+            top_p = ?self.top_p,
+            top_k = ?self.top_k,
+            repeat_penalty = ?self.repeat_penalty,
+            repeat_last_n = ?self.repeat_last_n,
+            presence_penalty = ?self.presence_penalty,
+            frequency_penalty = ?self.frequency_penalty,
+            seed = ?self.seed,
+            "sampling: operator override applied over the model's published defaults (#283)"
+        );
+        self
+    }
 }
 
 /// What the caller asked for. All optional — absent means "no opinion",
@@ -110,6 +208,8 @@ pub struct RequestedSampling {
     pub seed: Option<u64>,
     pub repeat_penalty: Option<f32>,
     pub repeat_last_n: Option<usize>,
+    pub presence_penalty: Option<f32>,
+    pub frequency_penalty: Option<f32>,
 }
 
 /// Fully-resolved sampling for one request.
@@ -121,6 +221,9 @@ pub struct SamplingParams {
     pub seed: u64,
     pub repeat_penalty: f32,
     pub repeat_last_n: usize,
+    /// Sequence-wide penalties, OpenAI semantics. `0.0` disables.
+    pub presence_penalty: f32,
+    pub frequency_penalty: f32,
 }
 
 impl SamplingParams {
@@ -146,9 +249,23 @@ impl SamplingParams {
             // whenever the caller stayed quiet.
             top_p: requested.top_p.or(model.top_p),
             top_k: requested.top_k.or(model.top_k),
-            seed: requested.seed.unwrap_or(random_seed),
-            repeat_penalty: requested.repeat_penalty.unwrap_or(DEFAULT_REPEAT_PENALTY),
-            repeat_last_n: requested.repeat_last_n.unwrap_or(DEFAULT_REPEAT_LAST_N),
+            seed: requested.seed.or(model.seed).unwrap_or(random_seed),
+            repeat_penalty: requested
+                .repeat_penalty
+                .or(model.repeat_penalty)
+                .unwrap_or(DEFAULT_REPEAT_PENALTY),
+            repeat_last_n: requested
+                .repeat_last_n
+                .or(model.repeat_last_n)
+                .unwrap_or(DEFAULT_REPEAT_LAST_N),
+            presence_penalty: requested
+                .presence_penalty
+                .or(model.presence_penalty)
+                .unwrap_or(0.0),
+            frequency_penalty: requested
+                .frequency_penalty
+                .or(model.frequency_penalty)
+                .unwrap_or(0.0),
         }
     }
 
@@ -176,12 +293,14 @@ impl SamplingParams {
 mod tests {
     use super::*;
     use candle_transformers::generation::Sampling;
+    use cortex_core::harness::SamplingOverride;
 
     fn qwen3_defaults() -> ModelGenerationDefaults {
         ModelGenerationDefaults {
             temperature: Some(1.0),
             top_p: Some(0.95),
             top_k: Some(20),
+            ..Default::default()
         }
     }
 
@@ -221,9 +340,207 @@ mod tests {
         assert_eq!(p.top_p, Some(0.95));
     }
 
+    // ── Operator override (#283) ──────────────────────────────
+
+    fn op(temperature: Option<f64>, top_p: Option<f64>, top_k: Option<usize>) -> SamplingOverride {
+        SamplingOverride {
+            temperature,
+            top_p,
+            top_k,
+            ..Default::default()
+        }
+    }
+
+    fn with_op(o: SamplingOverride) -> ModelGenerationDefaults {
+        qwen3_defaults().with_operator_override("Qwen/Qwen3.8-27B", Some(&o))
+    }
+
+    /// The #283 case: the model publishes `temperature = 1.0`, which
+    /// measured a 20% structural-defect rate on code generation. The
+    /// operator sets 0.6 and a silent caller must get 0.6.
+    #[test]
+    fn the_operator_override_replaces_what_the_model_published() {
+        let p = SamplingParams::resolve(
+            &RequestedSampling::default(),
+            &with_op(op(Some(0.6), None, None)),
+            42,
+        );
+        assert_eq!(p.temperature, 0.6);
+    }
+
+    /// Precedence is request > operator > model. An explicit API
+    /// parameter that is silently ignored is a contract break, so a
+    /// caller naming a temperature still wins over the operator.
+    #[test]
+    fn an_explicit_request_still_beats_the_operator() {
+        let requested = RequestedSampling {
+            temperature: Some(0.9),
+            ..Default::default()
+        };
+        let p = SamplingParams::resolve(&requested, &with_op(op(Some(0.6), None, None)), 42);
+        assert_eq!(p.temperature, 0.9);
+    }
+
+    /// Overriding one field must not blank the others back to "no
+    /// truncation" — that was the #271 failure mode, reintroduced by a
+    /// wholesale replace instead of a field-by-field overlay.
+    #[test]
+    fn overriding_temperature_alone_leaves_top_p_and_top_k_intact() {
+        let d = with_op(op(Some(0.6), None, None));
+        assert_eq!(d.temperature, Some(0.6));
+        assert_eq!(d.top_p, Some(0.95), "the model's top_p must survive");
+        assert_eq!(d.top_k, Some(20), "the model's top_k must survive");
+    }
+
+    /// Each field overlays independently.
+    #[test]
+    fn every_field_can_be_overridden_independently() {
+        let d = with_op(op(None, Some(0.8), Some(40)));
+        assert_eq!(d.temperature, Some(1.0), "untouched, so the model's");
+        assert_eq!(d.top_p, Some(0.8));
+        assert_eq!(d.top_k, Some(40));
+    }
+
+    /// No override, and an override that sets nothing, must both be
+    /// indistinguishable from the pre-#283 behaviour.
+    #[test]
+    fn absent_or_empty_override_changes_nothing() {
+        let base = qwen3_defaults();
+        let none = qwen3_defaults().with_operator_override("m", None);
+        let empty =
+            qwen3_defaults().with_operator_override("m", Some(&SamplingOverride::default()));
+        for d in [none, empty] {
+            assert_eq!(d.temperature, base.temperature);
+            assert_eq!(d.top_p, base.top_p);
+            assert_eq!(d.top_k, base.top_k);
+        }
+    }
+
+    /// An operator override on a model that ships no
+    /// `generation_config.json` must still take effect, rather than
+    /// falling through to the built-in 0.7.
+    #[test]
+    fn the_override_applies_even_without_a_generation_config() {
+        let d = ModelGenerationDefaults::default()
+            .with_operator_override("m", Some(&op(Some(0.6), None, None)));
+        let p = SamplingParams::resolve(&RequestedSampling::default(), &d, 42);
+        assert_eq!(p.temperature, 0.6);
+        assert_ne!(p.temperature, FALLBACK_TEMPERATURE);
+    }
+
+    /// `temperature = 0` from an operator must select greedy decoding,
+    /// not be treated as "unset" by an `unwrap_or`-shaped bug.
+    #[test]
+    fn an_operator_temperature_of_zero_is_greedy_not_unset() {
+        let p = SamplingParams::resolve(
+            &RequestedSampling::default(),
+            &with_op(op(Some(0.0), None, None)),
+            42,
+        );
+        assert_eq!(p.temperature, 0.0);
+        assert_eq!(p.to_sampling(), Sampling::ArgMax);
+    }
+
     /// A model with no `generation_config.json` must behave exactly as
     /// it did before this landed, so nothing regresses on the models
     /// that do not ship one.
+    /// An operator can pin the seed, because the client that needs it
+    /// pinned cannot send one: pi's request builder emits no `seed`, so
+    /// without this a parameter search runs unseeded and sampling
+    /// variance is indistinguishable from the effect being measured.
+    #[test]
+    fn an_operator_can_pin_the_seed() {
+        let op = cortex_core::harness::SamplingOverride {
+            seed: Some(424_242),
+            ..Default::default()
+        };
+        let defaults = ModelGenerationDefaults::default().with_operator_override("m", Some(&op));
+        // The random seed argument must be ignored once an operator pinned one.
+        let p = SamplingParams::resolve(&RequestedSampling::default(), &defaults, 999);
+        assert_eq!(p.seed, 424_242);
+    }
+
+    /// Precedence holds: a caller that names its own seed keeps it.
+    #[test]
+    fn a_requested_seed_outranks_the_operator() {
+        let op = cortex_core::harness::SamplingOverride {
+            seed: Some(424_242),
+            ..Default::default()
+        };
+        let defaults = ModelGenerationDefaults::default().with_operator_override("m", Some(&op));
+        let req = RequestedSampling {
+            seed: Some(7),
+            ..Default::default()
+        };
+        assert_eq!(SamplingParams::resolve(&req, &defaults, 999).seed, 7);
+    }
+
+    /// With neither, the per-request random seed still applies — an
+    /// unset override must not accidentally freeze the fleet's RNG.
+    #[test]
+    fn no_pinned_seed_leaves_generation_random() {
+        let d = ModelGenerationDefaults::default().with_operator_override("m", None);
+        assert_eq!(
+            SamplingParams::resolve(&RequestedSampling::default(), &d, 999).seed,
+            999
+        );
+    }
+
+    /// The gap this closes: `repeat_last_n` defaults to 64 — candle's
+    /// chat example value — while a reasoning model's think block runs
+    /// to tens of thousands of tokens. Anything restated more than 64
+    /// tokens later is invisible to the penalty, and before this there
+    /// was no operator route to the knob at all: `SamplingOverride`
+    /// carried only temperature/top_p/top_k, and OpenAI-shaped clients
+    /// do not send repetition parameters. The 64 was effectively
+    /// hardcoded in production.
+    #[test]
+    fn an_operator_can_widen_the_repetition_window() {
+        let op = cortex_core::harness::SamplingOverride {
+            repeat_penalty: Some(1.05),
+            repeat_last_n: Some(2048),
+            presence_penalty: None,
+            frequency_penalty: None,
+            ..Default::default()
+        };
+        let defaults = ModelGenerationDefaults::default().with_operator_override("m", Some(&op));
+        let p = SamplingParams::resolve(&RequestedSampling::default(), &defaults, 7);
+        assert_eq!(p.repeat_last_n, 2048);
+        assert!((p.repeat_penalty - 1.05).abs() < f32::EPSILON);
+    }
+
+    /// Precedence is request > operator > built-in, same as every other
+    /// sampling field — a caller that named a window keeps it.
+    #[test]
+    fn a_requested_repetition_window_outranks_the_operator() {
+        let op = cortex_core::harness::SamplingOverride {
+            repeat_last_n: Some(2048),
+            presence_penalty: None,
+            frequency_penalty: None,
+            ..Default::default()
+        };
+        let defaults = ModelGenerationDefaults::default().with_operator_override("m", Some(&op));
+        let req = RequestedSampling {
+            repeat_last_n: Some(128),
+            presence_penalty: None,
+            frequency_penalty: None,
+            ..Default::default()
+        };
+        assert_eq!(
+            SamplingParams::resolve(&req, &defaults, 7).repeat_last_n,
+            128
+        );
+    }
+
+    /// An operator who sets nothing must not change the constants.
+    #[test]
+    fn no_operator_setting_leaves_the_repetition_defaults_alone() {
+        let d = ModelGenerationDefaults::default().with_operator_override("m", None);
+        let p = SamplingParams::resolve(&RequestedSampling::default(), &d, 7);
+        assert_eq!(p.repeat_last_n, DEFAULT_REPEAT_LAST_N);
+        assert!((p.repeat_penalty - DEFAULT_REPEAT_PENALTY).abs() < f32::EPSILON);
+    }
+
     #[test]
     fn without_model_defaults_the_old_behaviour_is_preserved() {
         let p = SamplingParams::resolve(
@@ -325,6 +642,8 @@ mod tests {
             &RequestedSampling {
                 repeat_penalty: Some(1.05),
                 repeat_last_n: Some(512),
+                presence_penalty: None,
+                frequency_penalty: None,
                 ..Default::default()
             },
             &ModelGenerationDefaults::default(),

crates/neuron/src/harness/tp/tp_qwen3_5.rs

@@ -2445,7 +2445,7 @@ mod tests {
     }
 
     #[cfg(not(feature = "cuda"))]
-    fn vb_over(path: &std::path::Path) -> ShardedVarBuilder {
+    fn vb_over(path: &std::path::Path) -> ShardedVarBuilder<'_> {
         // SAFETY: mmap of a file the test just wrote; nothing mutates it.
         unsafe {
             candle_nn::var_builder::ShardedSafeTensors::var_builder(

crates/neuron/src/wire/openai_chat.rs

@@ -27,8 +27,8 @@
 //! propagate without us writing any logic.
 
 use cortex_core::openai::{
-    ChatCompletionChunk, ChunkChoice, CompletionTokensDetails, HelexaTiming, PromptTokensDetails,
-    Usage,
+    ChatCompletionChoice, ChatCompletionChunk, ChatCompletionResponse, ChatMessage, ChunkChoice,
+    CompletionTokensDetails, HelexaTiming, MessageContent, PromptTokensDetails, Usage,
 };
 use serde_json::json;
 use tokio::sync::mpsc;
@@ -450,10 +450,379 @@ fn usage_chunk(
     }
 }
 
+/// Drain an [`InferenceEvent`] stream into a single non-streaming
+/// [`ChatCompletionResponse`] (#285).
+///
+/// Non-streaming used to be a *separate inference path* that took the
+/// per-model `inference_lock` for the whole request. That made it
+/// invisible to the batch engine (#98), so concurrent non-streaming
+/// callers serialized completely — measured on beast at 1.00x aggregate
+/// throughput across 1/2/4/8 concurrency, against 3.98x for the same
+/// work streamed. `max_in_flight` bought those requests admission and
+/// nothing else.
+///
+/// Expressing non-streaming as "run the stream, buffer the result"
+/// collapses the two paths onto one producer, so batching, prefix-cache
+/// reuse, tool-call extraction, the reasoning governor and server-side
+/// timing all apply identically to both. It also removes a second place
+/// where the same wire object could be assembled differently — the
+/// shape of defect #252 and #272 both were.
+///
+/// Reasoning is kept out of `content` — it is not the assistant's
+/// reply — but it is **returned** on `message.reasoning_content`, the
+/// same field the streaming path emits as
+/// `choice.delta.reasoning_content`.
+///
+/// It used to be discarded here (#300). That was inherited from the
+/// path this replaced and was wrong on both counts: the tokens are
+/// metered in `usage.completion_tokens_details.reasoning_tokens`, so a
+/// caller paid for text it could never see; and a client cannot replay
+/// what it never received, so on the next turn the model saw its own
+/// prior turn as a bare assertion with the reasoning that produced it
+/// deleted.
+///
+/// `finish_reason` becomes `"tool_calls"` whenever any tool call was
+/// parsed.
+///
+/// Two fields that the old path hardcoded to `None` now populate,
+/// because the events carry them and the struct already had the slots:
+/// `prompt_tokens_details.cached_tokens` and `helexa_timing`.
+///
+/// A stream that ends without a `Finish` is an error, not an empty
+/// reply: that is how a dropped engine slot or a fatal worker fault
+/// presents, and returning a 200 with partial text would hide it.
+pub async fn collect_chat_completion(
+    mut rx: mpsc::Receiver<InferenceEvent>,
+    id: String,
+    created: u64,
+    model_id: String,
+) -> Result<ChatCompletionResponse, String> {
+    let mut content = String::new();
+    let mut reasoning = String::new();
+    let mut tool_calls: Vec<serde_json::Value> = Vec::new();
+    let mut finish: Option<(FinishReason, u32, u32, u32, u32, Option<FinishTiming>)> = None;
+
+    while let Some(event) = rx.recv().await {
+        match event {
+            InferenceEvent::Start => {}
+            InferenceEvent::TextDelta(t) => content.push_str(&t),
+            // Kept out of `content`, returned on `reasoning_content`.
+            InferenceEvent::ReasoningDelta(t) => reasoning.push_str(&t),
+            InferenceEvent::ToolCall {
+                index,
+                id: call_id,
+                name,
+                arguments,
+            } => tool_calls.push(json!({
+                "index": index,
+                "id": call_id,
+                "type": "function",
+                "function": { "name": name, "arguments": arguments },
+            })),
+            InferenceEvent::Finish {
+                reason,
+                prompt_tokens,
+                completion_tokens,
+                reasoning_tokens,
+                cached_tokens,
+                timing,
+            } => {
+                finish = Some((
+                    reason,
+                    prompt_tokens,
+                    completion_tokens,
+                    reasoning_tokens,
+                    cached_tokens,
+                    timing,
+                ));
+                break;
+            }
+        }
+    }
+
+    let Some((reason, prompt_tokens, completion_tokens, reasoning_tokens, cached_tokens, timing)) =
+        finish
+    else {
+        return Err(
+            "inference stream ended without a finish event (engine slot dropped?)".to_string(),
+        );
+    };
+
+    // A parsed tool call outranks whatever the sampler stopped on, so a
+    // model that emits a call and then hits EOS still reports
+    // `tool_calls` and the client dispatches instead of rendering.
+    let finish_reason = if tool_calls.is_empty() {
+        reason.as_openai_str().to_string()
+    } else {
+        "tool_calls".to_string()
+    };
+    let mut message_extra = serde_json::Map::new();
+    if !tool_calls.is_empty() {
+        message_extra.insert("tool_calls".into(), json!(tool_calls));
+    }
+    // Absent rather than empty when the model did not think: a
+    // non-reasoning model must keep the unchanged message shape.
+    if !reasoning.is_empty() {
+        message_extra.insert("reasoning_content".into(), json!(reasoning));
+    }
+    let message_extra = serde_json::Value::Object(message_extra);
+
+    Ok(ChatCompletionResponse {
+        id,
+        object: "chat.completion".into(),
+        created,
+        model: model_id,
+        choices: vec![ChatCompletionChoice {
+            index: 0,
+            message: ChatMessage {
+                role: "assistant".into(),
+                content: MessageContent::Text(content),
+                extra: message_extra,
+            },
+            finish_reason: Some(finish_reason),
+            extra: serde_json::Value::Object(Default::default()),
+        }],
+        usage: Some(Usage {
+            prompt_tokens: prompt_tokens as u64,
+            completion_tokens: completion_tokens as u64,
+            total_tokens: (prompt_tokens + completion_tokens) as u64,
+            completion_tokens_details: (reasoning_tokens > 0).then_some(CompletionTokensDetails {
+                reasoning_tokens: reasoning_tokens as u64,
+            }),
+            prompt_tokens_details: (cached_tokens > 0).then_some(PromptTokensDetails {
+                cached_tokens: cached_tokens as u64,
+            }),
+            helexa_timing: timing.map(|t| HelexaTiming {
+                prefill_ms: t.prefill_ms as u64,
+                decode_ms: t.decode_ms as u64,
+                prefill_tokens: t.prefill_tokens as u64,
+            }),
+        }),
+        extra: serde_json::Value::Object(Default::default()),
+    })
+}
+
 #[cfg(test)]
 mod tests {
+
+    /// The model's think block must come back on the non-streaming
+    /// path too (#300).
+    ///
+    /// It used to be dropped here while the streaming path emitted it
+    /// as `choice.delta.reasoning_content`. The tokens were metered
+    /// either way, so a non-streaming caller paid for text it could
+    /// never see — and, worse, could not replay the model's reasoning
+    /// on the next turn, leaving the model to face its own prior turn
+    /// as a bare assertion.
+    #[tokio::test]
+    async fn non_streaming_returns_the_models_reasoning() {
+        let (tx, rx) = mpsc::channel::<InferenceEvent>(8);
+        tx.send(InferenceEvent::Start).await.unwrap();
+        tx.send(InferenceEvent::ReasoningDelta(
+            "weighing the options".into(),
+        ))
+        .await
+        .unwrap();
+        tx.send(InferenceEvent::TextDelta("the answer".into()))
+            .await
+            .unwrap();
+        tx.send(InferenceEvent::Finish {
+            reason: FinishReason::Stop,
+            prompt_tokens: 5,
+            completion_tokens: 7,
+            reasoning_tokens: 3,
+            cached_tokens: 0,
+            timing: None,
+        })
+        .await
+        .unwrap();
+        drop(tx);
+
+        let r = collect_chat_completion(rx, "id".into(), 0, "m".into())
+            .await
+            .expect("collect");
+        let msg = &r.choices[0].message;
+        assert_eq!(
+            msg.extra["reasoning_content"], "weighing the options",
+            "reasoning must be returned, not discarded"
+        );
+        // And it must NOT be folded into the visible reply.
+        assert!(
+            matches!(&msg.content, MessageContent::Text(t) if t == "the answer"),
+            "reasoning leaked into content"
+        );
+    }
+
+    /// A non-reasoning model must keep the unchanged message shape —
+    /// absent, not an empty string.
+    #[tokio::test]
+    async fn no_reasoning_means_no_reasoning_field() {
+        let (tx, rx) = mpsc::channel::<InferenceEvent>(8);
+        tx.send(InferenceEvent::Start).await.unwrap();
+        tx.send(InferenceEvent::TextDelta("plain".into()))
+            .await
+            .unwrap();
+        tx.send(InferenceEvent::Finish {
+            reason: FinishReason::Stop,
+            prompt_tokens: 1,
+            completion_tokens: 1,
+            reasoning_tokens: 0,
+            cached_tokens: 0,
+            timing: None,
+        })
+        .await
+        .unwrap();
+        drop(tx);
+        let r = collect_chat_completion(rx, "id".into(), 0, "m".into())
+            .await
+            .expect("collect");
+        assert!(
+            r.choices[0]
+                .message
+                .extra
+                .get("reasoning_content")
+                .is_none()
+        );
+    }
     use super::*;
 
+    // ── Non-streaming collector (#285) ────────────────────────────
+
+    async fn collect_from(events: Vec<InferenceEvent>) -> Result<ChatCompletionResponse, String> {
+        let (tx, rx) = mpsc::channel(64);
+        for e in events {
+            tx.send(e).await.expect("send");
+        }
+        drop(tx);
+        collect_chat_completion(rx, "chatcmpl-test".into(), 1_700_000_000, "m".into()).await
+    }
+
+    fn finish(reason: FinishReason) -> InferenceEvent {
+        InferenceEvent::Finish {
+            reason,
+            prompt_tokens: 10,
+            completion_tokens: 4,
+            reasoning_tokens: 0,
+            cached_tokens: 0,
+            timing: None,
+        }
+    }
+
+    #[tokio::test]
+    async fn collector_concatenates_text_deltas() {
+        let r = collect_from(vec![
+            InferenceEvent::Start,
+            InferenceEvent::TextDelta("hel".into()),
+            InferenceEvent::TextDelta("lo".into()),
+            finish(FinishReason::Stop),
+        ])
+        .await
+        .expect("collected");
+        let c = &r.choices[0];
+        assert!(matches!(&c.message.content, MessageContent::Text(t) if t == "hello"));
+        assert_eq!(c.finish_reason.as_deref(), Some("stop"));
+        assert_eq!(r.object, "chat.completion");
+    }
+
+    /// Reasoning is counted but never rendered — the behaviour of the
+    /// path this replaced. A regression here would leak scratchpad into
+    /// `content` for every non-streaming caller.
+    #[tokio::test]
+    async fn collector_drops_reasoning_text_but_keeps_the_count() {
+        let (tx, rx) = mpsc::channel(64);
+        tx.send(InferenceEvent::ReasoningDelta("secret".into()))
+            .await
+            .expect("send");
+        tx.send(InferenceEvent::TextDelta("answer".into()))
+            .await
+            .expect("send");
+        tx.send(InferenceEvent::Finish {
+            reason: FinishReason::Stop,
+            prompt_tokens: 10,
+            completion_tokens: 9,
+            reasoning_tokens: 5,
+            cached_tokens: 0,
+            timing: None,
+        })
+        .await
+        .expect("send");
+        drop(tx);
+        let r = collect_chat_completion(rx, "id".into(), 1, "m".into())
+            .await
+            .expect("collected");
+        let c = &r.choices[0];
+        assert!(matches!(&c.message.content, MessageContent::Text(t) if t == "answer"));
+        let u = r.usage.expect("usage");
+        assert_eq!(
+            u.completion_tokens_details
+                .expect("details")
+                .reasoning_tokens,
+            5
+        );
+    }
+
+    /// A parsed tool call outranks the sampler's stop reason, matching
+    /// the streaming projector.
+    #[tokio::test]
+    async fn collector_promotes_finish_reason_to_tool_calls() {
+        let r = collect_from(vec![
+            InferenceEvent::ToolCall {
+                index: 0,
+                id: "call_1".into(),
+                name: "read_file".into(),
+                arguments: r#"{"path":"a.txt"}"#.into(),
+            },
+            finish(FinishReason::Stop),
+        ])
+        .await
+        .expect("collected");
+        let c = &r.choices[0];
+        assert_eq!(c.finish_reason.as_deref(), Some("tool_calls"));
+        let calls = c.message.extra.get("tool_calls").expect("tool_calls");
+        assert_eq!(calls[0]["function"]["name"], "read_file");
+        assert_eq!(calls[0]["type"], "function");
+        assert_eq!(calls[0]["id"], "call_1");
+    }
+
+    /// cached_tokens and helexa_timing were hardcoded to `None` on the
+    /// old non-streaming path; the events carried them all along.
+    #[tokio::test]
+    async fn collector_surfaces_cached_tokens_and_timing() {
+        let (tx, rx) = mpsc::channel(8);
+        tx.send(InferenceEvent::Finish {
+            reason: FinishReason::Stop,
+            prompt_tokens: 100,
+            completion_tokens: 5,
+            reasoning_tokens: 0,
+            cached_tokens: 64,
+            timing: Some(FinishTiming {
+                prefill_ms: 12,
+                decode_ms: 34,
+                prefill_tokens: 100,
+            }),
+        })
+        .await
+        .expect("send");
+        drop(tx);
+        let r = collect_chat_completion(rx, "id".into(), 1, "m".into())
+            .await
+            .expect("collected");
+        let u = r.usage.expect("usage");
+        assert_eq!(u.prompt_tokens_details.expect("details").cached_tokens, 64);
+        let t = u.helexa_timing.expect("timing");
+        assert_eq!((t.prefill_ms, t.decode_ms, t.prefill_tokens), (12, 34, 100));
+    }
+
+    /// A dropped engine slot ends the channel with no Finish. Returning
+    /// a 200 with partial text would hide a fatal worker fault.
+    #[tokio::test]
+    async fn collector_errors_when_the_stream_ends_without_finish() {
+        let err = collect_from(vec![InferenceEvent::TextDelta("partial".into())])
+            .await
+            .expect_err("must not succeed");
+        assert!(err.contains("without a finish"), "unexpected: {err}");
+    }
+
     /// Drain the projection's output into a Vec for assertion.
     async fn collect(mut rx: mpsc::Receiver<ChatCompletionChunk>) -> Vec<ChatCompletionChunk> {
         let mut out = Vec::new();

crates/neuron/tests/activation.rs

@@ -29,6 +29,8 @@ async fn test_load_default_models_skips_unknown_harness() {
             quant: None,
             tensor_parallel: None,
             devices: None,
+            sampling: None,
+            preserve_thinking: None,
         },
         ModelSpec {
             model_id: "model-b".into(),
@@ -36,6 +38,8 @@ async fn test_load_default_models_skips_unknown_harness() {
             quant: None,
             tensor_parallel: None,
             devices: None,
+            sampling: None,
+            preserve_thinking: None,
         },
     ];
 
@@ -95,6 +99,8 @@ async fn test_load_default_models_skipped_on_driver_mismatch() {
         quant: Some("q6k".into()),
         tensor_parallel: Some(2),
         devices: None,
+        sampling: None,
+        preserve_thinking: None,
     }];
     let activation = ActivationTracker::new(&specs);
     let reason = "host NVIDIA driver/library mismatch (userspace NVML 580.159 vs loaded \
@@ -196,6 +202,8 @@ fn qwen_spec() -> ModelSpec {
         quant: None,
         tensor_parallel: None,
         devices: None,
+        sampling: None,
+        preserve_thinking: None,
     }
 }
 

crates/neuron/tests/candle_lifecycle.rs

@@ -53,6 +53,8 @@ async fn test_candle_qwen3_load_unload_lifecycle() {
         quant: Some(quant),
         tensor_parallel: None,
         devices: Some(vec![0]),
+        sampling: None,
+        preserve_thinking: None,
     };
 
     registry

crates/neuron/tests/data/qwen3.8-27b.chat_template.jinja

@@ -0,0 +1,170 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- set reasoning_instructions = '' %}
+{%- if enable_thinking is undefined or enable_thinking is true %}
+    {%- set resolved_reasoning_effort = reasoning_effort|default('xhigh') %}
+    {%- if resolved_reasoning_effort not in ('xhigh', 'medium', 'low') %}
+        {{- raise_exception('Unexpected reasoning effort ' ~ reasoning_effort ~ '. Supported types are xhigh (default), medium, and low.') }}
+    {%- endif %}
+    {%- if resolved_reasoning_effort == 'xhigh' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer.' %}
+    {%- elif resolved_reasoning_effort == 'low' %}
+        {%- set reasoning_instructions = 'Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration.' %}
+    {%- endif %}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {%- if reasoning_instructions %}
+        {{- reasoning_instructions + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '<|im_start|>system\n' + (reasoning_instructions + '\n\n' if reasoning_instructions else '')  + content + '<|im_end|>\n' }}
+        {%- elif reasoning_instructions %}
+            {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+        {%- endif %}
+    {%- elif reasoning_instructions %}
+        {{- '<|im_start|>system\n' + reasoning_instructions + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if preserve_thinking is undefined or preserve_thinking is true or loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined and tool_call.arguments != '' %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

crates/neuron/tests/numerical_reference.rs

@@ -60,7 +60,9 @@ fn read_f32(path: &Path) -> Vec<f32> {
     let bytes = std::fs::read(path).unwrap_or_else(|e| panic!("read {path:?}: {e}"));
     assert!(bytes.len().is_multiple_of(4), "truncated f32 file {path:?}");
     bytes
-        .chunks_exact(4)
+        .as_chunks::<4>()
+        .0
+        .iter()
         .map(|b| f32::from_le_bytes([b[0], b[1], b[2], b[3]]))
         .collect()
 }

crates/neuron/tests/preflight.rs

@@ -106,6 +106,8 @@ fn spec(model_id: &str, tp: Option<u32>, quant: Option<&str>) -> ModelSpec {
         quant: quant.map(String::from),
         tensor_parallel: tp,
         devices: None,
+        sampling: None,
+        preserve_thinking: None,
     }
 }
 

crates/neuron/tests/qwen3_next_parity.rs

@@ -37,7 +37,9 @@ fn fixture_dir() -> PathBuf {
 fn read_f32(path: &Path) -> Vec<f32> {
     let bytes = std::fs::read(path).unwrap_or_else(|e| panic!("read {path:?}: {e}"));
     bytes
-        .chunks_exact(4)
+        .as_chunks::<4>()
+        .0
+        .iter()
         .map(|c| f32::from_le_bytes([c[0], c[1], c[2], c[3]]))
         .collect()
 }

models.example.toml

@@ -29,6 +29,27 @@
 #
 #                        Omit only when full precision is genuinely what
 #                        you want served.
+#   sampling           - operator sampling override (#283). Optional
+#                        sub-table `[models.sampling]` with any of
+#                        `temperature`, `top_p`, `top_k`.
+#
+#                        Precedence is request > operator > model's
+#                        generation_config.json > built-in fallback: this
+#                        replaces what the model published, but a caller
+#                        that names a value still wins.
+#
+#                        Use when a model's own published sampling is
+#                        wrong for the workload you serve. Qwen3.8-27B
+#                        publishes temperature = 1.0, which measured a
+#                        20% structural-defect rate on ~2k-token code
+#                        generation against 0/60 at <= 0.6 (Fisher exact
+#                        one-sided p = 0.0031, #283).
+#
+#                        Like `quant`, this must MATCH the same model's
+#                        `[default_models.sampling]` in the host's
+#                        neuron.toml — the same two-paths-to-one-model
+#                        hazard that made #252 possible.
+#                        `script/check-config-consistency.py` gates it.
 #   vram_mb            - rough estimate; advisory only, not enforced.
 #   min_devices        - GPU count this profile needs. TP profiles use
 #                        the same value as the tensor-parallel size.

models.toml

@@ -0,0 +1,219 @@
+# models.toml — helexa fleet catalogue
+#
+# Operator-owned and **tracked in git** (#283). It carries no secrets —
+# cortex.toml holds the API keys — and CI needs to see it: this file and
+# `asset/neuron/<host>.toml` both describe how a model is served, and
+# nothing forced them to agree until `script/check-config-consistency.py`
+# began gating on it. That gap is what #252 was.
+#
+# Deployed to the gateway by `.gitea/workflows/deploy.yml`, which is the
+# ONLY writer (operator decision 2026-08-25). Editing
+# /etc/cortex/models.toml on the host by hand is pointless — the next
+# deploy overwrites it from this file — and dangerous, because until it
+# does, the fleet is running something no commit describes.
+#
+# Fleet shape (sourced from asset/manifest.yml):
+#   beast      = 2x RTX 5090 (32 GB each)  → tp-2 capable
+#   benjy      = 1x RTX 4090 (24 GB)        → large single-GPU
+#   quadbrat   = 1x RTX 3060 (12 GB)        → small / quantised
+
+# Qwen3-Next single-GPU validation target (Stage 8c). Same architecture
+# as Qwen3.6-27B (`model_type = "qwen3_5"`, same layer_types pattern,
+# same Gated DeltaNet hyperparams) but small enough to fit on any
+# fleet GPU in bf16 (~1.6 GB). Stepping stone — if this answers the
+# Tbilisi probe correctly, the qwen3_5 code path is right and TP-2 on
+# Qwen3.6-27B is purely a sharding/collective problem.
+[[models]]
+id = "Qwen/Qwen3.5-0.8B"
+harness = "candle"
+vram_mb = 2000
+min_devices = 1
+min_device_vram_mb = 4000
+# Illustrative pricing (USD per 1M tokens). helexa has no paying customers
+# yet; these differ per model purely to validate spend surfaces in clients.
+cost.input = 0.02
+cost.output = 0.08
+
+# Frontier MoE target (epic #84, F2 #93): 80B total / 3B active,
+# 512 experts / 10 active + shared expert, same Gated DeltaNet hybrid
+# as the 27B. Weights ~45 GB @ Q4_K_M ISQ under TP-2 (~22.5 GB/GPU) —
+# beast only, and NOT co-resident with the 27B (cold-swap per F4e #99).
+# The serving quant belongs HERE as well as on any load request (#252):
+# a cortex cold-load builds its spec from this file, and without it the
+# model loads unquantised — ~149 GB, which cannot fit beast at all.
+# tensor_parallel is derived from min_devices, so only quant is needed.
+[[models]]
+id = "Qwen/Qwen3-Coder-Next"
+harness = "candle"
+quant = "q4k"
+vram_mb = 46000
+min_devices = 2
+min_device_vram_mb = 24000
+# Shares the big-node class with the flagship: either may cold-swap the
+# other off beast (#99), in both directions, so a comparison between
+# them can move traffic each way on demand.
+residency_priority = 300
+# Illustrative pricing (USD per 1M tokens) — frontier tier.
+cost.input = 0.40
+cost.output = 1.60
+
+# Thinking sibling of Coder-Next (same 80B-A3B arch) — on beast for the
+# F3 A/B evaluation (#94); one of the two frontier variants competes for
+# residency, the other stays cold-loadable.
+[[models]]
+id = "Qwen/Qwen3-Next-80B-A3B-Thinking"
+harness = "candle"
+quant = "q4k"
+vram_mb = 46000
+min_devices = 2
+min_device_vram_mb = 24000
+# Shares the big-node class with the flagship: either may cold-swap the
+# other off beast (#99), in both directions, so a comparison between
+# them can move traffic each way on demand.
+residency_priority = 300
+# Illustrative pricing (USD per 1M tokens) — frontier tier.
+cost.input = 0.40
+cost.output = 1.60
+
+# Tensor-parallel target. Beast is the only neuron with two big GPUs.
+[[models]]
+id = "Qwen/Qwen3.6-27B"
+harness = "candle"
+# REQUIRED, not cosmetic (#252). Without it a cortex cold-load builds an
+# unquantised request: the model loads bf16, leaves ~2 GB free per card
+# against a ~3.9 GB prefill requirement, reports `loaded`, and then fails
+# every request. Activation from neuron's default_models has always said
+# q6k; this is the same fact on the path cortex uses.
+quant = "q6k"
+# Measured q6k residency across the pair (17.9 + 22.9 GB), not the 51.7 GB
+# bf16 checkpoint size. The old 54000 described a model we do not serve.
+vram_mb = 42000
+min_devices = 2
+min_device_vram_mb = 24000
+pinned_on = ["beast"]
+# The big-node class. Above the everyday class, so image generation can
+# never take beast however idle the flagship is; equal to the frontier
+# tier, so those two swap freely. `pinned_on` here is affinity only --
+# it no longer implies immunity, so this number is what protects it.
+residency_priority = 300
+# Illustrative pricing (USD per 1M tokens) — top tier, the 27B flagship.
+cost.input = 0.30
+cost.output = 1.20
+# DELIBERATE, TEMPORARY operator override (issue #67): unblocks the opencode
+# helexa-provider implementation by giving /v1/models a limit to consume on
+# beast. Mirrors the enforced NEURON_MAX_PROMPT_TOKENS=131072 there.
+# To be REMOVED once neuron self-derives these — its removal is the
+# acceptance test that the computed values replace this hand-set one.
+#limit.context = 131072
+#limit.input = 122880
+#limit.output = 8192
+
+# Qwen3.8-27B (#251). Architecturally identical to the 3.6-27B — same
+# model_type, same layer counts, same vision tower; the two configs
+# differ only in `transformers_version` — so it serves through the same
+# qwen3_5 TP path with the same KV arithmetic.
+#
+# Same residency class as the 3.6-27B *deliberately*: equal rank is what
+# lets the two displace each other in BOTH directions, so whichever is
+# asked for takes beast and the other can take it back. A higher number
+# here would let this one claim the node permanently and make the
+# comparison between them impossible.
+[[models]]
+id = "Qwen/Qwen3.8-27B"
+harness = "candle"
+quant = "q6k"
+vram_mb = 42000
+min_devices = 2
+min_device_vram_mb = 24000
+pinned_on = ["beast"]
+residency_priority = 300
+cost.input = 0.30
+cost.output = 1.20
+
+# Sampling override (#283). The model publishes `temperature = 1.0` in
+# its generation_config.json, and honouring that (#272) is correct by
+# default — but measured on this stack it produced a 20% structural-
+# defect rate on ~2k-token code generation (4/20 samples failed
+# `node --check` or dropped required exports) against 0/60 at
+# temperature <= 0.6. Fisher exact one-sided p = 0.0031.
+#
+# Extrapolated at constant hazard that is a ~79% chance of at least one
+# defect in a 15k-token file, which is why this model had never once
+# completed the tower-defence task two cloud models did in ten minutes.
+#
+# 0.6 is Qwen's own published *inference* guidance for thinking mode —
+# a number that lives in a model card no machine can read, which is
+# exactly why it has to be set here by hand.
+#
+# Precedence is request > operator > model, so a caller that names a
+# temperature still gets it. MUST match asset/neuron/beast.toml;
+# script/check-config-consistency.py fails CI if it drifts.
+[models.sampling]
+temperature = 0.6
+
+# Mid-size dense bf16 — fits on benjy (24 GB) or beast (32 GB ea.).
+[[models]]
+id = "Qwen/Qwen3-8B"
+harness = "candle"
+vram_mb = 18000
+min_devices = 1
+min_device_vram_mb = 16000
+# Everyday class, shared with Z-Image — they take turns on benjy.
+residency_priority = 200
+# Illustrative pricing (USD per 1M tokens) — mid tier (helexa/balanced).
+cost.input = 0.10
+cost.output = 0.40
+
+# Single-GPU dense bf16 — fits anywhere with ≥4 GB.
+[[models]]
+id = "Qwen/Qwen3-1.7B"
+harness = "candle"
+vram_mb = 4000
+min_devices = 1
+min_device_vram_mb = 4000
+# Illustrative pricing (USD per 1M tokens) — small tier (helexa/small).
+cost.input = 0.04
+cost.output = 0.16
+
+# Small GGUF quantised — runs on quadbrat (12 GB) without breaking a sweat.
+[[models]]
+id = "unsloth/Qwen3-0.6B-GGUF"
+harness = "candle"
+quant = "Q4_K_M"
+vram_mb = 500
+min_devices = 1
+min_device_vram_mb = 4000
+# Illustrative pricing (USD per 1M tokens) — cheapest, the tiny GGUF.
+cost.input = 0.015
+cost.output = 0.06
+
+# -- Tier aliases ------------------------------------------------------------
+# Clients can request inference against an alias and cortex transparently
+# routes to the concrete model id below. Useful so client code can pick a
+# latency/quality tier without hardcoding specific model ids — operators
+# can swap the targets here without touching client config.
+[aliases]
+"helexa/small" = "Qwen/Qwen3-1.7B"
+"helexa/balanced" = "Qwen/Qwen3-8B"
+"helexa/large" = "Qwen/Qwen3.8-27B"
+"helexa/image" = "Tongyi-MAI/Z-Image-Turbo"
+
+# Text-to-image (#197/#203): Z-Image-Turbo, 6B S3-DiT + Qwen3-4B text
+# encoder (runs on host CPU by default — see [harness.candle.image] in
+# neuron.toml) + VAE. BF16 DiT+VAE resident ≈ 14.6 GB; fits benjy's
+# 4090 via cold-swap with Qwen3-8B (operator decision 2026-07-29,
+# #197). Not default-activated anywhere — cortex cold-loads it on
+# first images request and the evictor swaps it back out under text
+# pressure. Metered in megapixel-steps (#202), not tokens; the token
+# cost fields deliberately stay unset.
+[[models]]
+id = "Tongyi-MAI/Z-Image-Turbo"
+harness = "candle"
+vram_mb = 15000
+min_devices = 1
+min_device_vram_mb = 20000
+# Everyday class, shared with Qwen3-8B: an image request cold-swaps the
+# 8B off benjy and the next text request swaps it straight back. Below
+# the big-node class, so it can never take beast -- its device
+# constraints alone would otherwise let it land there.
+residency_priority = 200

neuron.example.toml

@@ -110,6 +110,33 @@ name = "candle"
 # harness = "candle"
 # quant = "Q4_K_M"
 # devices = [0]
+#
+# Operator sampling override (#283) — optional, per model.
+#
+# By default neuron honours what the model published in its
+# `generation_config.json`; that is the model authors' stated intent and
+# is right most of the time. It is not always right for the workload you
+# serve: Qwen3.8-27B publishes `temperature = 1.0`, which measured a 20%
+# structural-defect rate on ~2k-token code generation against 0/60 at
+# <= 0.6 (Fisher exact one-sided p = 0.0031).
+#
+# Precedence is request > operator > model > built-in fallback, so a
+# caller that names a temperature still gets what it asked for; this only
+# changes what a caller who names nothing receives. Each field overlays
+# independently — setting `temperature` alone leaves the model's `top_p`
+# and `top_k` in force.
+#
+# Deliberately explicit. neuron will never inspect a model and guess a
+# "better" temperature: a guesser is a footgun, a number you wrote is not.
+#
+# If cortex can also cold-load this model, the same values must appear in
+# its `models.toml` entry — `script/check-config-consistency.py` gates on
+# that, because two files describing one model is how #252 happened.
+#
+# [default_models.sampling]
+# temperature = 0.6
+# top_p = 0.95
+# top_k = 20
 
 # Text-to-image serving (#198). Applies to diffusers-format image
 # models (Z-Image family). All fields optional.

script/check-config-consistency.py

@@ -0,0 +1,185 @@
+#!/usr/bin/env python3
+"""Fail when the two files describing one model disagree.
+
+`models.toml` (cortex's catalogue, used for cold loads) and
+`asset/neuron/<host>.toml` (`[[default_models]]`, used for models a host
+holds resident) both describe *how a model is served*. Nothing forced
+them to agree, and #252 is what that costs: models.toml omitted `quant`,
+so a cortex cold-load served bf16 and could not prefill, while the same
+model was fine when the host loaded it from its own config.
+
+#283 adds a second such field — the operator sampling override — so this
+check exists before the second instance of the bug rather than after it.
+
+Checked, per model id present in both files:
+
+  * ``quant``      — the #252 field
+  * ``sampling``   — the #283 field, compared field-by-field
+  * ``preserve_thinking`` — the H7 A/B gate: a model that replays
+    prior reasoning when resident and not when cold-loaded would
+    make the A/B unreadable, which is #252's shape again
+
+A model in only one file is not an error: plenty of catalogue entries are
+never resident anywhere, and a host may hold a model the catalogue does
+not list. Only *disagreement* fails.
+
+Usage::
+
+    check-config-consistency.py                       # repo defaults
+    check-config-consistency.py --models /etc/cortex/models.toml \\
+                               --neuron asset/neuron/beast.toml
+
+Exit status is 0 when consistent, 1 on any disagreement, 2 on bad input.
+"""
+
+from __future__ import annotations
+
+import argparse
+import glob
+import os
+import sys
+
+try:
+    import tomllib
+except ModuleNotFoundError:  # pragma: no cover - Python < 3.11
+    print("error: needs Python 3.11+ for tomllib", file=sys.stderr)
+    raise SystemExit(2)
+
+REPO = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+# Fields that must agree wherever both files mention the same model.
+# Extend this list whenever a serving-critical field lands in both.
+SCALAR_FIELDS = ["quant", "preserve_thinking"]
+TABLE_FIELDS = ["sampling"]
+
+
+def load(path: str) -> dict:
+    try:
+        with open(path, "rb") as fh:
+            return tomllib.load(fh)
+    except FileNotFoundError:
+        print(f"error: {path} not found", file=sys.stderr)
+        raise SystemExit(2)
+    except tomllib.TOMLDecodeError as exc:
+        print(f"error: {path} is not valid TOML: {exc}", file=sys.stderr)
+        raise SystemExit(2)
+
+
+def catalogue_entries(path: str) -> dict[str, dict]:
+    """model id -> profile, from models.toml's `[[models]]`."""
+    return {m["id"]: m for m in load(path).get("models", []) if "id" in m}
+
+
+def resident_entries(path: str) -> dict[str, dict]:
+    """model id -> spec, from a neuron config's `[[default_models]]`."""
+    return {
+        m["model_id"]: m
+        for m in load(path).get("default_models", [])
+        if "model_id" in m
+    }
+
+
+def label(path: str) -> str:
+    """Repo-relative when the file is in the repo, basename otherwise.
+
+    Deploy runs pass paths outside the tree (a models.toml fetched off
+    the gateway), and `../../../..`-prefixed relpaths make the diff
+    unreadable in a CI log.
+    """
+    rel = os.path.relpath(path, REPO)
+    return rel if not rel.startswith("..") else os.path.basename(path)
+
+
+def compare(model_id: str, cat: dict, res: dict, catalogue_path: str, neuron_path: str):
+    """Yield one human-readable line per disagreement."""
+    cat_name = label(catalogue_path)
+    neu_name = label(neuron_path)
+
+    for field in SCALAR_FIELDS:
+        a, b = cat.get(field), res.get(field)
+        # Absent on one side is "unspecified", not "different" -- but for
+        # quant specifically that IS the #252 bug, so say so loudly.
+        if a != b:
+            if a is None or b is None:
+                detail = (
+                    "  (one side leaves it unset -- this is exactly #252: "
+                    "an unset quant makes a cold load serve bf16)"
+                    if field == "quant"
+                    else "  (one side leaves it unset)"
+                )
+            else:
+                detail = ""
+            yield (
+                f"{model_id}: {field} differs\n"
+                f"    {cat_name}: {a!r}\n"
+                f"    {neu_name}: {b!r}{detail}"
+            )
+
+    for field in TABLE_FIELDS:
+        a = cat.get(field) or {}
+        b = res.get(field) or {}
+        if a == b:
+            continue
+        keys = sorted(set(a) | set(b))
+        rows = "\n".join(
+            f"      {k}: {cat_name}={a.get(k)!r}  {neu_name}={b.get(k)!r}"
+            for k in keys
+            if a.get(k) != b.get(k)
+        )
+        yield (
+            f"{model_id}: {field} differs\n"
+            f"    a model must sample the same way whether cortex cold-loads it\n"
+            f"    or the host holds it resident:\n{rows}"
+        )
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--models", default=os.path.join(REPO, "models.toml"),
+                    help="path to models.toml (default: repo root)")
+    ap.add_argument("--neuron", nargs="*", default=None,
+                    help="neuron config(s) (default: asset/neuron/*.toml)")
+    args = ap.parse_args()
+
+    neuron_paths = args.neuron
+    if not neuron_paths:
+        neuron_paths = sorted(glob.glob(os.path.join(REPO, "asset", "neuron", "*.toml")))
+    if not neuron_paths:
+        print("error: no neuron configs found", file=sys.stderr)
+        return 2
+
+    catalogue = catalogue_entries(args.models)
+    problems: list[str] = []
+    compared = 0
+
+    for npath in neuron_paths:
+        for model_id, spec in resident_entries(npath).items():
+            profile = catalogue.get(model_id)
+            if profile is None:
+                # A host may hold a model the catalogue does not list.
+                # Not an error, but worth seeing -- cortex cannot route
+                # to what it has no profile for.
+                print(f"note: {model_id} is resident in {label(npath)} "
+                      f"but absent from {label(args.models)}")
+                continue
+            compared += 1
+            problems.extend(compare(model_id, profile, spec, args.models, npath))
+
+    if problems:
+        print(f"\nconfig inconsistency: {len(problems)} disagreement(s)\n",
+              file=sys.stderr)
+        for p in problems:
+            print(f"  {p}\n", file=sys.stderr)
+        print("Two config files describe one model. Make them agree, or the\n"
+              "model is served one way on a cold load and another when\n"
+              "resident -- see #252 and #283.", file=sys.stderr)
+        return 1
+
+    print(f"config consistency OK: {compared} model/host pairing(s) agree "
+          f"on {', '.join(SCALAR_FIELDS + TABLE_FIELDS)}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

script/cuda-check.sh

@@ -0,0 +1,83 @@
+#!/usr/bin/env bash
+# Type-check neuron's `cuda` feature locally, in CI's own image.
+#
+# Why this exists: `#[cfg(feature = "cuda")]` code is invisible to a
+# default-feature build, so `cargo build`, `cargo clippy --all-targets`
+# and the whole test suite can pass on code that does not compile for
+# the target the fleet actually serves on. The only gate that catches it
+# is CI's `cuda-check` job — a ~20 minute round trip, and the failure
+# arrives after the push rather than before it.
+#
+# This runs the same command, with the same env, in the same image.
+#
+# Do NOT try to install the CUDA toolkit on a Fedora 44 host instead:
+# CUDA 13.0's nvcc rejects host compilers newer than gcc 15 and F44
+# ships gcc 16 with no compat package. The image is Fedora 43 for
+# exactly that reason, and additionally patches a glibc 2.41+ /
+# CUDA-header `noexcept` mismatch that nvcc hard-errors on. See
+# gongfoo's images/runner-cuda-13.0/Containerfile.
+#
+# No GPU is required — this is a borrow-check, not a run.
+#
+# First run pulls the image (~GBs) and compiles the candle/flash-attn
+# kernels, so it costs about what CI costs. Afterwards the cargo target
+# volume makes an incremental check a fraction of that, which is the
+# whole point.
+set -euo pipefail
+
+IMAGE="${CUDA_CHECK_IMAGE:-git.lair.cafe/gongfoo/runner-cuda-13.0:latest}"
+ROOT="$(git rev-parse --show-toplevel)"
+
+# A dedicated target volume. Sharing the host's `target/` would make the
+# default-feature and cuda-feature builds evict each other's artifacts on
+# every alternation, which would cost more than it saves.
+TARGET_VOL="${CUDA_CHECK_TARGET_VOL:-helexa-cuda-check-target}"
+# Mounted at the registry rather than over CARGO_HOME: the toolchain
+# binaries live in /usr/local/cargo/bin and must stay visible.
+REGISTRY_VOL="${CUDA_CHECK_REGISTRY_VOL:-helexa-cuda-check-registry}"
+
+if [ "${1:-}" = "--pull" ]; then
+    shift
+    podman pull "${IMAGE}"
+fi
+
+# Default to the exact command ci.yml's cuda-check job runs. Extra
+# arguments replace it, for narrowing to one crate while iterating.
+if [ "$#" -gt 0 ]; then
+    CARGO_CMD=("$@")
+else
+    CARGO_CMD=(cargo check -p neuron --features cuda,flash-attn --all-targets)
+fi
+
+# Allocate a TTY only when there is one to allocate. `-it` against a
+# pipe or a CI log wedges the container before cargo ever starts, which
+# looks exactly like a very slow compile.
+TTY_ARGS=()
+if [ -t 0 ] && [ -t 1 ]; then
+    TTY_ARGS=(-it)
+fi
+
+exec podman run --rm "${TTY_ARGS[@]}" \
+    -v "${ROOT}:/work:Z" \
+    -v "${TARGET_VOL}:/target" \
+    -v "${REGISTRY_VOL}:/usr/local/cargo/registry" \
+    -w /work \
+    -e CARGO_TARGET_DIR=/target \
+    -e CARGO_TERM_COLOR=always \
+    `# candle-kernels' build script falls back to nvidia-smi for` \
+    `# compute-cap detection when this is unset, and there is no GPU` \
+    `# here. Any valid cap borrow-checks; the real per-flavour caps` \
+    `# live in build-prerelease.yml's matrix.` \
+    -e CUDA_COMPUTE_CAP=86 \
+    `# The image ships sccache, but the local run has no shared cache` \
+    `# to talk to; an unset wrapper avoids a hard cargo failure.` \
+    -e RUSTC_WRAPPER= \
+    "${IMAGE}" \
+    bash -lc '
+        set -euo pipefail
+        # CUDA_HOME/PATH/LD_LIBRARY_PATH come from the image. LIBRARY_PATH
+        # does not, and cudarc needs it at link time — ci.yml exports all
+        # three explicitly, so keep this in step with that job.
+        export LIBRARY_PATH="${CUDA_HOME}/targets/x86_64-linux/lib:${CUDA_HOME}/lib64:${LIBRARY_PATH:-}"
+        exec "$@"
+    ' _ "${CARGO_CMD[@]}"

script/infra-setup.sh

@@ -138,14 +138,28 @@ if ! ssh "${bench_host}" '
     echo "  failed to prepare ${bench_host} for helexa-bench"
 fi
 
-# Push application config to the fleet. The deploy workflow is
-# scoped to package install + service restart; config changes ride
-# along with this script instead, since:
-#   - cortex.toml and models.toml are gitignored (operator-owned, may
-#     include secrets), so CI never sees them
-#   - asset/neuron/<short>.toml is tracked but iterating locally is
-#     faster than pushing a commit and waiting for build-prerelease
-#     to roll over
+# Push the one config CI cannot own.
+#
+# Operator decision 2026-08-25 (#283): the automated deployment path is
+# the ONLY path. `.gitea/workflows/deploy.yml` now ships models.toml to
+# the gateway, asset/neuron/<short>.toml to each neuron, and
+# asset/helexa-bench/<host>.toml to the bench host — digest-compared,
+# restarting the service when the file changes.
+#
+# Those three syncs USED to live here, and that is exactly how the fleet
+# drifted: benjy and quadbrat were still running pre-#197 configs months
+# after the tracked assets had moved on, because deploying them meant
+# remembering to re-run this script. A hand-sync that exists "for local
+# iteration" is a footgun that silently overwrites, or silently fails to
+# overwrite, whatever CI just placed. Do not add one back.
+#
+# What remains is cortex.toml, which carries the inference API keys
+# ([[entitlements.keys]]) and therefore cannot be tracked in git for CI
+# to read. It is the last hand-synced config on the fleet and the last
+# place this drift can still happen; closing it needs the secrets to
+# come from Gitea secrets at deploy time rather than from a working
+# copy.
+#
 # Missing source files are skipped silently — re-run after editing.
 # Optional 4th/5th args set the destination group and mode. Configs that
 # carry secrets (API keys, JWT signing keys, SMTP/DB passwords) must land
@@ -174,23 +188,15 @@ sync_config() {
     fi
 }
 
-echo "==> ${cortex_host}: syncing gateway configs"
-# cortex.toml holds the inference API keys ([[entitlements.keys]]).
+echo "==> ${cortex_host}: syncing gateway secrets config"
+# cortex.toml holds the inference API keys ([[entitlements.keys]]), so it
+# is gitignored and CI cannot see it. The only config still synced by
+# hand — everything else is deploy.yml's.
 sync_config "${cortex_host}" "${repo_path}/cortex.toml" /etc/cortex/cortex.toml cortex 0640
-sync_config "${cortex_host}" "${repo_path}/models.toml" /etc/cortex/models.toml
 
-for neuron_host in "${neuron_hosts[@]}"; do
-    short="${neuron_host%%.*}"
-    echo "==> ${neuron_host}: syncing per-host neuron config"
-    sync_config "${neuron_host}" \
-        "${repo_path}/asset/neuron/${short}.toml" \
-        /etc/neuron/neuron.toml
-done
-
-echo "==> ${bench_host}: syncing bench config"
-sync_config "${bench_host}" \
-    "${repo_path}/asset/helexa-bench/${bench_host%%.*}.toml" \
-    /etc/helexa-bench/helexa-bench.toml
+# models.toml, asset/neuron/<short>.toml and asset/helexa-bench/<host>.toml
+# are deliberately NOT synced here. deploy.yml owns them; see the note
+# above sync_config().
 
 # ── bench UI: public nginx vhost on the gateway (bench.helexa.ai) ──────
 # The built SPA is rsynced to the webroot by deploy.yml; nginx serves it
@@ -600,6 +606,24 @@ if ! ssh "${svc_host}" '
     echo "  failed to prepare ${svc_host}"
 fi
 
+# ── STILL HAND-SYNCED, and tracked as such ────────────────────────────
+# Everything below this line is a config CI does not yet own, which by
+# the 2026-08-25 decision (#283) makes it a footgun still loaded. Two
+# distinct reasons, needing two distinct fixes:
+#
+#   * helexa-router.toml / helexa-upstream.toml / helexa-angels.toml are
+#     gitignored because they carry JWT secrets, SMTP passwords and
+#     database URLs. Same class as cortex.toml: CI cannot read them until
+#     the secrets come from Gitea secrets and the file is rendered at
+#     deploy time.
+#   * asset/searxng/{settings.yml,searxng.container,helexa-searxng.xml}
+#     ARE tracked and could move to deploy.yml today but for one thing:
+#     asset/sudoers.d/gallumbits-host.conf grants gitea_ci no rsync at
+#     all, so the deploy would fail until the grants land on the host.
+#     Grants first, deploy step second — wiring it the other way round
+#     breaks every deploy in between.
+#
+# See the tracking issue; do not close the gap by adding syncs here.
 echo "==> ${svc_host}: syncing service configs"
 sync_config "${svc_host}" "${repo_path}/helexa-router.toml" /etc/helexa-router/helexa-router.toml
 # helexa-upstream.toml holds the session JWT secret, the SMTP submission