bump version to v4.0.2 (#93 )

Co-authored-by: n13 <n13@users.noreply.github.com>
Different kernels for Apple Metal vs Nvidia and everything else (#92 )
2026-09-01 23:25:24 +08:00 · 2026-09-01 23:21:43 +08:00 · 2026-09-01 19:03:54 +08:00 · 2026-09-01 19:03:05 +08:00 · 2026-08-31 22:50:27 +08:00 · 2026-08-28 14:11:24 +08:00
62 changed files with 8104 additions and 6231 deletions
@@ -1,42 +0,0 @@
-# Git
-.git
-.gitignore
-.github
-
-# Build artifacts
-target/
-*.rs.bk
-*.pdb
-
-# IDE
-.vscode
-.idea
-*.swp
-*.swo
-*~
-
-# Documentation
-docs/
-*.md
-!README.md
-
-# Examples and tests (not needed for runtime)
-examples/
-
-# Other
-.DS_Store
-Thumbs.db
-
-# CI/CD
-.github/
-
-# License
-LICENSE
-
-# Dockerfiles themselves
-Dockerfile*
-.dockerignore
-
-# Cargo cache
-.cargo/
-
@@ -45,28 +45,17 @@ jobs:
    name: 🛠️ Build & Test
    needs: fast-checks
    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        features: ["", "gpu"]
    steps:
      - uses: actions/checkout@v4
      - name: setup rust
        uses: actions-rust-lang/setup-rust-toolchain@v1
-      - name: compile (${{ matrix.features || 'cpu-only' }})
+      - name: compile (gpu)
        run: |
-          if [ -n "${{ matrix.features }}" ]; then
-            cargo build --locked --workspace --features ${{ matrix.features }}
-          else
-            cargo build --locked --workspace
-          fi
+          cargo build --locked --workspace
        timeout-minutes: 90
-      - name: test (${{ matrix.features || 'cpu-only' }})
+      - name: test (gpu)
        run: |
-          if [ -n "${{ matrix.features }}" ]; then
-            cargo test --locked --workspace --features ${{ matrix.features }}
-          else
-            cargo test --locked --workspace
-          fi
+          cargo test --locked --workspace
        timeout-minutes: 15

  analysis:
@@ -95,8 +84,13 @@ jobs:
      - name: setup rust
        uses: actions-rust-lang/setup-rust-toolchain@v1
      - name: build benchmark binary
-        run: cargo build -p miner-cli --release --features gpu
+        run: cargo build -p miner-cli --release
        timeout-minutes: 60
+      - name: upload linux release binary
+        uses: actions/upload-artifact@v4
+        with:
+          name: quantus-miner-linux-x86_64
+          path: target/release/quantus-miner
      - name: run cpu benchmark
        run: ./target/release/quantus-miner benchmark --cpu-workers 2 --duration 5
        timeout-minutes: 10
@@ -0,0 +1,13 @@
+name: Dependency cooldown audit
+
+on:
+  schedule:
+    - cron: '0 6 * * 1'
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  audit:
+    uses: Quantus-Network/shared-workflows/.github/workflows/dependency-cooldown-audit.yml@v1
@@ -0,0 +1,14 @@
+name: Dependency cooldown
+
+on:
+  pull_request:
+    # labeled/unlabeled/edited are required so the gate re-runs when an
+    # emergency bypass label or reason is added.
+    types: [opened, synchronize, reopened, labeled, unlabeled, edited]
+
+permissions:
+  contents: read
+
+jobs:
+  dependency-cooldown:
+    uses: Quantus-Network/shared-workflows/.github/workflows/dependency-cooldown.yml@v1
@@ -1,145 +0,0 @@
-name: Docker Image
-
-on:
-  workflow_dispatch:
-    inputs:
-      specific_version:
-        description: "Optional: Specify a full version (e.g., v0.3.0 or v0.3.1-beta.1) to build. If empty, uses the latest release tag. MUST start with 'v'."
-        required: false
-        default: ""
-
-env:
-  CARGO_TERM_COLOR: always
-
-jobs:
-  determine_version:
-    runs-on: ubuntu-latest
-    outputs:
-      version_with_v: ${{ steps.version_info.outputs.version_with_v }}
-      is_valid_version: ${{ steps.version_info.outputs.is_valid_version }}
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
-      - name: Get version information
-        id: version_info
-        run: |
-          set -ex
-          TARGET_VERSION=""
-          SPECIFIC_VERSION="${{ github.event.inputs.specific_version }}"
-
-          if [[ -n "$SPECIFIC_VERSION" ]]; then
-            if [[ ! "$SPECIFIC_VERSION" =~ ^v[0-9]+\.[0-9]+\.[0-9]+(-[a-zA-Z0-9.-]+)?$ ]]; then
-              echo "::error::Specified version '$SPECIFIC_VERSION' is not a valid format. It must start with 'v' (e.g., v0.3.0 or v0.3.1-beta.1)."
-              echo "is_valid_version=false" >> $GITHUB_OUTPUT
-              exit 1
-            fi
-            TARGET_VERSION="$SPECIFIC_VERSION"
-            echo "Using specified version: $TARGET_VERSION"
-          else
-            echo "No specific version provided, determining latest release tag..."
-            TARGET_VERSION=$(git tag --list 'v*' --sort=-v:refname | head -n 1)
-            if [[ -z "$TARGET_VERSION" ]]; then
-              echo "::error::No version tags starting with 'v' (e.g., vX.Y.Z) found in the repository."
-              echo "is_valid_version=false" >> $GITHUB_OUTPUT
-              exit 1
-            fi
-            echo "Latest release version found: $TARGET_VERSION"
-          fi
-
-          echo "version_with_v=$TARGET_VERSION" >> $GITHUB_OUTPUT
-          echo "is_valid_version=true" >> $GITHUB_OUTPUT
-
-  build_and_publish_image:
-    needs: determine_version
-    if: needs.determine_version.outputs.is_valid_version == 'true'
-    runs-on: ubuntu-latest
-    permissions:
-      contents: read
-      packages: write
-    env:
-      GHCR_IMAGE_PATH: ghcr.io/quantus-network/quantus-miner
-      TARGET_VERSION_WITH_V: ${{ needs.determine_version.outputs.version_with_v }}
-
-    steps:
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
-    
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Log in to GitHub Container Registry
-        uses: docker/login-action@v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Check if Docker image already exists and fail if so
-        id: check_image
-        run: |
-          set -ex
-          IMAGE_TO_CHECK="${{ env.GHCR_IMAGE_PATH }}:${{ env.TARGET_VERSION_WITH_V }}"
-          echo "Checking for image: $IMAGE_TO_CHECK"
-          
-          if docker manifest inspect "$IMAGE_TO_CHECK" > /dev/null 2>&1; then
-            echo "::error::Image $IMAGE_TO_CHECK already exists. Aborting."
-            exit 1
-          else
-            echo "Image $IMAGE_TO_CHECK does not exist. Proceeding with build."
-          fi
-
-      - name: Checkout code at tag version
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ env.TARGET_VERSION_WITH_V }}
-
-      - name: Check if Dockerfile exists in tag
-        id: check_dockerfile
-        run: |
-          if [ ! -f Dockerfile ]; then
-            echo "::warning::Dockerfile not found in tag ${{ env.TARGET_VERSION_WITH_V }}, falling back to main branch Dockerfile"
-            echo "use_main=true" >> $GITHUB_OUTPUT
-          else
-            echo "Dockerfile found in tag"
-            echo "use_main=false" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Checkout main branch for Dockerfile (if needed)
-        if: steps.check_dockerfile.outputs.use_main == 'true'
-        uses: actions/checkout@v4
-        with:
-          ref: main
-          sparse-checkout: |
-            Dockerfile
-            .dockerignore
-          sparse-checkout-cone-mode: false
-          path: dockerfile-source
-
-      - name: Copy Dockerfile from main (if needed)
-        if: steps.check_dockerfile.outputs.use_main == 'true'
-        run: |
-          cp dockerfile-source/Dockerfile ./
-          [ -f dockerfile-source/.dockerignore ] && cp dockerfile-source/.dockerignore ./ || true
-
-      - name: Build and push Docker image
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          file: ./Dockerfile
-          push: true
-          platforms: linux/amd64,linux/arm64
-          provenance: false
-          tags: |
-            ${{ env.GHCR_IMAGE_PATH }}:${{ env.TARGET_VERSION_WITH_V }}
-            ${{ env.GHCR_IMAGE_PATH }}:latest
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
-
-      - name: Print completion message
-        run: |
-          echo "Successfully built and pushed ${{ env.GHCR_IMAGE_PATH }}:${{ env.TARGET_VERSION_WITH_V }}"
-          echo "Image is also tagged as latest: ${{ env.GHCR_IMAGE_PATH }}:latest"
-
@@ -71,45 +71,21 @@ jobs:
            target: x86_64-unknown-linux-gnu
            binary_name: quantus-miner
            asset_name: quantus-miner-linux-x86_64
-            features: ""
-          - os: ubuntu-latest
-            target: x86_64-unknown-linux-gnu
-            binary_name: quantus-miner
-            asset_name: quantus-miner-linux-x86_64-gpu
-            features: "gpu"
          # Windows builds
          - os: windows-latest
            target: x86_64-pc-windows-msvc
            binary_name: quantus-miner.exe
            asset_name: quantus-miner-windows-x86_64.exe
-            features: ""
-          - os: windows-latest
-            target: x86_64-pc-windows-msvc
-            binary_name: quantus-miner.exe
-            asset_name: quantus-miner-windows-x86_64-gpu.exe
-            features: "gpu"
          # macOS builds (Intel)
-          - os: macos-13
+          - os: macos-15-intel
            target: x86_64-apple-darwin
            binary_name: quantus-miner
            asset_name: quantus-miner-macos-x86_64
-            features: ""
-          - os: macos-13
-            target: x86_64-apple-darwin
-            binary_name: quantus-miner
-            asset_name: quantus-miner-macos-x86_64-gpu
-            features: "gpu"
          # macOS builds (Apple Silicon)
          - os: macos-latest
            target: aarch64-apple-darwin
            binary_name: quantus-miner
            asset_name: quantus-miner-macos-aarch64
-            features: ""
-          - os: macos-latest
-            target: aarch64-apple-darwin
-            binary_name: quantus-miner
-            asset_name: quantus-miner-macos-aarch64-gpu
-            features: "gpu"

    steps:
      - uses: actions/checkout@v4
@@ -122,6 +98,7 @@ jobs:
          target: ${{ matrix.target }}

      - name: build binary
+        shell: bash
        run: |
          if [[ -n "${{ matrix.features }}" ]]; then
            cargo build --release --locked --target ${{ matrix.target }} --features ${{ matrix.features }}
@@ -6,7 +6,10 @@ members = [
 	"crates/miner-cli",
 	"crates/miner-service",
 	"crates/miner-telemetry",
+	"crates/pool-service",
 	"crates/pow-core",
+	"crates/quic-transport",
+	"crates/solver-wasm",
 ]
 default-members = ["crates/miner-cli"]
 resolver = "2"
@@ -15,7 +18,7 @@ resolver = "2"
 edition = "2021"
 authors = ["Quantus Network"]
 description = "Quantus External Miner Workspace"
-version = "2.0.0"
+version = "4.0.2"

 [workspace.dependencies]
 anyhow = "1"
@@ -29,18 +32,15 @@ num-bigint = { version = "0.4", features = ["rand"] }
 num-traits = "0.2"
 num_cpus = "1.16"
 primitive-types = { version = "0.13.1", default-features = false }
-qp-poseidon-core = { version = "1.0.1", default-features = false }
-qpow-math = { git = "https://github.com/Quantus-Network/chain.git", package = "qpow-math", default-features = false }
-quantus-miner-api = { version = "0.0.3", default-features = false }
-rand = { version = "0.8.5", default-features = false }
+qp-poseidon-core = { version = "3.1.0", default-features = false }
+qpow-math = { git = "https://github.com/Quantus-Network/chain.git", tag = "v0.7.1-q-day-2", package = "qpow-math", default-features = false }
+quantus-miner-api = "0.3.0"
+rand = { version = "0.9", default-features = false, features = ["std", "std_rng", "thread_rng"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = { version = "1.0.132", default-features = false }
 thiserror = "1"
 tokio = { version = "1.36", features = ["full"] }
 warp = "0.3"
-# pow-core now delegates to qpow-math for core functionality while providing optimized mining scaffolding.
-
-# GPU backend dependencies are declared in individual engine crates (no optional deps in workspace)

 [profile.release]
 opt-level = 3
@@ -1,55 +0,0 @@
-# syntax=docker/dockerfile:1
-
-############################
-# Builder stage
-############################
-FROM rust:1.85-slim-bookworm AS builder
-
-# Install build dependencies
-RUN apt-get update \
- && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-    pkg-config \
-    libssl-dev \
-    ca-certificates \
- && rm -rf /var/lib/apt/lists/*
-
-WORKDIR /build
-
-# Copy workspace files
-COPY Cargo.toml Cargo.lock rust-toolchain taplo.toml ./
-COPY crates ./crates
-COPY tests ./tests
-
-# Build the miner-cli in release mode
-RUN cargo build --release -p miner-cli --locked
-
-# Strip debug symbols to reduce binary size
-RUN strip target/release/quantus-miner || true
-
-############################
-# Runtime-only stage
-############################
-FROM debian:bookworm-slim
-
-# Install runtime dependencies
-RUN apt-get update \
- && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-    ca-certificates \
- && rm -rf /var/lib/apt/lists/*
-
-# Copy binary from builder stage
-COPY --from=builder /build/target/release/quantus-miner /usr/local/bin/quantus-miner
-
-# Expose miner API port and metrics port
-EXPOSE 9833 9900
-
-# Run as unprivileged user
-RUN useradd --system --uid 10001 quantus
-USER 10001:10001
-
-# Default working directory
-WORKDIR /data
-
-# Start the miner
-ENTRYPOINT ["quantus-miner"]
-
@@ -1,256 +0,0 @@
-# External Miner Protocol Specification
-
-This document defines the JSON-based HTTP protocol for communication between the Resonance Network node and an external QPoW miner service.
-
-## Overview
-
-The node delegates the mining task (finding a valid nonce) to an external service. The node provides the necessary parameters (header hash, difficulty, nonce range) and the external miner searches for a valid nonce according to the QPoW rules defined in the `qpow-math` crate. The miner returns the result, including the winning nonce, when found.
-
-## Data Types
-
-See the `resonance-miner-api` crate for the canonical Rust definitions of these structures.
-
- `job_id`: String (UUID recommended) - Unique identifier for a specific mining task, generated by the node.
- `mining_hash`: String (64 hex chars, no 0x) - The header hash for which to find a nonce.
- `difficulty`: String (u64 as string) - The target difficulty for the mining job.
- `nonce_start`: String (128 hex chars, no 0x) - The starting nonce value (inclusive) for the search range.
- `nonce_end`: String (128 hex chars, no 0x) - The ending nonce value (inclusive) for the search range.
- `status`: Enum (`ApiResponseStatus`) - Indicates the state or result of an API call.
- `message`: String (optional) - Provides details for `Error` status responses.
- `nonce`: String (Hex, no 0x) - Represents the `U512` value of the current or winning nonce.
- `work`: String (128 hex chars, no 0x) - Represents the winning nonce as `[u8; 64]`. This is the value the node needs for verification.
- `hash_count`: Number (u64) - Number of nonces checked by the miner for the job.
- `elapsed_time`: Number (f64) - Time in seconds the miner spent on the job.
-
-## Endpoints
-
-### 1. Submit Mining Job
-
- **Endpoint:** `POST /mine`
- **Description:** The node requests the external miner to start searching for a valid nonce.
- **Request Body (`MiningRequest`):**
-  ```json
-  {
-    "job_id": "...",
-    "mining_hash": "...",
-    "difficulty": "...",
-    "nonce_start": "...",
-    "nonce_end": "..."
-  }
-  ```
- **Response Body (`MiningResponse`):**
-  - Success (200 OK):
-    ```json
-    {
-      "status": "accepted",
-      "job_id": "..."
-    }
-    ```
-  - Error (400 Bad Request - Invalid Input / 409 Conflict - Duplicate Job ID):
-    ```json
-    {
-      "status": "error",
-      "job_id": "...",
-      "message": "..." // e.g., "Job already exists", "Invalid mining_hash (...)"
-    }
-    ```
-
-### 2. Get Job Result
-
- **Endpoint:** `GET /result/{job_id}`
- **Description:** The node polls the external miner to check the status and retrieve the result.
- **Path Parameter:**
-  - `job_id`: String (UUID) - The ID of the job to query.
- **Response Body (`MiningResult`):**
-  - Job Completed (200 OK):
-    ```json
-    {
-      "status": "completed",
-      "job_id": "...",
-      "nonce": "...",   // U512 hex value of winning nonce
-      "work": "...",    // [u8; 64] hex value of winning nonce
-      "hash_count": ..., // u64
-      "elapsed_time": ... // f64 seconds
-    }
-    ```
-  - Job Still Running (200 OK):
-    ```json
-    {
-      "status": "running",
-      "job_id": "...",
-      "nonce": "...",   // Current nonce being checked (U512 hex)
-      "work": null,
-      "hash_count": ..., // u64
-      "elapsed_time": ... // f64 seconds
-    }
-    ```
-  - Job Failed (e.g., nonce range exhausted) (200 OK):
-    ```json
-    {
-      "status": "failed",
-      "job_id": "...",
-      "nonce": "...",   // Final nonce checked (U512 hex)
-      "work": null,
-      "hash_count": ..., // u64
-      "elapsed_time": ... // f64 seconds
-    }
-    ```
-  - Job Not Found (404 Not Found):
-    ```json
-    {
-      "status": "not_found",
-      "job_id": "...",
-      "nonce": null,
-      "work": null,
-      "hash_count": 0,
-      "elapsed_time": 0.0
-    }
-    ```
-
-### 3. Cancel Mining Job
-
- **Endpoint:** `POST /cancel/{job_id}`
- **Description:** The node requests the external miner to stop working on a specific job.
- **Path Parameter:**
-  - `job_id`: String (UUID) - The ID of the job to cancel.
- **Request Body:** (Empty)
- **Response Body (`MiningResponse`):
-  - Success (200 OK):
-    ```json
-    {
-      "status": "cancelled",
-      "job_id": "..."
-    }
-    ```
-  - Job Not Found (404 Not Found):
-    ```json
-    {
-      "status": "not_found",
-      "job_id": "..."
-    }
-    ```
-
-## Notes
-
- All hex values (`mining_hash`, `nonce_start`, `nonce_end`, `nonce`, `work`) should be sent **without** the `0x` prefix.
- The miner must implement the validation logic defined in `qpow_math::is_valid_nonce`.
- The node relies primarily on the `work` field in the `MiningResult` (when status is `completed`) for constructing the `QPoWSeal`.
-
-# External Miner Protocol Specification
-
-This document defines the JSON-based HTTP protocol for communication between the node and an external QPoW miner.
-
-## Overview
-
-The node delegates the mining task to an external service. The node provides the necessary parameters (mining hash, difficulty, and a nonce range) and the external miner searches for a valid nonce within that range. The miner returns the nonce and the resulting work hash when a solution is found.
-
-## Data Types
-
- `job_id`: String (UUID recommended) - Identifier for a specific mining task.
- `mining_hash`: String (Hex-encoded, 32-byte hash, H256) - The hash derived from the block header data that the miner needs to solve.
- `difficulty`: String (Decimal representation of u64) - The target difficulty for the block.
- `nonce_start`: String (Hex-encoded, 64-byte value, U512) - The starting nonce value (inclusive).
- `nonce_end`: String (Hex-encoded, 64-byte value, U512) - The ending nonce value (inclusive).
- `nonce`: String (Hex-encoded, 64-byte value, U512) - The solution found by the miner.
- `work`: String (Hex-encoded, 32-byte hash, H256) - The hash resulting from the combination of `mining_hash` and `nonce`, meeting the difficulty requirement.
- `status`: String Enum - Indicates the state or result of an API call.
-
-## Endpoints
-
-### 1. Start Mining Job
-
- **Endpoint:** `POST /mine`
- **Description:** The node requests the external miner to start searching for a valid nonce within the specified range for the given parameters.
- **Request Body (application/json):**
-  ```json
-  {
-    "job_id": "...",      // String (UUID), generated by the node
-    "mining_hash": "...", // Hex String (H256)
-    "difficulty": "...",  // String (u64 decimal)
-    "nonce_start": "...", // Hex String (U512 hex)
-    "nonce_end": "..."    // Hex String (U512 hex)
-  }
-  ```
- **Response Body (application/json):**
-  - Success (200 OK):
-    ```json
-    {
-      "status": "accepted",
-      "job_id": "..." // String (UUID), confirming the job ID received
-    }
-    ```
-  - Error (e.g., 400 Bad Request, 500 Internal Server Error):
-    ```json
-    {
-      "status": "rejected",
-      "reason": "..." // String (Description of error)
-    }
-    ```
-
-### 2. Get Job Result
-
- **Endpoint:** `GET /result/{job_id}`
- **Description:** The node polls the external miner to check the status and retrieve the result of a previously submitted job.
- **Path Parameter:**
-  - `job_id`: String (UUID) - The ID of the job to query.
- **Response Body (application/json):**
-  - Solution Found (200 OK):
-    ```json
-    {
-      "status": "found",
-      "job_id": "...", // String (UUID)
-      "nonce": "...", // Hex String (U512 hex)
-      "work": "CAFEBABE01.."   // Hex String (H256 hex)
-    }
-    ```
-  - Still Working (200 OK):
-    ```json
-    {
-      "status": "working",
-      "job_id": "..." // String (UUID)
-    }
-    ```
-  - Job Stale/Cancelled (200 OK): Indicates the job is no longer valid (e.g., the node requested cancellation or submitted work for a newer block).
-    ```json
-    {
-      "status": "stale",
-      "job_id": "..." // String (UUID)
-    }
-    ```
-  - Job Not Found (404 Not Found):
-    ```json
-    {
-      "status": "not_found",
-      "job_id": "..." // String (UUID)
-    }
-    ```
-
-### 3. Cancel Mining Job
-
- **Endpoint:** `POST /cancel/{job_id}`
- **Description:** The node requests the external miner to stop working on a specific job. This is typically used when the node receives a new block or its mining parameters change, making the old job obsolete.
- **Path Parameter:**
-  - `job_id`: String (UUID) - The ID of the job to cancel.
- **Request Body:** (Empty)
- **Response Body (application/json):**
-  - Success (200 OK):
-    ```json
-    {
-      "status": "cancelled",
-      "job_id": "..." // String (UUID)
-    }
-    ```
-  - Job Not Found (404 Not Found):
-    ```json
-    {
-      "status": "not_found",
-      "job_id": "..." // String (UUID)
-    }
-    ```
-
-## Notes
-
- The external miner should iterate from `nonce_start` up to and including `nonce_end` when searching for a valid nonce.
- The miner should return the `nonce` and the calculated `work` hash when a solution is found.
- The node uses the returned `nonce` and `work` (along with the fetched `difficulty`) to construct the `QPoWSeal` and submit it.
- The external miner should not need to know anything about the runtime or the code; it only needs to perform the nonce search and return the results. 
@@ -9,35 +9,58 @@ High-performance external mining service for Quantus Network with support for CP
 cargo build -p miner-cli --release

 # With GPU support (recommended)
-cargo build -p miner-cli --features gpu --release
+cargo build -p miner-cli --release
 ```

 The binary will be available at `target/release/quantus-miner`.

 ## Running

-```bash
-# CPU-only mining (default: auto-detected CPU cores)
-./target/release/quantus-miner --cpu-workers 4
+The node requires a shared auth token and TLS cert pin. Both live under the
+node's chain config dir (`<base-path>/chains/<chain>/`):

-# GPU-only mining (requires --features gpu build)
-./target/release/quantus-miner --gpu-workers 1
+- `miner-auth-token` — shared secret (**not** logged by the node; read this file)
+- `miner-tls-cert-sha256` — SHA-256 of the miner TLS cert (also printed in node logs)
+
+```bash
+# Preferred: mount/read the node's chain config files
+./target/release/quantus-miner serve \
+  --node-addr 127.0.0.1:9833 \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --cpu-workers 4
+
+# Or pass values directly (token from miner-auth-token; fingerprint also in node logs)
+./target/release/quantus-miner serve \
+  --node-addr 127.0.0.1:9833 \
+  --auth-token <TOKEN> \
+  --tls-cert-sha256 <FINGERPRINT> \
+  --gpu-devices 1

 # Hybrid CPU+GPU mining
-./target/release/quantus-miner --cpu-workers 4 --gpu-workers 1
-
-# Custom port and metrics
-./target/release/quantus-miner --cpu-workers 2 --port 8000 --metrics-port 9900
+./target/release/quantus-miner serve \
+  --node-addr 127.0.0.1:9833 \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --cpu-workers 4 \
+  --gpu-devices 1
 ```

 ## Configuration

 | Argument | Environment Variable | Description | Default |
 |----------|---------------------|-------------|---------|
+| `--node-addr <ADDR>` | `MINER_NODE_ADDR` | Node address to connect to | `127.0.0.1:9833` |
+| `--auth-token <TOKEN>` | `MINER_AUTH_TOKEN` | Shared secret from the node's `miner-auth-token` file (not logged) | required |
+| `--auth-token-file <PATH>` | `MINER_AUTH_TOKEN_FILE` | Read the shared secret from a file (preferred) | — |
+| `--tls-cert-sha256 <HEX>` | `MINER_TLS_CERT_SHA256` | SHA-256 of the node's miner TLS cert (`miner-tls-cert-sha256` / node logs) | required |
+| `--tls-cert-sha256-file <PATH>` | `MINER_TLS_CERT_SHA256_FILE` | Read the TLS cert fingerprint from a file | — |
 | `--cpu-workers <N>` | `MINER_CPU_WORKERS` | Number of CPU worker threads | Auto-detect |
-| `--gpu-workers <N>` | `MINER_GPU_WORKERS` | Number of GPU worker threads | 0 |
-| `--port <PORT>` | `MINER_PORT` | HTTP API port | 9833 |
-| `--metrics-port <PORT>` | `MINER_METRICS_PORT` | Prometheus metrics port | Disabled |
+| `--gpu-devices <N>` | `MINER_GPU_DEVICES` | Number of GPU devices | Auto-detect |
+| `--gpu-batch-size <N>` | `MINER_GPU_BATCH_SIZE` | GPU batch size in nonces | 1000000 |
+| `--cpu-batch-size <N>` | `MINER_CPU_BATCH_SIZE` | CPU batch size in hashes | 10000 |
+| `--gpu-throttle-ms <MS>` | `MINER_GPU_THROTTLE_MS` | Sleep duration (ms) between GPU batches | 0 |
+| `--metrics-port <PORT>` | `MINER_METRICS_PORT` | Prometheus metrics port | 9900 |

 ## GPU Mining

@@ -51,7 +74,7 @@ GPU support uses WGPU for cross-platform acceleration:

 **Build with GPU support:**
 ```bash
-cargo build -p miner-cli --features gpu --release
+cargo build -p miner-cli --release
 ```

 **Platform requirements:**
@@ -67,48 +90,60 @@ cargo build -p miner-cli --features gpu --release

 ## Examples

+All `serve` examples need the auth token and TLS pin (files or inline values).
+
 ```bash
 # CPU mining with 8 workers
-./target/release/quantus-miner --cpu-workers 8
+./target/release/quantus-miner serve \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --cpu-workers 8

 # Pure GPU mining
-./target/release/quantus-miner --gpu-workers 1
+./target/release/quantus-miner serve \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --gpu-devices 1
+
+# GPU mining with throttle (reduce GPU utilization)
+./target/release/quantus-miner serve \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --gpu-devices 1 --gpu-throttle-ms 50

 # Hybrid mining: 4 CPU + 1 GPU workers
-./target/release/quantus-miner --cpu-workers 4 --gpu-workers 1
+./target/release/quantus-miner serve \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --cpu-workers 4 --gpu-devices 1

 # With verbose logging
-RUST_LOG=debug ./target/release/quantus-miner --cpu-workers 2 --gpu-workers 1
+RUST_LOG=debug ./target/release/quantus-miner serve \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
+  --cpu-workers 2 --gpu-devices 1

 # Production setup with metrics
-./target/release/quantus-miner \
+./target/release/quantus-miner serve \
+  --node-addr 127.0.0.1:9833 \
+  --auth-token-file /path/to/miner-auth-token \
+  --tls-cert-sha256-file /path/to/miner-tls-cert-sha256 \
  --cpu-workers 6 \
-  --gpu-workers 1 \
-  --port 9833 \
+  --gpu-devices 1 \
  --metrics-port 9900
 ```

-## API Endpoints
+## Protocol

- `POST /mine`: Submit mining job
- `GET /result/{job_id}`: Get job status/result
- `POST /cancel/{job_id}`: Cancel job
+The miner uses a QUIC-based protocol for communication with the node:

-Full API specification: `api/openapi.yaml`
+- **Transport**: QUIC with TLS 1.3 (self-signed certificate, pinned by SHA-256)
+- **Auth**: `Ready { token }` must match the node's `miner-auth-token`
+- **ALPN**: `quantus-miner/2`
+- **Port**: 9833 (default)
+- **Messages**: `Ready` (miner→node), `NewJob` (node→miner), `JobResult` (miner→node)

-## Docker
-
-```bash
-# Quick start
-docker pull ghcr.io/quantus-network/quantus-miner:latest
-docker run -d -p 9833:9833 -p 9900:9900 \
-  ghcr.io/quantus-network/quantus-miner:latest \
-  --cpu-workers 4 --metrics-port 9900
-
-# Build from source
-docker build -t quantus-miner .
-docker run -d -p 9833:9833 quantus-miner --cpu-workers 4
-```
+For full protocol specification, see the node's `MINING.md`.

 ## Benchmarking

@@ -117,8 +152,8 @@ docker run -d -p 9833:9833 quantus-miner --cpu-workers 4
 ./target/release/quantus-miner benchmark --cpu-workers 8 --duration 30

 # Benchmark GPU performance  
-./target/release/quantus-miner benchmark --gpu-workers 1 --duration 30
+./target/release/quantus-miner benchmark --gpu-devices 1 --duration 30

 # Benchmark hybrid performance
-./target/release/quantus-miner benchmark --cpu-workers 4 --gpu-workers 1 --duration 30
+./target/release/quantus-miner benchmark --cpu-workers 4 --gpu-devices 1 --duration 30
 ```
@@ -0,0 +1,80 @@
+# Security Policy
+
+Quantus Network takes the security of our users seriously and welcomes reports
+from security researchers.
+
+## Reporting a vulnerability
+
+**Please do not open a public issue, pull request, or social-media post for a
+security vulnerability.** Public disclosure before a fix is available puts
+users' funds at risk. Use one of the private channels below and we will
+coordinate a fix and disclosure with you.
+
+1. **GitHub private security advisory (preferred).**
+   [Report a vulnerability](https://github.com/quantus-network/quantus-miner/security/advisories/new).
+   This gives you a private, structured thread with the maintainers and is the
+   fastest way to reach us.
+2. **Email.** Send details to **security@quantus.com**.
+
+### What to include
+
+- A clear description of the issue and its security impact.
+- A working proof of concept is required. It must run directly against this
+  project, using a release build or locally built version, and demonstrate the
+  reported behavior and security impact. Standalone code, mathematical
+  examples, or simulations that only reproduce the theory without exercising
+  the project do not satisfy this requirement.
+- Step-by-step reproduction instructions.
+- Affected platforms and versions.
+- Any relevant logs, addresses, or transaction IDs (for on-chain issues).
+
+If you used AI tooling to find or write up the report, please say so.
+
+## Our commitment (safe harbor)
+
+We consider security research conducted in good faith under this policy to be
+authorized. We will not pursue or support legal action against researchers who:
+
+- make a good-faith effort to avoid privacy violations, data destruction, and
+  interruption or degradation of our services;
+- only interact with accounts they own or have explicit permission to access;
+  and
+- give us a reasonable opportunity to fix an issue before disclosing it
+  publicly.
+
+If in doubt about whether an action is authorized, ask us first at
+security@quantus.com.
+
+## What to expect
+
+- **Acknowledgement:** within **2 business days**.
+- **Triage and initial assessment:** within **7 business days**.
+- **Coordinated disclosure:** we aim to ship a fix and coordinate public
+  disclosure within **90 days** of the report. We will keep you updated on
+  progress and agree on a disclosure date with you.
+- **Credit:** with your permission, we are happy to publicly credit you for the
+  report once a fix is released.
+
+## Scope
+
+**In scope:** the external mining service (CPU and GPU) in this repository — anything that could lead to theft or misattribution of mining rewards, invalid blocks being produced, or remote code execution or crash via its service API.
+
+**Out of scope:** issues in third-party services or nodes we do not operate;
+reports generated solely by automated scanners without a demonstrated impact;
+low-severity or informational issues on our marketing and landing websites;
+and social-engineering or physical attacks.
+
+## Rewards
+
+At our **sole discretion**, we may offer a reward for a valid report. To be
+eligible, a report must be submitted **privately** through one of the channels
+above and identify a genuine vulnerability with real impact on users. Trivial
+or low-impact findings, automated-scanner output without a working proof of
+concept, and already-known issues are **not** eligible. There is no fixed
+bounty and no guaranteed payout; whether a report qualifies, and any amount,
+are determined solely by Quantus Network.
+
+## Supported versions
+
+Only the latest release is supported; security fixes are delivered in new
+versions. Please keep your installation up to date.
@@ -1,242 +0,0 @@
-openapi: 3.0.0
-info:
-  title: Quantus External Miner API
-  version: 0.1.0
-  description: API for managing QPoW mining jobs for the Resonance Network external miner service.
-
-paths:
-  /mine:
-    post:
-      summary: Submit a new mining job
-      description: Accepts a mining job request and adds it to the queue if valid.
-      requestBody:
-        required: true
-        content:
-          application/json:
-            schema:
-              $ref: '#/components/schemas/MiningRequest'
-      responses:
-        '200':
-          description: Job accepted successfully.
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResponseAccepted'
-        '400':
-          description: Invalid request (e.g., missing fields, invalid format).
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResponseError'
-        '409':
-          description: Conflict (e.g., duplicate job ID).
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResponseError'
-
-  /result/{job_id}:
-    get:
-      summary: Get mining job status and result
-      description: Retrieves the current status and potentially the found nonce for a given job ID.
-      parameters:
-        - name: job_id
-          in: path
-          required: true
-          description: The ID of the mining job to retrieve.
-          schema:
-            type: string
-            format: uuid
-      responses:
-        '200':
-          description: Job status retrieved successfully (could be running, completed, or failed).
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResult'
-        '404':
-          description: Job ID not found.
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResultNotFound'
-
-  /cancel/{job_id}:
-    post:
-      summary: Cancel a mining job
-      description: Attempts to cancel an ongoing mining job.
-      parameters:
-        - name: job_id
-          in: path
-          required: true
-          description: The ID of the mining job to cancel.
-          schema:
-            type: string
-            format: uuid
-      responses:
-        '200':
-          description: Job cancelled successfully.
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResponseCancelled'
-        '404':
-          description: Job ID not found.
-          content:
-            application/json:
-              schema:
-                $ref: '#/components/schemas/MiningResponseNotFound'
-
-components:
-  schemas:
-    ApiResponseStatus:
-      type: string
-      description: Status code indicating the result of an API call or the state of a mining job.
-      enum:
-        - accepted
-        - running
-        - completed
-        - failed
-        - cancelled
-        - not_found
-        - error
-      example: running
-
-    MiningRequest:
-      type: object
-      required:
-        - job_id
-        - mining_hash
-        - difficulty
-        - nonce_start
-        - nonce_end
-      properties:
-        job_id:
-          type: string
-          format: uuid
-          description: Unique identifier for the mining job (UUID format recommended).
-          example: "123e4567-e89b-12d3-a456-426614174000"
-        mining_hash:
-          type: string
-          description: The header hash to mine on (hex encoded, 64 characters, no 0x).
-          example: "a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0a0"
-          pattern: '^[a-fA-F0-9]{64}$'
-        difficulty:
-          type: string
-          description: Target difficulty for the mining job (string representation of a u64).
-          example: "10000000000"
-        nonce_start:
-          type: string
-          description: Starting nonce value (hex encoded, 128 characters, no 0x).
-          example: "00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000"
-          pattern: '^[a-fA-F0-9]{128}$'
-        nonce_end:
-          type: string
-          description: Ending nonce value (hex encoded, 128 characters, no 0x).
-          example: "ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff"
-          pattern: '^[a-fA-F0-9]{128}$'
-
-    MiningResponseBase:
-      type: object
-      required:
-        - status
-        - job_id
-      properties:
-        status:
-          $ref: '#/components/schemas/ApiResponseStatus'
-        job_id:
-          type: string
-          format: uuid
-          description: The ID of the job associated with the response.
-          example: "123e4567-e89b-12d3-a456-426614174000"
-
-    MiningResponseAccepted:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          properties:
-            status:
-              example: accepted
-
-    MiningResponseCancelled:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          properties:
-            status:
-              example: cancelled
-
-    MiningResponseError:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          properties:
-            status:
-              example: error
-            message:
-              type: string
-              description: Details about the error that occurred.
-              example: "Job already exists"
-
-    MiningResponseNotFound:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          properties:
-            status:
-              example: not_found
-
-    MiningResult:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          required:
-            - hash_count
-            - elapsed_time
-          properties:
-            nonce:
-              type: string
-              nullable: true
-              description: The current or winning nonce (U512 hex, no 0x).
-              example: "1a2b3c"
-              pattern: '^[a-fA-F0-9]+$'
-            work:
-              type: string
-              nullable: true
-              description: The winning nonce ([u8; 64] hex, 128 chars, no 0x). Only present if status is 'completed'.
-              example: "000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001a2b3c"
-              pattern: '^[a-fA-F0-9]{128}$'
-            hash_count:
-              type: integer
-              format: uint64
-              description: Number of nonces checked so far for this job.
-              example: 123456
-            elapsed_time:
-              type: number
-              format: double
-              description: Time elapsed (in seconds) since the job started.
-              example: 15.732
-
-    MiningResultNotFound:
-      allOf:
-        - $ref: '#/components/schemas/MiningResponseBase'
-        - type: object
-          properties:
-            status:
-              example: not_found
-            nonce:
-              type: string
-              nullable: true
-              example: null
-            work:
-              type: string
-              nullable: true
-              example: null
-            hash_count:
-              type: integer
-              format: uint64
-              example: 0
-            elapsed_time:
-              type: number
-              format: double
-              example: 0.0
@@ -0,0 +1,7 @@
+#!/usr/bin/env bash
+# Run the same clippy command as CI (see .github/workflows/ci.yml)
+set -euo pipefail
+
+cargo +nightly fmt
+taplo format
+cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
@@ -20,9 +20,6 @@ simd-poseidon2 = ["pow-core/simd-poseidon2"]
 [dependencies]
 pow-core = { path = "../pow-core" }
 primitive-types = { workspace = true }
-log = { workspace = true }
-thiserror = { workspace = true }
-anyhow = { workspace = true }

 [dev-dependencies]
 hex = { workspace = true }
@@ -1,14 +1,15 @@
 use criterion::{black_box, criterion_group, criterion_main, Criterion};
-use engine_cpu::{FastCpuEngine, MinerEngine, Range};
+use engine_cpu::{AtomicBoolCancelCheck, FastCpuEngine, MinerEngine, Range};
 use pow_core::{hash_from_nonce, JobContext};
 use primitive_types::U512;
 use rand::RngCore;
 use std::sync::atomic::AtomicBool;

 fn bench_cpu_fast_engine(c: &mut Criterion) {
-    // Create the engine
-    let engine = FastCpuEngine::new();
+    // Create the engine with batch size of 10000
+    let engine = FastCpuEngine::new(10_000);
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    let large_range = Range {
        start: U512::from(0u64),
@@ -18,14 +19,14 @@ fn bench_cpu_fast_engine(c: &mut Criterion) {
    c.bench_function("cpu_fast_large_range", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(10_000_000u64);
            let ctx = JobContext::new(header, difficulty);

            let result = engine.search_range(
                black_box(&ctx),
                black_box(large_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -35,7 +36,7 @@ fn bench_cpu_fast_engine(c: &mut Criterion) {
 fn bench_hash_from_nonce(c: &mut Criterion) {
    // Create a test job context
    let mut header = [0u8; 32];
-    rand::thread_rng().fill_bytes(&mut header);
+    rand::rng().fill_bytes(&mut header);
    let difficulty = U512::from(1000u64);
    let ctx = JobContext::new(header, difficulty);

@@ -1,23 +1,14 @@
 #![deny(rust_2018_idioms)]
 #![forbid(unsafe_code)]

-//! CPU mining engine scaffolding and trait definition.
+//! CPU mining engine for Quantus.
 //!
-//! This crate defines a minimal `MinerEngine` trait so the service layer can
-//! orchestrate mining without knowing about specific math/device details,
-//! plus a baseline CPU implementation that uses the reference path in `pow-core`.
-//!
-//! The baseline engine performs a straightforward linear scan across an inclusive
-//! nonce range, computing distance per nonce using the context (header-derived)
-//! constants. It is intended as a correctness reference and may be replaced at
-//! runtime by faster engines (e.g., incremental CPU path, CUDA/OpenCL engines).
+//! This crate defines the `MinerEngine` trait for mining orchestration,
+//! plus a fast CPU implementation using optimized paths from `pow-core`.

-use core::cmp::Ordering;
-
-use pow_core::{is_valid_nonce, JobContext};
+use pow_core::JobContext;
 use primitive_types::U512;
-use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
-use std::time::Duration;
+use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};

 /// An inclusive nonce range to search.
 #[derive(Clone, Debug)]
@@ -44,13 +35,6 @@ pub enum FoundOrigin {
 }

 /// Status from an engine search.
-///
-/// For synchronous `search_range` calls, the final outcome will be one of:
-/// - `Found`
-/// - `Exhausted`
-/// - `Cancelled`
-///
-/// `Running` is included for potential future async/streaming engines.
 #[derive(Clone, Debug)]
 pub enum EngineStatus {
    Running {
@@ -67,117 +51,76 @@ pub enum EngineStatus {
    Cancelled {
        hash_count: u64,
    },
+    /// GPU device lost or unresponsive. Worker should exit permanently.
+    DeviceLost {
+        hash_count: u64,
+    },
+}
+
+/// Cancellation checker passed to search_range.
+/// Returns true if the search should be cancelled.
+pub trait CancelCheck: Send + Sync {
+    fn is_cancelled(&self) -> bool;
+}
+
+/// Simple cancel check using an AtomicBool flag.
+/// Useful for benchmarks and simple scenarios.
+pub struct AtomicBoolCancelCheck<'a>(pub &'a AtomicBool);
+
+impl CancelCheck for AtomicBoolCancelCheck<'_> {
+    fn is_cancelled(&self) -> bool {
+        self.0.load(Ordering::Relaxed)
+    }
+}
+
+/// Cancel check using job ID comparison.
+/// When a new job starts, the current_job_id is incremented.
+/// Workers compare their job_id against the current to detect cancellation.
+pub struct JobIdCancelCheck<'a> {
+    pub current_job_id: &'a AtomicU64,
+    pub my_job_id: u64,
+}
+
+impl CancelCheck for JobIdCancelCheck<'_> {
+    fn is_cancelled(&self) -> bool {
+        self.current_job_id.load(Ordering::Relaxed) != self.my_job_id
+    }
 }

 /// Abstract mining engine interface.
 ///
 /// The service layer depends only on this trait to manage jobs.
-/// Different engines (baseline CPU, optimized CPU, CUDA, OpenCL) can implement
-/// this trait and be selected at runtime via configuration.
+/// Different engines (CPU, CUDA, OpenCL) can implement this trait.
 pub trait MinerEngine: Send + Sync {
-    /// Human-readable engine name (for logs/metrics).
+    /// Human-readable engine name (for logs).
    fn name(&self) -> &'static str;

    /// Prepare a precomputed context for a job (header + difficulty).
    fn prepare_context(&self, header_hash: [u8; 32], difficulty: U512) -> JobContext;

    /// Search an inclusive nonce range with cancellation support.
-    ///
-    /// Implementations should:
-    /// - Respect `cancel` promptly to minimize wasted work after a solution is found.
-    /// - Return `Found` with a `Candidate` on success.
-    /// - Return `Exhausted` if the range is fully searched without a solution.
-    /// - Return `Cancelled` if `cancel` was observed during the search.
-    fn search_range(&self, ctx: &JobContext, range: Range, cancel: &AtomicBool) -> EngineStatus;
+    fn search_range(
+        &self,
+        ctx: &JobContext,
+        range: Range,
+        cancel: &dyn CancelCheck,
+    ) -> EngineStatus;

    /// Enable downcasting to concrete engine types.
    fn as_any(&self) -> &dyn std::any::Any;
 }

-/// Baseline CPU engine.
-///
-/// This implementation uses the reference path from `pow-core` for distance
-/// computation and scans the range linearly, one nonce at a time.
-#[derive(Default)]
-pub struct BaselineCpuEngine;
-
-impl BaselineCpuEngine {
-    pub fn new() -> Self {
-        Self
-    }
+/// Fast CPU engine using optimized pow-core helpers.
+pub struct FastCpuEngine {
+    /// How often to check for cancellation (in hashes)
+    batch_size: u64,
 }

-impl MinerEngine for BaselineCpuEngine {
-    fn name(&self) -> &'static str {
-        "cpu-baseline"
-    }
-
-    fn prepare_context(&self, header_hash: [u8; 32], difficulty: U512) -> JobContext {
-        JobContext::new(header_hash, difficulty)
-    }
-
-    fn as_any(&self) -> &dyn std::any::Any {
-        self
-    }
-
-    fn search_range(&self, ctx: &JobContext, range: Range, cancel: &AtomicBool) -> EngineStatus {
-        // Ensure start <= end (inclusive range). If not, treat as exhausted.
-        if range.start > range.end {
-            return EngineStatus::Exhausted { hash_count: 0 };
-        }
-
-        let mut current = range.start;
-        let mut hash_count: u64 = 0;
-
-        loop {
-            // Cancellation check
-            if cancel.load(AtomicOrdering::Relaxed) {
-                return EngineStatus::Cancelled { hash_count };
-            }
-
-            // Compute hash for this nonce using Bitcoin-style double Poseidon2
-            let (is_valid, hash) =
-                is_valid_nonce(ctx.header, current.to_big_endian(), ctx.difficulty);
-            hash_count = hash_count.saturating_add(1);
-
-            // Check if it meets difficulty target
-            if is_valid {
-                let work = current.to_big_endian();
-                let candidate = Candidate {
-                    nonce: current,
-                    work,
-                    hash,
-                };
-                return EngineStatus::Found {
-                    candidate,
-                    hash_count,
-                    origin: FoundOrigin::Cpu,
-                };
-            }
-
-            // Advance or finish
-            match current.cmp(&range.end) {
-                Ordering::Less => {
-                    current = current.saturating_add(U512::one());
-                }
-                _ => {
-                    // End of inclusive range reached
-                    break EngineStatus::Exhausted { hash_count };
-                }
-            }
-        }
-    }
-}
-
-// Re-export commonly used items for convenience by consumers.
-
-// Fast CPU engine using incremental pow-core helpers (init_worker_y0 + step_mul)
-#[derive(Default)]
-pub struct FastCpuEngine;
-
 impl FastCpuEngine {
-    pub fn new() -> Self {
-        Self
+    pub fn new(batch_size: u64) -> Self {
+        Self {
+            batch_size: batch_size.max(1), // Ensure at least 1
+        }
    }
 }

@@ -194,24 +137,34 @@ impl MinerEngine for FastCpuEngine {
        self
    }

-    fn search_range(&self, ctx: &JobContext, range: Range, cancel: &AtomicBool) -> EngineStatus {
+    fn search_range(
+        &self,
+        ctx: &JobContext,
+        range: Range,
+        cancel: &dyn CancelCheck,
+    ) -> EngineStatus {
        use pow_core::{hash_from_nonce, is_valid_hash, step_nonce};

-        // Ensure start <= end (inclusive range). If not, treat as exhausted.
        if range.start > range.end {
            return EngineStatus::Exhausted { hash_count: 0 };
        }

        let mut current = range.start;
        let mut hash_count: u64 = 0;
+        // Use decrementing counter to avoid modulo division in hot loop
+        // Initialize to 0 so we check cancellation immediately on first iteration
+        let mut until_check: u64 = 0;

        loop {
-            // Cancellation check
-            if cancel.load(AtomicOrdering::Relaxed) {
-                return EngineStatus::Cancelled { hash_count };
+            // Check for cancellation every batch_size hashes
+            if until_check == 0 {
+                if cancel.is_cancelled() {
+                    return EngineStatus::Cancelled { hash_count };
+                }
+                until_check = self.batch_size;
            }
+            until_check -= 1;

-            // Compute hash using Bitcoin-style double Poseidon2
            let hash = hash_from_nonce(ctx, current);
            hash_count = hash_count.saturating_add(1);

@@ -232,7 +185,6 @@ impl MinerEngine for FastCpuEngine {
                break;
            }

-            // Advance to next nonce
            current = step_nonce(current);
        }

@@ -240,140 +192,8 @@ impl MinerEngine for FastCpuEngine {
    }
 }

-#[derive(Default)]
-pub struct ChainManipulatorEngine {
-    /// Base sleep per batch in nanoseconds; actual sleep increases linearly with job count.
-    pub base_delay_ns: u64,
-    /// Number of nonce attempts between sleeps.
-    pub step_batch: u64,
-    /// Optional cap for solved-block throttling (sleep index will not exceed this).
-    pub throttle_cap: Option<u64>,
-    /// Monotonically increasing solved-block counter used to scale throttling.
-    /// Public so the service can initialize from CLI (pick up where we left off).
-    pub job_index: AtomicU64,
-}
-
-impl ChainManipulatorEngine {
-    pub fn new() -> Self {
-        // Start fast: first block has no throttle (job_index = 0 -> 0ns sleep),
-        // then 0.5ms per batch at block 1, 1.0ms at block 2, etc.
-        Self {
-            base_delay_ns: 500_000, // 0.5 ms
-            step_batch: 10_000,     // sleep every 10k nonce checks
-            throttle_cap: None,     // unlimited by default
-            job_index: AtomicU64::new(0),
-        }
-    }
-}
-
-impl MinerEngine for ChainManipulatorEngine {
-    fn name(&self) -> &'static str {
-        "cpu-chain-manipulator"
-    }
-
-    fn prepare_context(&self, header_hash: [u8; 32], difficulty: U512) -> JobContext {
-        // Per-block throttling: do NOT increment here. We increment on Found (i.e., when a block is solved).
-        let ctx = JobContext::new(header_hash, difficulty);
-        // Debug: log current throttle state at job start
-        log::debug!(
-            target: "miner",
-            "manipulator throttle start: solved_blocks={}, sleep_ns_per_batch={}, step_batch={}",
-            self.job_index.load(AtomicOrdering::Relaxed),
-            self.base_delay_ns.saturating_mul(self.job_index.load(AtomicOrdering::Relaxed)),
-            self.step_batch
-        );
-        ctx
-    }
-
-    fn as_any(&self) -> &dyn std::any::Any {
-        self
-    }
-
-    fn search_range(&self, ctx: &JobContext, range: Range, cancel: &AtomicBool) -> EngineStatus {
-        use pow_core::{hash_from_nonce, is_valid_hash, step_nonce};
-
-        if range.start > range.end {
-            return EngineStatus::Exhausted { hash_count: 0 };
-        }
-
-        // Bitcoin-style hashing path
-        let mut current = range.start;
-        let mut hash_count: u64 = 0;
-        let mut batch_counter: u64 = 0;
-
-        // Per-block throttling: derived from blocks solved so far (apply cap if configured).
-        let mut solved_blocks = self.job_index.load(AtomicOrdering::Relaxed);
-        if let Some(cap) = self.throttle_cap {
-            if solved_blocks > cap {
-                solved_blocks = cap;
-            }
-        }
-        let sleep_ns = self.base_delay_ns.saturating_mul(solved_blocks);
-        let do_sleep = sleep_ns > 0;
-
-        loop {
-            if cancel.load(AtomicOrdering::Relaxed) {
-                return EngineStatus::Cancelled { hash_count };
-            }
-
-            // Compute hash using Bitcoin-style double Poseidon2
-            let hash = hash_from_nonce(ctx, current);
-            hash_count = hash_count.saturating_add(1);
-            batch_counter = batch_counter.saturating_add(1);
-
-            // Optional detailed debug of throttle progression within a job
-            #[allow(unused_variables)]
-            let _dbg_batch = batch_counter;
-
-            if is_valid_hash(ctx, hash) {
-                let work = current.to_big_endian();
-                // Increment solved-block counter so the NEXT block throttles more.
-                let _new_idx = self.job_index.fetch_add(1, AtomicOrdering::Relaxed) + 1;
-                {
-                    let capped = if let Some(cap) = self.throttle_cap {
-                        std::cmp::min(_new_idx, cap)
-                    } else {
-                        _new_idx
-                    };
-                    log::debug!(
-                        target: "miner",
-                        "manipulator throttle increment: solved_blocks={} (next sleep_ns_per_batch={}, cap={:?})",
-                        _new_idx,
-                        self.base_delay_ns.saturating_mul(capped),
-                        self.throttle_cap
-                    );
-                }
-                return EngineStatus::Found {
-                    candidate: Candidate {
-                        nonce: current,
-                        work,
-                        hash,
-                    },
-                    hash_count,
-                    origin: FoundOrigin::Cpu,
-                };
-            }
-
-            // Throttle after each batch to artificially slow down based on blocks solved so far.
-            if do_sleep && batch_counter >= self.step_batch {
-                std::thread::sleep(Duration::from_nanos(sleep_ns));
-                batch_counter = 0;
-            }
-
-            // Advance
-            if current < range.end {
-                current = step_nonce(current);
-            } else {
-                break EngineStatus::Exhausted { hash_count };
-            }
-        }
-    }
-}
-
-pub use {
-    BaselineCpuEngine as DefaultEngine, Candidate as EngineCandidate,
-    ChainManipulatorEngine as ChainEngine, FastCpuEngine as FastEngine, Range as EngineRange,
-};
+// Re-exports for convenience
+pub use {Candidate as EngineCandidate, Range as EngineRange};

 #[cfg(test)]
 mod tests {
@@ -381,33 +201,26 @@ mod tests {
    use primitive_types::U512;
    use std::sync::atomic::AtomicBool;

-    fn make_ctx() -> JobContext {
-        let header = [1u8; 32];
-        let difficulty = U512::from(1u64); // easy difficulty for "found" parity test
-        JobContext::new(header, difficulty)
-    }
-
    #[test]
    fn engine_returns_exhausted_when_no_solution_in_range() {
-        // Use a very hard difficulty to make solutions effectively impossible in a tiny range.
        let header = [2u8; 32];
-        let difficulty = U512::MAX;
+        let difficulty = U512::MAX; // impossible difficulty
        let ctx = JobContext::new(header, difficulty);

        let range = Range {
            start: U512::from(1u64),
-            end: U512::from(1000u64), // small range; probability of accidental match is negligible
+            end: U512::from(1000u64),
        };

        let cancel = AtomicBool::new(false);
-        let baseline = BaselineCpuEngine::new();
+        let cancel_check = AtomicBoolCancelCheck(&cancel);
+        let engine = FastCpuEngine::new(1000); // check every 1000 hashes

-        let status = baseline.search_range(&ctx, range.clone(), &cancel);
+        let status = engine.search_range(&ctx, range.clone(), &cancel_check);
        match status {
            EngineStatus::Exhausted { hash_count } => {
-                // Inclusive range length = end - start + 1
                let expected = (range.end - range.start + U512::one()).as_u64();
-                assert_eq!(hash_count, expected, "hash_count should equal range length");
+                assert_eq!(hash_count, expected);
            }
            other => panic!("expected Exhausted, got {other:?}"),
        }
@@ -415,19 +228,23 @@ mod tests {

    #[test]
    fn engine_respects_immediate_cancellation() {
-        let ctx = make_ctx();
+        let header = [1u8; 32];
+        let difficulty = U512::from(1u64);
+        let ctx = JobContext::new(header, difficulty);
+
        let range = Range {
            start: U512::from(0u64),
            end: U512::from(1_000_000u64),
        };
-        let cancel = AtomicBool::new(true); // cancelled before starting
-        let baseline = BaselineCpuEngine::new();

-        let status = baseline.search_range(&ctx, range, &cancel);
+        let cancel = AtomicBool::new(true); // pre-cancelled
+        let cancel_check = AtomicBoolCancelCheck(&cancel);
+        let engine = FastCpuEngine::new(1); // check every hash for immediate cancellation
+
+        let status = engine.search_range(&ctx, range, &cancel_check);
        match status {
            EngineStatus::Cancelled { hash_count } => {
-                // Cancellation was pre-set; allow zero or near-zero work depending on timing.
-                assert_eq!(hash_count, 0, "expected no work when cancelled immediately");
+                assert_eq!(hash_count, 0);
            }
            other => panic!("expected Cancelled, got {other:?}"),
        }
@@ -11,38 +11,36 @@ description = "GPU mining engine implementation for the Quantus External Miner"

 [features]
 # Default to the baseline/reference implementation; enable others as needed.
-default = ["baseline", "metrics"]
+default = ["baseline"]

 # Map engine features (pow-core dependency removed for now)
 baseline = []
 simd-poseidon2 = []
-metrics = ["dep:metrics"]

 [dependencies]
 engine-cpu = { path = "../engine-cpu" }
 pow-core = { path = "../pow-core" }
-metrics = { path = "../metrics", optional = true }
 primitive-types = { workspace = true }
 log = { workspace = true }
-thiserror = { workspace = true }
-anyhow = { workspace = true }
-qp-poseidon-core = { version = "1.0.2", features = ["p2"] }
-qp-poseidon-constants = { version = "1.0.2" }
-qp-plonky2 = { version = "1.1.1" }
+qp-poseidon-core = { workspace = true }
+qp-poseidon-constants = "1.1"
 qp-plonky2-field = { version = "1.1.1" }
+plonky2 = { package = "qp-plonky2", version = "1.1.3" }

-
-wgpu = { version = "27.0.1" }                              # GPU compute library
-futures = "0.3"                                            # For async executor
-bytemuck = "1.16"                                          # For buffer mapping
+wgpu = { version = "27.0.1" }                                              # GPU compute library
+tokio = { workspace = true, features = ["rt", "rt-multi-thread", "time"] } # Async runtime with timeout support
+futures = "0.3"                                                            # For async oneshot channels in test modules
+bytemuck = "1.16"                                                          # For buffer mapping
 rand = { workspace = true, features = ["std", "std_rng"] }
-rand_chacha = "0.3"                                        # For deterministic random test generation
+rand_chacha = "0.9"                                                        # For deterministic random test generation
+regex = "1"                                                                # For GPU name pattern matching

 [dev-dependencies]
 hex = { workspace = true }
 criterion = "0.5"
 rand = { workspace = true, features = ["std", "std_rng"] }
 env_logger.workspace = true
+tokio = { workspace = true, features = ["macros", "rt", "rt-multi-thread", "time"] } # Add macros for #[tokio::main]

 [[bench]]
 name = "gpu_engine_bench"
@@ -1,15 +1,24 @@
 use criterion::{black_box, criterion_group, criterion_main, Criterion};
-use engine_cpu::{FastCpuEngine, MinerEngine, Range};
+use engine_cpu::{AtomicBoolCancelCheck, FastCpuEngine, MinerEngine, Range};
 use engine_gpu::GpuEngine;
 use pow_core::JobContext;
 use primitive_types::U512;
 use rand::RngCore;
 use std::sync::atomic::AtomicBool;

+/// Drop thread-local wgpu buffers before `GpuEngine` is dropped. Criterion
+/// creates a fresh engine per group; without this, TLS buffers outlive the
+/// device and the next group panics (`Buffer[…] does not exist`).
+fn teardown_gpu(engine: GpuEngine) {
+    GpuEngine::clear_worker_resources();
+    drop(engine);
+}
+
 fn bench_cpu_vs_gpu_small(c: &mut Criterion) {
-    let cpu_engine = FastCpuEngine::new();
-    let gpu_engine = GpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    // Small range: 10K nonces - reasonable for benchmarking
    let small_range = Range {
@@ -24,14 +33,14 @@ fn bench_cpu_vs_gpu_small(c: &mut Criterion) {
    group.bench_function("cpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = cpu_engine.search_range(
                black_box(&ctx),
                black_box(small_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -40,26 +49,28 @@ fn bench_cpu_vs_gpu_small(c: &mut Criterion) {
    group.bench_function("gpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(small_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 fn bench_cpu_vs_gpu_medium(c: &mut Criterion) {
-    let cpu_engine = FastCpuEngine::new();
-    let gpu_engine = GpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    // Medium range: 100K nonces
    let medium_range = Range {
@@ -74,14 +85,14 @@ fn bench_cpu_vs_gpu_medium(c: &mut Criterion) {
    group.bench_function("cpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = cpu_engine.search_range(
                black_box(&ctx),
                black_box(medium_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -90,26 +101,28 @@ fn bench_cpu_vs_gpu_medium(c: &mut Criterion) {
    group.bench_function("gpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(medium_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 fn bench_cpu_vs_gpu_large(c: &mut Criterion) {
-    let cpu_engine = FastCpuEngine::new();
-    let gpu_engine = GpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    // Large range: 1M nonces - where GPU should really shine
    let large_range = Range {
@@ -124,14 +137,14 @@ fn bench_cpu_vs_gpu_large(c: &mut Criterion) {
    group.bench_function("cpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = cpu_engine.search_range(
                black_box(&ctx),
                black_box(large_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -140,26 +153,28 @@ fn bench_cpu_vs_gpu_large(c: &mut Criterion) {
    group.bench_function("gpu", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(large_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 fn bench_solution_finding(c: &mut Criterion) {
-    let cpu_engine = FastCpuEngine::new();
-    let gpu_engine = GpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    // Range where we expect to find solutions quickly
    let solution_range = Range {
@@ -174,14 +189,14 @@ fn bench_solution_finding(c: &mut Criterion) {
    group.bench_function("cpu_find_solution", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(10_000u64); // Easy difficulty - should find solution
            let ctx = JobContext::new(header, difficulty);

            let result = cpu_engine.search_range(
                black_box(&ctx),
                black_box(solution_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -190,26 +205,28 @@ fn bench_solution_finding(c: &mut Criterion) {
    group.bench_function("gpu_find_solution", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(10_000u64); // Easy difficulty - should find solution
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(solution_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 fn bench_throughput_per_second(c: &mut Criterion) {
-    let cpu_engine = FastCpuEngine::new();
-    let gpu_engine = GpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    // Fixed time benchmark - see how many hashes we can do in 1 second
    let throughput_range = Range {
@@ -224,14 +241,14 @@ fn bench_throughput_per_second(c: &mut Criterion) {
    group.bench_function("cpu_throughput", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = cpu_engine.search_range(
                black_box(&ctx),
                black_box(throughput_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -240,25 +257,27 @@ fn bench_throughput_per_second(c: &mut Criterion) {
    group.bench_function("gpu_throughput", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(throughput_range.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 fn bench_gpu_batch_efficiency(c: &mut Criterion) {
-    let gpu_engine = GpuEngine::new();
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
    let cancel_flag = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel_flag);

    let mut group = c.benchmark_group("gpu_batch_sizes");
    group.sample_size(10);
@@ -283,14 +302,14 @@ fn bench_gpu_batch_efficiency(c: &mut Criterion) {
    group.bench_function("gpu_1k_batch", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX);
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(small_batch.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -299,14 +318,14 @@ fn bench_gpu_batch_efficiency(c: &mut Criterion) {
    group.bench_function("gpu_50k_batch", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX);
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(medium_batch.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
@@ -315,20 +334,21 @@ fn bench_gpu_batch_efficiency(c: &mut Criterion) {
    group.bench_function("gpu_500k_batch", |b| {
        b.iter(|| {
            let mut header = [0u8; 32];
-            rand::thread_rng().fill_bytes(&mut header);
+            rand::rng().fill_bytes(&mut header);
            let difficulty = U512::from(u64::MAX);
            let ctx = JobContext::new(header, difficulty);

            let result = gpu_engine.search_range(
                black_box(&ctx),
                black_box(large_batch.clone()),
-                black_box(&cancel_flag),
+                black_box(&cancel_check),
            );
            black_box(result)
        })
    });

    group.finish();
+    teardown_gpu(gpu_engine);
 }

 criterion_group!(
@@ -0,0 +1,63 @@
+use engine_cpu::{AtomicBoolCancelCheck, EngineStatus, MinerEngine, Range};
+use engine_gpu::GpuEngine;
+use primitive_types::U512;
+use rand::RngCore;
+use std::sync::atomic::AtomicBool;
+
+fn main() {
+    env_logger::init();
+    let jobs: usize = std::env::args()
+        .nth(1)
+        .map(|s| s.parse().expect("job count"))
+        .unwrap_or(25);
+
+    let engine = GpuEngine::try_new(1_000_000, 0, false).expect("GPU init failed");
+    let cancel_flag = AtomicBool::new(false);
+    let cancel = AtomicBoolCancelCheck(&cancel_flag);
+    let mut rng = rand::rng();
+
+    let mut found = 0usize;
+    for job in 0..jobs {
+        let mut header = [0u8; 32];
+        rng.fill_bytes(&mut header);
+        let difficulty = U512::from(100_000u64);
+        let ctx = engine.prepare_context(header, difficulty);
+
+        let start = if job == 0 {
+            // Cross a 2^256 boundary: the high nonce half changes mid-range,
+            // exercising the midstate batch clamp.
+            (U512::from(3u64) << 256) - U512::from(1_000u64)
+        } else {
+            let mut start_bytes = [0u8; 64];
+            rng.fill_bytes(&mut start_bytes);
+            // Keep clear of the very top so range arithmetic cannot wrap
+            start_bytes[0] = 0;
+            U512::from_big_endian(&start_bytes)
+        };
+        let range = Range {
+            start,
+            end: start + U512::from(10_000_000u64),
+        };
+
+        match engine.search_range(&ctx, range.clone(), &cancel) {
+            EngineStatus::Found { candidate, .. } => {
+                let cpu_hash = pow_core::hash_from_nonce(&ctx, candidate.nonce);
+                assert_eq!(
+                    cpu_hash, candidate.hash,
+                    "job {job}: GPU hash != CPU hash for nonce {}",
+                    candidate.nonce
+                );
+                assert!(cpu_hash < ctx.target, "job {job}: hash not below target");
+                assert!(
+                    candidate.nonce >= range.start && candidate.nonce <= range.end,
+                    "job {job}: nonce outside range"
+                );
+                found += 1;
+            }
+            EngineStatus::Exhausted { .. } => {}
+            other => panic!("job {job}: unexpected status {other:?}"),
+        }
+    }
+    assert!(found > 0, "no solutions found across {jobs} jobs");
+    println!("PARITY OK: {found}/{jobs} jobs found solutions, all verified against CPU");
+}
@@ -0,0 +1,30 @@
+fn main() {
+    let instance = wgpu::Instance::new(&wgpu::InstanceDescriptor {
+        backends: wgpu::Backends::PRIMARY,
+        ..Default::default()
+    });
+    for adapter in instance.enumerate_adapters(wgpu::Backends::PRIMARY) {
+        let info = adapter.get_info();
+        let features = adapter.features();
+        println!("{} ({:?}, {:?})", info.name, info.device_type, info.backend);
+        println!(
+            "  SHADER_INT64: {}",
+            features.contains(wgpu::Features::SHADER_INT64)
+        );
+        println!(
+            "  SUBGROUP: {}",
+            features.contains(wgpu::Features::SUBGROUP)
+        );
+        println!(
+            "  TIMESTAMP_QUERY: {}",
+            features.contains(wgpu::Features::TIMESTAMP_QUERY)
+        );
+        let limits = adapter.limits();
+        println!(
+            "  max_workgroup_size_x: {}, max_invocations: {}, max_workgroups_per_dim: {}",
+            limits.max_compute_workgroup_size_x,
+            limits.max_compute_invocations_per_workgroup,
+            limits.max_compute_workgroups_per_dimension
+        );
+    }
+}
@@ -0,0 +1,52 @@
+use engine_cpu::{AtomicBoolCancelCheck, MinerEngine, Range};
+use engine_gpu::GpuEngine;
+use primitive_types::U512;
+use std::sync::atomic::AtomicBool;
+
+fn main() {
+    env_logger::init();
+    let args: Vec<String> = std::env::args().collect();
+    let total: u64 = args
+        .get(1)
+        .map(|s| s.parse().expect("total nonces"))
+        .unwrap_or(4_000_000);
+    let batch: u32 = args
+        .get(2)
+        .map(|s| s.parse().expect("batch size"))
+        .unwrap_or(1_000_000);
+
+    let engine = GpuEngine::try_new(batch, 0, false).expect("GPU init failed");
+    let cancel_flag = AtomicBool::new(false);
+    let cancel = AtomicBoolCancelCheck(&cancel_flag);
+
+    let header = [42u8; 32];
+    let ctx = engine.prepare_context(header, U512::from(u64::MAX));
+
+    let range = Range {
+        start: U512::from(1u64) << 200,
+        end: (U512::from(1u64) << 200) + U512::from(total - 1),
+    };
+
+    // Warmup
+    let warm = Range {
+        start: range.start,
+        end: range.start + U512::from(200_000u64),
+    };
+    engine.search_range(&ctx, warm, &cancel);
+
+    let start = std::time::Instant::now();
+    let status = engine.search_range(&ctx, range, &cancel);
+    let elapsed = start.elapsed().as_secs_f64();
+
+    let hashes = match status {
+        engine_cpu::EngineStatus::Exhausted { hash_count } => hash_count,
+        engine_cpu::EngineStatus::Found { hash_count, .. } => hash_count,
+        other => panic!("unexpected status: {other:?}"),
+    };
+    println!(
+        "{} hashes in {:.3}s = {:.3} MH/s",
+        hashes,
+        elapsed,
+        hashes as f64 / elapsed / 1e6
+    );
+}
@@ -0,0 +1,229 @@
+use pow_core::{hash_from_nonce, mining_midstate, JobContext};
+use primitive_types::U512;
+use std::time::Instant;
+
+fn u512_to_u32s_le(v: U512) -> [u32; 16] {
+    let bytes = v.to_little_endian();
+    let mut out = [0u32; 16];
+    for i in 0..16 {
+        out[i] = u32::from_le_bytes(bytes[i * 4..(i + 1) * 4].try_into().unwrap());
+    }
+    out
+}
+
+struct Runner {
+    device: wgpu::Device,
+    queue: wgpu::Queue,
+    pipeline: wgpu::ComputePipeline,
+    results: wgpu::Buffer,
+    midstate: wgpu::Buffer,
+    start_nonce: wgpu::Buffer,
+    target: wgpu::Buffer,
+    cfg: wgpu::Buffer,
+    staging: wgpu::Buffer,
+}
+
+impl Runner {
+    async fn new(trusted: bool) -> Self {
+        let instance = wgpu::Instance::new(&wgpu::InstanceDescriptor {
+            backends: wgpu::Backends::PRIMARY,
+            ..Default::default()
+        });
+        let adapter = instance
+            .request_adapter(&wgpu::RequestAdapterOptions::default())
+            .await
+            .expect("no adapter");
+        assert!(
+            adapter.features().contains(wgpu::Features::SHADER_INT64),
+            "SHADER_INT64 required"
+        );
+        let (device, queue) = adapter
+            .request_device(&wgpu::DeviceDescriptor {
+                label: None,
+                required_features: wgpu::Features::SHADER_INT64,
+                ..Default::default()
+            })
+            .await
+            .unwrap();
+
+        let kernel = engine_gpu::Kernel::for_adapter(&adapter);
+        let desc = wgpu::ShaderModuleDescriptor {
+            label: Some(kernel.label()),
+            source: wgpu::ShaderSource::Wgsl(kernel.source().into()),
+        };
+        let shader = if trusted {
+            unsafe {
+                device.create_shader_module_trusted(desc, wgpu::ShaderRuntimeChecks::unchecked())
+            }
+        } else {
+            device.create_shader_module(desc)
+        };
+        let pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+            label: None,
+            layout: None,
+            module: &shader,
+            entry_point: Some("mining_main"),
+            compilation_options: Default::default(),
+            cache: None,
+        });
+        let mk = |size: u64, usage| {
+            device.create_buffer(&wgpu::BufferDescriptor {
+                label: None,
+                size,
+                usage,
+                mapped_at_creation: false,
+            })
+        };
+        use wgpu::BufferUsages as U;
+        Runner {
+            pipeline,
+            results: mk(132, U::STORAGE | U::COPY_SRC | U::COPY_DST),
+            midstate: mk(96, U::STORAGE | U::COPY_DST),
+            start_nonce: mk(64, U::STORAGE | U::COPY_DST),
+            target: mk(64, U::STORAGE | U::COPY_DST),
+            cfg: mk(12, U::STORAGE | U::COPY_DST),
+            staging: mk(132, U::MAP_READ | U::COPY_DST),
+            device,
+            queue,
+        }
+    }
+
+    fn run_batch(&self, header: [u8; 32], start: U512, batch: u32, target: U512) -> f64 {
+        let nonce_be = start.to_big_endian();
+        let mid = mining_midstate(header, nonce_be[..32].try_into().unwrap());
+        let mut mid_u32 = [0u32; 24];
+        for (i, f) in mid.iter().enumerate() {
+            mid_u32[2 * i] = *f as u32;
+            mid_u32[2 * i + 1] = (*f >> 32) as u32;
+        }
+        self.queue
+            .write_buffer(&self.midstate, 0, bytemuck::cast_slice(&mid_u32));
+        self.queue.write_buffer(
+            &self.start_nonce,
+            0,
+            bytemuck::cast_slice(&u512_to_u32s_le(start)),
+        );
+        self.queue.write_buffer(
+            &self.target,
+            0,
+            bytemuck::cast_slice(&u512_to_u32s_le(target)),
+        );
+        self.queue
+            .write_buffer(&self.cfg, 0, bytemuck::cast_slice(&[batch, 1u32, batch]));
+        self.queue.write_buffer(&self.results, 0, &[0u8; 132]);
+
+        let layout = self.pipeline.get_bind_group_layout(0);
+        let entries: Vec<wgpu::BindGroupEntry<'_>> = [
+            (0, &self.results),
+            (1, &self.midstate),
+            (2, &self.start_nonce),
+            (3, &self.target),
+            (4, &self.cfg),
+        ]
+        .iter()
+        .map(|(i, b)| wgpu::BindGroupEntry {
+            binding: *i,
+            resource: b.as_entire_binding(),
+        })
+        .collect();
+        let bind_group = self.device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: None,
+            layout: &layout,
+            entries: &entries,
+        });
+
+        let t0 = Instant::now();
+        let mut encoder = self
+            .device
+            .create_command_encoder(&wgpu::CommandEncoderDescriptor { label: None });
+        {
+            let mut cpass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
+                label: None,
+                timestamp_writes: None,
+            });
+            cpass.set_pipeline(&self.pipeline);
+            cpass.set_bind_group(0, &bind_group, &[]);
+            cpass.dispatch_workgroups(batch.div_ceil(256), 1, 1);
+        }
+        encoder.copy_buffer_to_buffer(&self.results, 0, &self.staging, 0, 132);
+        self.queue.submit(Some(encoder.finish()));
+
+        let slice = self.staging.slice(..);
+        let (tx, rx) = std::sync::mpsc::channel();
+        slice.map_async(wgpu::MapMode::Read, move |r| tx.send(r).unwrap());
+        loop {
+            let _ = self.device.poll(wgpu::PollType::Wait {
+                submission_index: None,
+                timeout: None,
+            });
+            if rx.try_recv().is_ok() {
+                break;
+            }
+        }
+        let elapsed = t0.elapsed().as_secs_f64();
+        self.staging.unmap();
+        elapsed
+    }
+
+    fn read_results(&self) -> [u32; 33] {
+        let slice = self.staging.slice(..);
+        let (tx, rx) = std::sync::mpsc::channel();
+        slice.map_async(wgpu::MapMode::Read, move |r| tx.send(r).unwrap());
+        let _ = self.device.poll(wgpu::PollType::Wait {
+            submission_index: None,
+            timeout: None,
+        });
+        rx.recv().unwrap().unwrap();
+        let data = slice.get_mapped_range();
+        let mut out = [0u32; 33];
+        out.copy_from_slice(bytemuck::cast_slice(&data));
+        drop(data);
+        self.staging.unmap();
+        out
+    }
+}
+
+fn main() {
+    let batch: u32 = std::env::args()
+        .nth(1)
+        .map(|s| s.parse().unwrap())
+        .unwrap_or(8_000_000);
+    let iters = 4u32;
+    let rt = tokio::runtime::Runtime::new().unwrap();
+
+    for trusted in [false, true] {
+        let label = if trusted {
+            "trusted (no checks)"
+        } else {
+            "checked (default)"
+        };
+        let runner = rt.block_on(Runner::new(trusted));
+
+        let header = [9u8; 32];
+        let ctx = JobContext::new(header, U512::one());
+        let start = (U512::from(7u64) << 300) | U512::from(123456789u64);
+        runner.run_batch(header, start, 256, ctx.target);
+        let r = runner.read_results();
+        assert_eq!(r[0], 1, "{label}: no solution with target=MAX");
+        let nonce = U512::from_little_endian(bytemuck::cast_slice(&r[1..17]));
+        let hash = U512::from_little_endian(bytemuck::cast_slice(&r[17..33]));
+        assert_eq!(hash, hash_from_nonce(&ctx, nonce), "{label}: hash mismatch");
+        println!("{label}: correctness OK");
+
+        let header = [42u8; 32];
+        let start = U512::from(1u64) << 200;
+        let target = U512::one();
+        runner.run_batch(header, start, batch, target);
+        let mut total = 0.0;
+        for i in 0..iters {
+            total += runner.run_batch(
+                header,
+                start + U512::from((i as u64 + 1) * batch as u64),
+                batch,
+                target,
+            );
+        }
+        let mhs = (batch as f64 * iters as f64) / total / 1e6;
+        println!("{label}: {batch} nonces x{iters} in {total:.3}s = {mhs:.3} MH/s");
+    }
+}
@@ -1,4 +1,4 @@
-use engine_cpu::{BaselineCpuEngine, EngineStatus, MinerEngine, Range};
+use engine_cpu::{AtomicBoolCancelCheck, EngineStatus, FastCpuEngine, MinerEngine, Range};
 use engine_gpu::GpuEngine;
 use primitive_types::U512;
 use std::sync::atomic::AtomicBool;
@@ -16,16 +16,17 @@ fn main() {
    // Use a fixed header and easy difficulty (1) so any nonce is valid
    let header = [1u8; 32];
    let difficulty = U512::from(u64::MAX); // High difficulty - no solutions expected
-    let cpu_engine = BaselineCpuEngine::new();
+    let cpu_engine = FastCpuEngine::new(10_000);
    let ctx = cpu_engine.prepare_context(header, difficulty);

    log::info!("Context prepared. Difficulty: {}", difficulty);

    let cancel = AtomicBool::new(false);
+    let cancel_check = AtomicBoolCancelCheck(&cancel);

    // 3. Verify with GPU engine
    log::info!("Initializing GPU engine...");
-    let gpu_engine = GpuEngine::new();
+    let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");

    // Search a small range around the valid nonce
    let gpu_range = Range {
@@ -39,7 +40,7 @@ fn main() {
        gpu_range.end
    );
    let start = std::time::Instant::now();
-    let gpu_result = gpu_engine.search_range(&ctx, gpu_range, &cancel);
+    let gpu_result = gpu_engine.search_range(&ctx, gpu_range, &cancel_check);
    let elapsed = start.elapsed();

    log::info!("GPU search took {:?}", elapsed);
@@ -55,6 +56,9 @@ fn main() {
        EngineStatus::Cancelled { .. } => {
            log::error!("FAILURE: GPU search cancelled!");
        }
+        EngineStatus::DeviceLost { .. } => {
+            log::error!("FAILURE: GPU device lost!");
+        }
        EngineStatus::Running { .. } => {
            log::error!("FAILURE: GPU returned Running status!");
        }
@@ -7,6 +7,7 @@ use wgpu::util::DeviceExt;
 pub async fn test_end_to_end_mining(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    println!("Running End-to-End Mining Test...");

@@ -28,15 +29,17 @@ pub async fn test_end_to_end_mining(

    // 4. Run GPU Mining for this specific nonce

-    // Header Buffer
-    let mut header_u32s = [0u32; 8];
-    for (i, item) in header_u32s.iter_mut().enumerate() {
-        let chunk = &ctx.header[i * 4..(i + 1) * 4];
-        *item = u32::from_le_bytes([chunk[0], chunk[1], chunk[2], chunk[3]]);
+    // Midstate Buffer (header + high nonce half absorbed on CPU)
+    let nonce_be = nonce_val.to_big_endian();
+    let midstate = pow_core::mining_midstate(ctx.header, nonce_be[..32].try_into().unwrap());
+    let mut midstate_u32s = [0u32; 24];
+    for (i, felt) in midstate.iter().enumerate() {
+        midstate_u32s[2 * i] = *felt as u32;
+        midstate_u32s[2 * i + 1] = (*felt >> 32) as u32;
    }
-    let header_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
-        label: Some("Header Buffer"),
-        contents: bytemuck::cast_slice(&header_u32s),
+    let midstate_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("Midstate Buffer"),
+        contents: bytemuck::cast_slice(&midstate_u32s),
        usage: wgpu::BufferUsages::STORAGE,
    });

@@ -82,11 +85,11 @@ pub async fn test_end_to_end_mining(
    let zeros = vec![0u8; results_size];
    queue.write_buffer(&results_buffer, 0, &zeros);

-    // Dispatch config buffer: [total_threads, nonces_per_thread, work_per_batch, threads_per_workgroup]
-    let dispatch_config_data: [u32; 4] = [256, 1, 1, 256];
+    // Dispatch config buffer: [total_threads, nonces_per_thread, total_nonces]
+    let dispatch_config_data: [u32; 3] = [256, 1, 256];
    let dispatch_config_buffer = device.create_buffer(&wgpu::BufferDescriptor {
        label: Some("Dispatch Config Buffer"),
-        size: 16,
+        size: 12,
        usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
        mapped_at_creation: false,
    });
@@ -96,11 +99,9 @@ pub async fn test_end_to_end_mining(
        bytemuck::cast_slice(&dispatch_config_data),
    );

-    // Load Shader
-    let shader_source = include_str!("mining.wgsl");
    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
        label: Some("Mining Shader"),
-        source: wgpu::ShaderSource::Wgsl(shader_source.into()),
+        source: wgpu::ShaderSource::Wgsl(shader_src.into()),
    });

    let pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
@@ -123,7 +124,7 @@ pub async fn test_end_to_end_mining(
            },
            wgpu::BindGroupEntry {
                binding: 1,
-                resource: header_buffer.as_entire_binding(),
+                resource: midstate_buffer.as_entire_binding(),
            },
            wgpu::BindGroupEntry {
                binding: 2,
@@ -193,18 +194,27 @@ pub async fn test_end_to_end_mining(
            let found_nonce = U512::from_little_endian(&nonce_bytes);
            println!("GPU Nonce: {}", found_nonce);

-            assert_eq!(found_nonce, nonce_val, "Nonce mismatch!");
-
-            // Parse hash
+            // Parse hash from GPU result
            let mut hash_bytes = [0u8; 64];
            for i in 0..16 {
                let bytes = result_u32s[17 + i].to_le_bytes();
                hash_bytes[i * 4..(i + 1) * 4].copy_from_slice(&bytes);
            }
-            let found_hash = U512::from_little_endian(&hash_bytes);
-            println!("GPU Hash: {:x}", found_hash);
+            let gpu_hash = U512::from_little_endian(&hash_bytes);
+            println!("GPU Hash: {:x}", gpu_hash);
+
+            // Verify: compute hash on CPU for the found nonce and compare
+            let cpu_verified_hash = hash_from_nonce(&ctx, found_nonce);
+            println!("CPU verified hash: {:x}", cpu_verified_hash);
+
+            assert_eq!(
+                gpu_hash, cpu_verified_hash,
+                "GPU hash doesn't match CPU verification for the found nonce!"
+            );
+
+            // Verify hash is below target (should always pass with difficulty 1)
+            assert!(gpu_hash <= ctx.target, "Hash is not below target!");

-            assert_eq!(found_hash, expected_hash, "Hash mismatch!");
            println!("✅ End-to-End Test Passed!");
        } else {
            println!("❌ GPU did not find solution (should have passed with MAX target)");
@@ -0,0 +1,748 @@
+//! GPU tier detection using regex patterns
+//!
+//! This module provides a table-driven approach to GPU detection,
+//! using regex patterns with word boundaries to avoid substring matching issues.
+
+use regex::Regex;
+use std::sync::LazyLock;
+
+/// GPU tier configuration
+struct GpuTier {
+    /// Regex pattern to match GPU name (case-insensitive)
+    pattern: &'static str,
+    /// Human-readable tier name
+    name: &'static str,
+    /// Divisor for max_workgroups (higher = more conservative)
+    workgroup_divisor: u32,
+    /// Minimum workgroups to use
+    min_workgroups: u32,
+}
+
+/// Compiled GPU tier with regex
+struct CompiledGpuTier {
+    regex: Regex,
+    name: &'static str,
+    workgroup_divisor: u32,
+    min_workgroups: u32,
+}
+
+impl CompiledGpuTier {
+    fn from_tier(tier: &GpuTier) -> Self {
+        Self {
+            regex: Regex::new(&format!("(?i){}", tier.pattern)).expect("Invalid GPU tier regex"),
+            name: tier.name,
+            workgroup_divisor: tier.workgroup_divisor,
+            min_workgroups: tier.min_workgroups,
+        }
+    }
+}
+
+// GPU tiers are checked in order - first match wins
+// Use word boundaries (\b) to avoid substring issues like "550" matching "5500"
+
+const NVIDIA_TIERS: &[GpuTier] = &[
+    // Blackwell (RTX 50 series)
+    GpuTier {
+        pattern: r"\b50[89]0\b",
+        name: "NVIDIA RTX 50 Flagship (Blackwell)",
+        workgroup_divisor: 6,
+        min_workgroups: 5120,
+    },
+    GpuTier {
+        pattern: r"\b50[67]0\b|rtx 50",
+        name: "NVIDIA RTX 50 (Blackwell)",
+        workgroup_divisor: 7,
+        min_workgroups: 4608,
+    },
+    // Ada Lovelace (RTX 40 series)
+    GpuTier {
+        pattern: r"\b40[89]0\b",
+        name: "NVIDIA RTX 40 Flagship (Ada)",
+        workgroup_divisor: 8,
+        min_workgroups: 4096,
+    },
+    GpuTier {
+        pattern: r"\b40[67]0\b|rtx 40",
+        name: "NVIDIA RTX 40 (Ada)",
+        workgroup_divisor: 10,
+        min_workgroups: 3072,
+    },
+    // Ampere/Turing (RTX 30/20 series)
+    GpuTier {
+        pattern: r"\b30[5-9]0\b|\b20[6-8]0\b|rtx 30|rtx 20",
+        name: "NVIDIA RTX 30/20 (Ampere/Turing)",
+        workgroup_divisor: 12,
+        min_workgroups: 2048,
+    },
+    // Turing/Pascal (GTX 16/10 series)
+    GpuTier {
+        pattern: r"\b16[56]0\b|\b10[3-8]0\b|gtx 16|gtx 10",
+        name: "NVIDIA GTX 16/10 (Turing/Pascal)",
+        workgroup_divisor: 16,
+        min_workgroups: 1024,
+    },
+    // Maxwell (GTX 900 series)
+    GpuTier {
+        pattern: r"\b9[5-8]0\b|gtx 9",
+        name: "NVIDIA GTX 900 (Maxwell)",
+        workgroup_divisor: 18,
+        min_workgroups: 768,
+    },
+    // Kepler/Maxwell (GTX 700 series)
+    GpuTier {
+        pattern: r"\b7[5-8]0\b|gtx 7",
+        name: "NVIDIA GTX 700 (Kepler/Maxwell)",
+        workgroup_divisor: 20,
+        min_workgroups: 512,
+    },
+    // Legacy GTX
+    GpuTier {
+        pattern: r"gtx [456]",
+        name: "NVIDIA GTX Legacy (Fermi/Kepler)",
+        workgroup_divisor: 24,
+        min_workgroups: 384,
+    },
+    GpuTier {
+        pattern: r"\bgtx\b",
+        name: "NVIDIA GTX (Unknown)",
+        workgroup_divisor: 20,
+        min_workgroups: 512,
+    },
+    // Mobile
+    GpuTier {
+        pattern: r"\bmx[1-5]\d0|geforce mx",
+        name: "NVIDIA MX (Mobile)",
+        workgroup_divisor: 24,
+        min_workgroups: 384,
+    },
+    GpuTier {
+        pattern: r"geforce gt\b|\bgt \d{3}\b",
+        name: "NVIDIA GT (Entry-Level)",
+        workgroup_divisor: 28,
+        min_workgroups: 256,
+    },
+    // Professional
+    GpuTier {
+        pattern: r"quadro|rtx a\d|tesla|\ba100\b|\bh100\b|\bl4\b",
+        name: "NVIDIA Quadro/Professional",
+        workgroup_divisor: 10,
+        min_workgroups: 2560,
+    },
+];
+
+const AMD_TIERS: &[GpuTier] = &[
+    // RDNA 4 (RX 9000 series)
+    GpuTier {
+        pattern: r"rx 9|\b90[78]0\b",
+        name: "AMD RX 9000 (RDNA 4)",
+        workgroup_divisor: 8,
+        min_workgroups: 4096,
+    },
+    // RDNA 3 Discrete
+    GpuTier {
+        pattern: r"\b7900\b",
+        name: "AMD RX 7900 (RDNA 3 Flagship)",
+        workgroup_divisor: 9,
+        min_workgroups: 3584,
+    },
+    GpuTier {
+        pattern: r"rx 7|\b7[6-8]00\b",
+        name: "AMD RX 7000 (RDNA 3)",
+        workgroup_divisor: 10,
+        min_workgroups: 3072,
+    },
+    // RDNA 3 APUs - check before discrete to avoid substring match
+    GpuTier {
+        pattern: r"\b780m\b|radeon 780m",
+        name: "AMD Radeon 780M (RDNA 3 APU)",
+        workgroup_divisor: 12,
+        min_workgroups: 2048,
+    },
+    GpuTier {
+        pattern: r"\b7[46]0m\b|radeon 7[46]0m",
+        name: "AMD Radeon 7x0M (RDNA 3 APU)",
+        workgroup_divisor: 16,
+        min_workgroups: 1024,
+    },
+    // RDNA 2 Discrete
+    GpuTier {
+        pattern: r"\b6[89][05]0\b",
+        name: "AMD RX 6900/6800 (RDNA 2 Flagship)",
+        workgroup_divisor: 12,
+        min_workgroups: 2560,
+    },
+    GpuTier {
+        pattern: r"\b6[67][05]0\b",
+        name: "AMD RX 6700/6600 (RDNA 2)",
+        workgroup_divisor: 14,
+        min_workgroups: 2048,
+    },
+    GpuTier {
+        pattern: r"\b6[45]00\b",
+        name: "AMD RX 6500/6400 (RDNA 2 Entry)",
+        workgroup_divisor: 22,
+        min_workgroups: 512,
+    },
+    GpuTier {
+        pattern: r"rx 6\d{3}",
+        name: "AMD RX 6000 (RDNA 2)",
+        workgroup_divisor: 14,
+        min_workgroups: 2048,
+    },
+    // RDNA 2 APUs - check before discrete
+    GpuTier {
+        pattern: r"\b680m\b|radeon 680m",
+        name: "AMD Radeon 680M (RDNA 2 APU)",
+        workgroup_divisor: 16,
+        min_workgroups: 1536,
+    },
+    GpuTier {
+        pattern: r"\b6[16]0m\b|radeon 6[16]0m",
+        name: "AMD Radeon 6x0M (RDNA 2 APU)",
+        workgroup_divisor: 22,
+        min_workgroups: 768,
+    },
+    // RDNA 1 (RX 5000 series) - 4-digit patterns to avoid matching Polaris 3-digit
+    GpuTier {
+        pattern: r"\b5700\b",
+        name: "AMD RX 5700 (RDNA 1)",
+        workgroup_divisor: 16,
+        min_workgroups: 1536,
+    },
+    GpuTier {
+        pattern: r"\b5[56]00\b|rx 5\d{3}",
+        name: "AMD RX 5000 (RDNA 1)",
+        workgroup_divisor: 18,
+        min_workgroups: 1024,
+    },
+    // Polaris (RX 400/500 series) - 3-digit models with boundaries
+    GpuTier {
+        pattern: r"rx [45]\d0\b|\b[45][6-9]0\b|\b590\b|rx 5.0x",
+        name: "AMD RX 500/400 (Polaris)",
+        workgroup_divisor: 20,
+        min_workgroups: 768,
+    },
+    // Vega
+    GpuTier {
+        pattern: r"radeon vii\b",
+        name: "AMD Radeon VII (Vega 20)",
+        workgroup_divisor: 12,
+        min_workgroups: 2048,
+    },
+    GpuTier {
+        pattern: r"vega\s*64",
+        name: "AMD Vega 64 (Discrete)",
+        workgroup_divisor: 14,
+        min_workgroups: 1536,
+    },
+    GpuTier {
+        pattern: r"vega\s*56",
+        name: "AMD Vega 56 (Discrete)",
+        workgroup_divisor: 16,
+        min_workgroups: 1280,
+    },
+    GpuTier {
+        pattern: r"vega",
+        name: "AMD Vega (APU)",
+        workgroup_divisor: 28,
+        min_workgroups: 384,
+    },
+    // GCN
+    GpuTier {
+        pattern: r"fury|nano",
+        name: "AMD R9 Fury/Nano (Fiji)",
+        workgroup_divisor: 16,
+        min_workgroups: 1280,
+    },
+    GpuTier {
+        pattern: r"r9.*(3[89]0|2[89]0)|\b[23][89]0x?\b",
+        name: "AMD R9 (GCN)",
+        workgroup_divisor: 20,
+        min_workgroups: 768,
+    },
+    GpuTier {
+        pattern: r"r7.*(3[67]0|2[67]0)|\b[23][67]0x?\b",
+        name: "AMD R7 (GCN)",
+        workgroup_divisor: 22,
+        min_workgroups: 512,
+    },
+    // Professional
+    GpuTier {
+        pattern: r"radeon pro|instinct mi|mi[123]\d0|firepro|w[567]\d{3}",
+        name: "AMD Radeon Pro/Instinct",
+        workgroup_divisor: 10,
+        min_workgroups: 2560,
+    },
+    // OEM/APU fallbacks
+    GpuTier {
+        pattern: r"radeon\s*\(tm\)\s*[67]\d{2}\b|radeon [67]\d{2}\b",
+        name: "AMD Radeon OEM (Polaris Rebrand)",
+        workgroup_divisor: 24,
+        min_workgroups: 512,
+    },
+    GpuTier {
+        pattern: r"radeon\s*(\(tm\)\s*)?graphics",
+        name: "AMD Radeon Graphics (APU)",
+        workgroup_divisor: 26,
+        min_workgroups: 384,
+    },
+];
+
+const INTEL_TIERS: &[GpuTier] = &[
+    // Battlemage (Arc B-Series)
+    GpuTier {
+        pattern: r"arc b|\bb5[78]0\b",
+        name: "Intel Arc B-Series (Battlemage)",
+        workgroup_divisor: 10,
+        min_workgroups: 2560,
+    },
+    // Alchemist Mobile - check BEFORE desktop (a770m contains a770)
+    GpuTier {
+        pattern: r"\ba7[37]0m\b",
+        name: "Intel Arc A7 Mobile (Alchemist)",
+        workgroup_divisor: 14,
+        min_workgroups: 1536,
+    },
+    GpuTier {
+        pattern: r"\ba5[57]0m\b",
+        name: "Intel Arc A5 Mobile (Alchemist)",
+        workgroup_divisor: 16,
+        min_workgroups: 1024,
+    },
+    GpuTier {
+        pattern: r"\ba3[57]0m\b",
+        name: "Intel Arc A3 Mobile (Alchemist)",
+        workgroup_divisor: 20,
+        min_workgroups: 512,
+    },
+    // Alchemist Desktop
+    GpuTier {
+        pattern: r"\ba7[57]0\b",
+        name: "Intel Arc A7 (Alchemist)",
+        workgroup_divisor: 12,
+        min_workgroups: 2048,
+    },
+    GpuTier {
+        pattern: r"\ba580\b",
+        name: "Intel Arc A5 (Alchemist)",
+        workgroup_divisor: 14,
+        min_workgroups: 1536,
+    },
+    GpuTier {
+        pattern: r"\ba3[18]0\b",
+        name: "Intel Arc A3 (Alchemist)",
+        workgroup_divisor: 18,
+        min_workgroups: 768,
+    },
+    GpuTier {
+        pattern: r"arc a|\barc\b",
+        name: "Intel Arc (Unknown)",
+        workgroup_divisor: 16,
+        min_workgroups: 1024,
+    },
+    // Integrated
+    GpuTier {
+        pattern: r"iris xe max",
+        name: "Intel Iris Xe Max (Discrete)",
+        workgroup_divisor: 20,
+        min_workgroups: 512,
+    },
+    GpuTier {
+        pattern: r"iris xe",
+        name: "Intel Iris Xe (Integrated)",
+        workgroup_divisor: 24,
+        min_workgroups: 384,
+    },
+    GpuTier {
+        pattern: r"iris pro",
+        name: "Intel Iris Pro (Integrated)",
+        workgroup_divisor: 26,
+        min_workgroups: 256,
+    },
+    GpuTier {
+        pattern: r"iris plus|iris\b",
+        name: "Intel Iris Plus (Integrated)",
+        workgroup_divisor: 26,
+        min_workgroups: 320,
+    },
+    GpuTier {
+        pattern: r"uhd.*(7\d{2}|graphics 7)",
+        name: "Intel UHD 700 (Integrated)",
+        workgroup_divisor: 26,
+        min_workgroups: 320,
+    },
+    GpuTier {
+        pattern: r"uhd.*(6\d{2}|graphics 6)",
+        name: "Intel UHD 600 (Integrated)",
+        workgroup_divisor: 28,
+        min_workgroups: 256,
+    },
+    GpuTier {
+        pattern: r"uhd",
+        name: "Intel UHD Graphics (Integrated)",
+        workgroup_divisor: 28,
+        min_workgroups: 256,
+    },
+    GpuTier {
+        pattern: r"hd graphics|hd [456]\d{2}",
+        name: "Intel HD Graphics (Integrated)",
+        workgroup_divisor: 30,
+        min_workgroups: 192,
+    },
+];
+
+const QUALCOMM_TIERS: &[GpuTier] = &[
+    // Snapdragon X (Adreno X1)
+    GpuTier {
+        pattern: r"x elite|x plus|adreno x|x1-[89]",
+        name: "Qualcomm Adreno X1 (Snapdragon X)",
+        workgroup_divisor: 14,
+        min_workgroups: 1536,
+    },
+    // Adreno 700 series (730, 740, 750, etc.)
+    GpuTier {
+        pattern: r"adreno\s*7[0-9]{2}\b|\b7[345]0\b",
+        name: "Qualcomm Adreno 700 Series",
+        workgroup_divisor: 16,
+        min_workgroups: 1024,
+    },
+    // Adreno 600 series (610, 612, 615, 618, 619, 620, 630, 640, 650, 660, etc.)
+    GpuTier {
+        pattern: r"adreno\s*6[0-9]{2}\b|\b6[1-6][0-9]\b",
+        name: "Qualcomm Adreno 600 Series",
+        workgroup_divisor: 20,
+        min_workgroups: 512,
+    },
+    // Adreno 500 series (505, 506, 508, 509, 510, 512, 530, 540)
+    GpuTier {
+        pattern: r"adreno\s*5[0-4][0-9]\b|\b5[0-4][0-9]\b",
+        name: "Qualcomm Adreno 500 Series",
+        workgroup_divisor: 24,
+        min_workgroups: 384,
+    },
+];
+
+const APPLE_TIERS: &[GpuTier] = &[
+    // M5 series
+    GpuTier {
+        pattern: r"m5 ultra",
+        name: "Apple M5 Ultra",
+        workgroup_divisor: 4,
+        min_workgroups: 1600,
+    },
+    GpuTier {
+        pattern: r"m5 max",
+        name: "Apple M5 Max",
+        workgroup_divisor: 4,
+        min_workgroups: 800,
+    },
+    GpuTier {
+        pattern: r"m5 pro",
+        name: "Apple M5 Pro",
+        workgroup_divisor: 4,
+        min_workgroups: 400,
+    },
+    GpuTier {
+        pattern: r"\bm5\b",
+        name: "Apple M5",
+        workgroup_divisor: 4,
+        min_workgroups: 200,
+    },
+    // M4 series
+    GpuTier {
+        pattern: r"m4 ultra",
+        name: "Apple M4 Ultra",
+        workgroup_divisor: 4,
+        min_workgroups: 1600,
+    },
+    GpuTier {
+        pattern: r"m4 max",
+        name: "Apple M4 Max",
+        workgroup_divisor: 4,
+        min_workgroups: 800,
+    },
+    GpuTier {
+        pattern: r"m4 pro",
+        name: "Apple M4 Pro",
+        workgroup_divisor: 4,
+        min_workgroups: 400,
+    },
+    GpuTier {
+        pattern: r"\bm4\b",
+        name: "Apple M4",
+        workgroup_divisor: 4,
+        min_workgroups: 200,
+    },
+    // M3 series
+    GpuTier {
+        pattern: r"m3 ultra",
+        name: "Apple M3 Ultra",
+        workgroup_divisor: 4,
+        min_workgroups: 1520,
+    },
+    GpuTier {
+        pattern: r"m3 max",
+        name: "Apple M3 Max",
+        workgroup_divisor: 4,
+        min_workgroups: 800,
+    },
+    GpuTier {
+        pattern: r"m3 pro",
+        name: "Apple M3 Pro",
+        workgroup_divisor: 4,
+        min_workgroups: 360,
+    },
+    GpuTier {
+        pattern: r"\bm3\b",
+        name: "Apple M3",
+        workgroup_divisor: 4,
+        min_workgroups: 200,
+    },
+    // M2 series
+    GpuTier {
+        pattern: r"m2 ultra",
+        name: "Apple M2 Ultra",
+        workgroup_divisor: 4,
+        min_workgroups: 1520,
+    },
+    GpuTier {
+        pattern: r"m2 max",
+        name: "Apple M2 Max",
+        workgroup_divisor: 4,
+        min_workgroups: 760,
+    },
+    GpuTier {
+        pattern: r"m2 pro",
+        name: "Apple M2 Pro",
+        workgroup_divisor: 4,
+        min_workgroups: 380,
+    },
+    GpuTier {
+        pattern: r"\bm2\b",
+        name: "Apple M2",
+        workgroup_divisor: 4,
+        min_workgroups: 200,
+    },
+    // M1 series
+    GpuTier {
+        pattern: r"m1 ultra",
+        name: "Apple M1 Ultra",
+        workgroup_divisor: 4,
+        min_workgroups: 1280,
+    },
+    GpuTier {
+        pattern: r"m1 max",
+        name: "Apple M1 Max",
+        workgroup_divisor: 4,
+        min_workgroups: 640,
+    },
+    GpuTier {
+        pattern: r"m1 pro",
+        name: "Apple M1 Pro",
+        workgroup_divisor: 4,
+        min_workgroups: 320,
+    },
+    GpuTier {
+        pattern: r"\bm1\b",
+        name: "Apple M1",
+        workgroup_divisor: 4,
+        min_workgroups: 160,
+    },
+];
+
+/// Compiled GPU tier tables (lazily initialized)
+struct GpuTierTables {
+    nvidia: Vec<CompiledGpuTier>,
+    amd: Vec<CompiledGpuTier>,
+    intel: Vec<CompiledGpuTier>,
+    qualcomm: Vec<CompiledGpuTier>,
+    apple: Vec<CompiledGpuTier>,
+}
+
+static GPU_TIERS: LazyLock<GpuTierTables> = LazyLock::new(|| GpuTierTables {
+    nvidia: NVIDIA_TIERS
+        .iter()
+        .map(CompiledGpuTier::from_tier)
+        .collect(),
+    amd: AMD_TIERS.iter().map(CompiledGpuTier::from_tier).collect(),
+    intel: INTEL_TIERS.iter().map(CompiledGpuTier::from_tier).collect(),
+    qualcomm: QUALCOMM_TIERS
+        .iter()
+        .map(CompiledGpuTier::from_tier)
+        .collect(),
+    apple: APPLE_TIERS.iter().map(CompiledGpuTier::from_tier).collect(),
+});
+
+/// Result of GPU tier matching
+pub struct GpuTierMatch {
+    pub name: &'static str,
+    pub workgroup_divisor: u32,
+    pub min_workgroups: u32,
+    pub is_fallback: bool,
+}
+
+/// Find matching tier from a list of compiled tiers
+fn find_matching_tier(name: &str, tiers: &[CompiledGpuTier]) -> Option<GpuTierMatch> {
+    tiers
+        .iter()
+        .find(|tier| tier.regex.is_match(name))
+        .map(|tier| GpuTierMatch {
+            name: tier.name,
+            workgroup_divisor: tier.workgroup_divisor,
+            min_workgroups: tier.min_workgroups,
+            is_fallback: false,
+        })
+}
+
+/// Detect GPU tier based on adapter info
+///
+/// Returns (tier_name, workgroup_divisor, min_workgroups, is_fallback)
+pub fn detect_gpu_tier(vendor_name: &str, vendor_id: u32, is_metal_backend: bool) -> GpuTierMatch {
+    let name_lower = vendor_name.to_lowercase();
+
+    // Determine vendor and find matching tier
+    if name_lower.contains("nvidia") || vendor_id == 0x10DE {
+        find_matching_tier(&name_lower, &GPU_TIERS.nvidia).unwrap_or(GpuTierMatch {
+            name: "NVIDIA Unknown",
+            workgroup_divisor: 20,
+            min_workgroups: 512,
+            is_fallback: true,
+        })
+    } else if name_lower.contains("amd") || name_lower.contains("radeon") || vendor_id == 0x1002 {
+        find_matching_tier(&name_lower, &GPU_TIERS.amd).unwrap_or(GpuTierMatch {
+            name: "AMD Unknown",
+            workgroup_divisor: 24,
+            min_workgroups: 512,
+            is_fallback: true,
+        })
+    } else if name_lower.contains("intel") || vendor_id == 0x8086 {
+        find_matching_tier(&name_lower, &GPU_TIERS.intel).unwrap_or(GpuTierMatch {
+            name: "Intel Unknown",
+            workgroup_divisor: 24,
+            min_workgroups: 256,
+            is_fallback: true,
+        })
+    } else if name_lower.contains("qualcomm")
+        || name_lower.contains("adreno")
+        || vendor_id == 0x5143
+    {
+        find_matching_tier(&name_lower, &GPU_TIERS.qualcomm).unwrap_or(GpuTierMatch {
+            name: "Qualcomm Adreno (Unknown)",
+            workgroup_divisor: 24,
+            min_workgroups: 384,
+            is_fallback: true,
+        })
+    } else if is_metal_backend {
+        // Apple Silicon - detected by Metal backend
+        find_matching_tier(&name_lower, &GPU_TIERS.apple).unwrap_or(GpuTierMatch {
+            name: "Apple Silicon Unknown",
+            workgroup_divisor: 4,
+            min_workgroups: 160,
+            is_fallback: true,
+        })
+    } else {
+        GpuTierMatch {
+            name: "Unknown GPU",
+            workgroup_divisor: 16,
+            min_workgroups: 512,
+            is_fallback: true,
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_nvidia_detection() {
+        // RTX 40 series
+        let tier = detect_gpu_tier("NVIDIA GeForce RTX 4090", 0x10DE, false);
+        assert_eq!(tier.name, "NVIDIA RTX 40 Flagship (Ada)");
+        assert!(!tier.is_fallback);
+
+        let tier = detect_gpu_tier("NVIDIA GeForce RTX 4070 Ti", 0x10DE, false);
+        assert_eq!(tier.name, "NVIDIA RTX 40 (Ada)");
+
+        // RTX 30 series
+        let tier = detect_gpu_tier("NVIDIA GeForce RTX 3080", 0x10DE, false);
+        assert_eq!(tier.name, "NVIDIA RTX 30/20 (Ampere/Turing)");
+
+        // GTX 10 series
+        let tier = detect_gpu_tier("NVIDIA GeForce GTX 1080 Ti", 0x10DE, false);
+        assert_eq!(tier.name, "NVIDIA GTX 16/10 (Turing/Pascal)");
+    }
+
+    #[test]
+    fn test_amd_rdna_vs_polaris() {
+        // RDNA 1 - should NOT match Polaris
+        let tier = detect_gpu_tier("AMD Radeon RX 5500 XT", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 5000 (RDNA 1)");
+        assert!(!tier.is_fallback);
+
+        let tier = detect_gpu_tier("AMD Radeon RX 5600 XT", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 5000 (RDNA 1)");
+
+        let tier = detect_gpu_tier("AMD Radeon RX 5700 XT", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 5700 (RDNA 1)");
+
+        // Polaris - should match Polaris
+        let tier = detect_gpu_tier("AMD Radeon RX 580", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 500/400 (Polaris)");
+
+        let tier = detect_gpu_tier("AMD Radeon RX 560X", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 500/400 (Polaris)");
+
+        let tier = detect_gpu_tier("AMD Radeon RX 550", 0x1002, false);
+        assert_eq!(tier.name, "AMD RX 500/400 (Polaris)");
+    }
+
+    #[test]
+    fn test_amd_vega_apu() {
+        let tier = detect_gpu_tier("AMD Radeon(TM) Vega 8 Graphics", 0x1002, false);
+        assert_eq!(tier.name, "AMD Vega (APU)");
+        assert!(!tier.is_fallback);
+    }
+
+    #[test]
+    fn test_intel_arc_mobile_vs_desktop() {
+        // Mobile should match mobile tier
+        let tier = detect_gpu_tier("Intel Arc A770M", 0x8086, false);
+        assert_eq!(tier.name, "Intel Arc A7 Mobile (Alchemist)");
+        assert!(!tier.is_fallback);
+
+        // Desktop should match desktop tier
+        let tier = detect_gpu_tier("Intel Arc A770", 0x8086, false);
+        assert_eq!(tier.name, "Intel Arc A7 (Alchemist)");
+        assert!(!tier.is_fallback);
+    }
+
+    #[test]
+    fn test_apple_silicon() {
+        let tier = detect_gpu_tier("Apple M1 Pro", 0, true);
+        assert_eq!(tier.name, "Apple M1 Pro");
+        assert!(!tier.is_fallback);
+
+        let tier = detect_gpu_tier("Apple M3 Max", 0, true);
+        assert_eq!(tier.name, "Apple M3 Max");
+        assert!(!tier.is_fallback);
+    }
+
+    #[test]
+    fn test_qualcomm_adreno_series() {
+        // Adreno 627 should match 600 series, NOT 700 series
+        let tier = detect_gpu_tier("Qualcomm Adreno 627", 0x5143, false);
+        assert_eq!(tier.name, "Qualcomm Adreno 600 Series");
+        assert!(!tier.is_fallback);
+
+        // Adreno 730 should match 700 series
+        let tier = detect_gpu_tier("Qualcomm Adreno 730", 0x5143, false);
+        assert_eq!(tier.name, "Qualcomm Adreno 700 Series");
+        assert!(!tier.is_fallback);
+
+        // Adreno 540 should match 500 series
+        let tier = detect_gpu_tier("Qualcomm Adreno 540", 0x5143, false);
+        assert_eq!(tier.name, "Qualcomm Adreno 500 Series");
+        assert!(!tier.is_fallback);
+    }
+}
@@ -9,7 +9,7 @@ const GOLDILOCKS_PRIME_HIGH: u32 = 4294967295u;  // High 32 bits (2^32 - 1)

 // Poseidon2 constants
 const WIDTH: u32 = 12u;
-const RATE: u32 = 4u;
+const RATE: u32 = 8u;
 const EXTERNAL_ROUNDS: u32 = 4u;
 const INTERNAL_ROUNDS: u32 = 22u;

@@ -191,10 +191,12 @@ const MDS_MATRIX_DIAG_12: array<array<u32, 2>, 12> = array<array<u32, 2>, 12>(

 // Storage buffers
@group(0) @binding(0) var<storage, read_write> results: array<atomic<u32>>;
-@group(0) @binding(1) var<storage, read> header: array<u32, 8>;     // 32 bytes
+// Sponge state after absorbing header + high nonce half (12 felts as LE u32 pairs),
+// precomputed on the host per batch. See pow_core::mining_midstate.
+@group(0) @binding(1) var<storage, read> midstate: array<u32, 24>;
@group(0) @binding(2) var<storage, read> start_nonce: array<u32, 16>; // 64 bytes
@group(0) @binding(3) var<storage, read> difficulty_target: array<u32, 16>;    // 64 bytes (U512 target)
-@group(0) @binding(4) var<storage, read> dispatch_config: array<u32, 4>; // [total_threads (logical), nonces_per_thread, total_nonces (logical nonces in this dispatch), threads_per_workgroup]
+@group(0) @binding(4) var<storage, read> dispatch_config: array<u32, 3>; // [total_threads, nonces_per_thread, total_nonces]

 // Goldilocks field element represented as [limb0, limb1]
 // where the value is limb0 + limb1*2^32
@@ -676,17 +678,17 @@ fn poseidon2_hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
        state[i] = gf_zero();
    }

-    // Convert input to field elements (25 total)
+    // Convert input to field elements (25 total: 24 from input + 1 terminator)
    let input_felts = bytes_to_field_elements(input);

    // Sponge construction matching CPU reference exactly:
-    // CPU processes field elements using push_to_buf() which absorbs in chunks of RATE=4
+    // CPU processes field elements using push_to_buf() which absorbs in chunks of RATE=8

-    // Process first 24 elements (6 complete chunks of 4)
-    for (var chunk = 0u; chunk < 6u; chunk++) {
-        // Absorb 4 elements for this chunk
-        for (var i = 0u; i < 4u; i++) {
-            let felt_idx = chunk * 4u + i;
+    // Process first 24 elements (3 complete chunks of 8)
+    for (var chunk = 0u; chunk < 3u; chunk++) {
+        // Absorb 8 elements for this chunk
+        for (var i = 0u; i < 8u; i++) {
+            let felt_idx = chunk * 8u + i;
            state[i] = gf_add(state[i], input_felts[felt_idx]);
        }
        // Permute after each complete chunk
@@ -694,17 +696,16 @@ fn poseidon2_hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
    }

    // Process remaining 1 element (partial chunk)
-    // Now we have element 24 (the padding marker = 1) remaining
-    // This simulates CPU's finalize_twice adding ONE to buffer position 0
-    state[0] = gf_add(state[0], input_felts[24u]); // Add the padding marker (should be 1)
+    // Now we have element 24 (the terminator = 1) remaining
+    state[0] = gf_add(state[0], input_felts[24u]); // Add the terminator (value 1)

    // Add sponge padding marker (ONE) to the next position
    state[1] = gf_add(state[1], gf_one());

-    // Final permutation (CPU calls permute after completing the block)
+    // Final permutation
    poseidon2_permute(&state);

-    // First squeeze - get first 4 field elements
+    // First squeeze - get first 4 field elements (POSEIDON2_OUTPUT=4 -> 32 bytes)
    let first_output = field_elements_to_bytes(array<GoldilocksField, 4>(
        state[0], state[1], state[2], state[3]
    ));
@@ -783,23 +784,13 @@ fn poseidon2_hash_squeeze_twice_64(input: array<u32, 16>) -> array<u32, 16> {
    return result;
 }

-// Double Poseidon2 hash (like Bitcoin's double SHA256)
-fn double_hash(input: array<u32, 24>) -> array<u32, 16> {
-    let first_hash = poseidon2_hash_squeeze_twice(input);
-    return poseidon2_hash_squeeze_twice_64(first_hash);
+// Single Poseidon2 hash with double squeeze (512-bit output)
+fn hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
+    return poseidon2_hash_squeeze_twice(input);
 }

-// Check if hash < target (U512 comparison)
-fn is_below_target(hash: array<u32, 16>, difficulty_tgt: array<u32, 16>) -> bool {
-    // Compare from most significant to least significant
-    for (var i = 0u; i < 16u; i++) {
-        if (hash[15u - i] < difficulty_tgt[15u - i]) {
-            return true;
-        } else if (hash[15u - i] > difficulty_tgt[15u - i]) {
-            return false;
-        }
-    }
-    return false; // Equal, not below
+fn bswap32(v: u32) -> u32 {
+    return ((v & 0xFFu) << 24u) | ((v & 0xFF00u) << 8u) | ((v >> 8u) & 0xFF00u) | (v >> 24u);
 }

 // Main mining kernel
@@ -836,61 +827,90 @@ fn mining_main(@builtin(global_invocation_id) global_id: vec3<u32>) {
            return;
        }

-        // current_nonce = start_nonce + logical_index
+        // current_nonce = start_nonce + logical_index. The host guarantees a batch
+        // never carries into the high nonce half (limbs 8..15).
        var current_nonce: array<u32, 16>;
        var carry: u32 = 0u;

-        // Add logical_index into the low limb and propagate carry through the U512 nonce
        let val0 = start_nonce[0];
        let sum0 = val0 + logical_index;
        current_nonce[0] = sum0;
        carry = select(0u, 1u, sum0 < val0);

-        // Propagate carry through remaining limbs
-        for (var i = 1u; i < 16u; i++) {
+        for (var i = 1u; i < 8u; i++) {
            let val = start_nonce[i];
            let sum = val + carry;
            current_nonce[i] = sum;
            carry = select(0u, 1u, sum < val);
        }
-
-        // Construct input (96 bytes = 24 u32s)
-        // Header (32 bytes = 8 u32s) followed by Nonce (64 bytes = 16 u32s)
-        var input: array<u32, 24>;
-        for (var i = 0u; i < 8u; i++) {
-            input[i] = header[i];
+        for (var i = 8u; i < 16u; i++) {
+            current_nonce[i] = start_nonce[i];
        }
-        // Nonce needs to be Big Endian in the byte stream for hashing.
-        // current_nonce is Little Endian words.
-        for (var i = 0u; i < 16u; i++) {
-            let val = current_nonce[15u - i];
-            // Reverse bytes
+
+        // Resume the sponge from the precomputed midstate: absorb the low nonce
+        // half (as Big Endian words), pad, squeeze twice (3 permutations instead of 5).
+        var state: array<GoldilocksField, 12>;
+        for (var i = 0u; i < 12u; i++) {
+            state[i] = gf_from_limbs(midstate[2u * i], midstate[2u * i + 1u]);
+        }
+        for (var i = 0u; i < 8u; i++) {
+            let val = current_nonce[7u - i];
            let rev = ((val & 0xFFu) << 24u) |
                      ((val & 0xFF00u) << 8u) |
                      ((val & 0xFF0000u) >> 8u) |
                      ((val & 0xFF000000u) >> 24u);
-            input[8u + i] = rev;
+            state[i] = gf_add(state[i], gf_from_u32(rev));
        }
+        poseidon2_permute(&state);
+        state[0] = gf_add(state[0], gf_one());
+        state[1] = gf_add(state[1], gf_one());
+        poseidon2_permute(&state);

-        // Hash (Big Endian)
-        let hash_be = double_hash(input);
-
-        // Convert to Little Endian for difficulty check and storage
+        // First squeeze yields the most significant 256 bits of the hash, which
+        // decide hash-vs-target on their own unless they exactly equal the
+        // target's high half. Only candidates pay for the second squeeze.
+        let first_output = field_elements_to_bytes(array<GoldilocksField, 4>(
+            state[0], state[1], state[2], state[3]
+        ));
        var hash_le: array<u32, 16>;
-        for (var i = 0u; i < 16u; i++) {
-            let val = hash_be[15u - i];
-            // Reverse bytes in u32
-            hash_le[i] = ((val & 0xFFu) << 24u) |
-                         ((val & 0xFF00u) << 8u) |
-                         ((val & 0xFF0000u) >> 8u) |
-                         ((val & 0xFF000000u) >> 24u);
+        for (var i = 0u; i < 8u; i++) {
+            hash_le[15u - i] = bswap32(first_output[i]);
+        }
+        var cmp = 0u;
+        for (var i = 0u; i < 8u; i++) {
+            let h = hash_le[15u - i];
+            let t = difficulty_target[15u - i];
+            if (h != t) {
+                cmp = select(2u, 1u, h > t);
+                break;
+            }
+        }
+        if (cmp == 1u) {
+            continue;
        }

-        // Check target
-        if (is_below_target(hash_le, difficulty_target)) {
+        poseidon2_permute(&state);
+        let second_output = field_elements_to_bytes(array<GoldilocksField, 4>(
+            state[0], state[1], state[2], state[3]
+        ));
+        for (var i = 0u; i < 8u; i++) {
+            hash_le[i] = bswap32(second_output[7u - i]);
+        }
+        var below = cmp == 2u;
+        if (!below) {
+            for (var i = 0u; i < 8u; i++) {
+                let h = hash_le[7u - i];
+                let t = difficulty_target[7u - i];
+                if (h != t) {
+                    below = h < t;
+                    break;
+                }
+            }
+        }
+
+        if (below) {
            // Try to claim the solution
            if (atomicExchange(&results[0], 1u) == 0u) {
-                // We won! Write nonce and hash
                // results layout: [0]=found, [1..16]=nonce, [17..32]=hash
                for (var i = 0u; i < 16u; i++) {
                    atomicStore(&results[1u + i], current_nonce[i]);
@@ -0,0 +1,573 @@
+// Native-u64 Poseidon2 mining kernel. Default path for non-Apple GPUs.
+// Requires wgpu Features::SHADER_INT64. Bit-exact with mining.wgsl / pow_core.
+
+@group(0) @binding(0) var<storage, read_write> results: array<atomic<u32>>;
+// Sponge state after absorbing header + high nonce half (12 felts as LE u32 pairs),
+// precomputed on the host per batch. See pow_core::mining_midstate.
+@group(0) @binding(1) var<storage, read> midstate: array<u32, 24>;
+@group(0) @binding(2) var<storage, read> start_nonce: array<u32, 16>;
+@group(0) @binding(3) var<storage, read> difficulty_target: array<u32, 16>;
+@group(0) @binding(4) var<storage, read> dispatch_config: array<u32, 3>;
+
+const P64: u64 = 0xFFFFFFFF00000001lu;
+// EPS64 = 2^32 - 1 = 2^64 mod P
+const EPS64: u64 = 0xFFFFFFFFlu;
+
+const RC_INTERNAL: array<u64, 22> = array<u64, 22>(
+    0x97f7798a784ad863lu, 0xd1d2bf082f60d4f0lu, 0x69a377a79f9ad206lu, 0xa9d06906a3858e24lu, 0x295275001eede5b5lu, 0x5874e441117bd746lu, 0x8a084bbba8ed86cclu, 0x3defd7645cde6425lu, 0x3998cfe6871cc137lu, 0x3e52ef8bca48314alu, 0x964a209f85dc9ecclu, 0x3fcc9ee82cc4577elu, 0x8e79b4a5d0096d6dlu, 0x8492362ad2392556lu, 0xee72f470262574d6lu, 0x1e0e18496da2444alu, 0x0f3a74bf215eaac6lu, 0x1b061b76a1c0ded3lu, 0x192c42d86803d7a6lu, 0xf6d49ff997ae0260lu, 0x3ec372e7a0fa3786lu, 0x5538cdf4f23445d3lu
+);
+const RC_INITIAL: array<array<u64, 12>, 4> = array<array<u64, 12>, 4>(
+    array<u64, 12>(0xc002e770975b1607lu, 0xbca51a8dfe14593alu, 0x72938dfbe774f7f9lu, 0xe4f2fe29e03234aclu, 0xd5e0ba2f541b6449lu, 0xec33b868f3cc46c1lu, 0x486dcb55419d475alu, 0x6c1cb2a358cc24f1lu, 0xe3f30d509a1436bblu, 0xd9a64f068dca7c29lu, 0xe59b3f57aabba1aelu, 0x2a3dd4505b478fdclu),
+    array<u64, 12>(0xada1f8dc7676ed25lu, 0x2711aa8b5509d516lu, 0x4ae6acd0c9c92897lu, 0x56eb3d6b5256d67alu, 0x1f7a9d55923bf51elu, 0x3600427d397a7f68lu, 0xe5076df75b72c3d0lu, 0xfcd59aa12c6090adlu, 0xcd895e8c68b57a9elu, 0x41df7ef9d730ae3elu, 0xee3e2b889abe977dlu, 0xd29bb7edbeb9c405lu),
+    array<u64, 12>(0x7d5c08eef608e382lu, 0x89ae889caaf0802clu, 0xb35a8e976d2af617lu, 0xdb14234eafaf5173lu, 0x78f04462d48b1c98lu, 0x265293b0e47ce88alu, 0x999a649b69b9d32flu, 0x64b0a186698e01d3lu, 0xee0b22d0dfae8bb8lu, 0x4fd53e50ca04a7eelu, 0x5762bfe181f25047lu, 0xf51593e2beb5e3bdlu),
+    array<u64, 12>(0x1e5e2b5760e32477lu, 0x622462a1f9aaaeedlu, 0xaa284b3ecdb222aelu, 0x63c8e72f542bf3fclu, 0x3ba588cacb43b5e0lu, 0x23eda6f3c99150ddlu, 0xaad3bea4baac9a5alu, 0xe9da8d699b94184alu, 0xcdb13f4cd93e024clu, 0x902cbd0956f655e3lu, 0x5b4e40ffc759532flu, 0xde795c20a2357af7lu)
+);
+const RC_TERMINAL: array<array<u64, 12>, 4> = array<array<u64, 12>, 4>(
+    array<u64, 12>(0x7b72c539e0ea4c6elu, 0x144573dae2ce9976lu, 0x802028b68f35fc88lu, 0x6d36c5022c4fe7c2lu, 0xa205d0ffa9b9def3lu, 0xf6e7e38b1ea6ba2flu, 0x34f7909ae5258d64lu, 0xb0464d9d77b97fcalu, 0x64ddb9d5de7e00a6lu, 0x0ed0d75c27975d97lu, 0x1cbb36f11127338blu, 0x6673e505cfd0b6balu),
+    array<u64, 12>(0x605f902830872e01lu, 0x3fd5eb927e95fe4flu, 0xe81025b5a24c69cdlu, 0xf7d0ce75de23f74elu, 0xf39942b6a8585089lu, 0x6d808a08f7b71df6lu, 0xf8806b6588f49a8blu, 0x57df2d8c2a32107alu, 0x16e7c2074d654a2dlu, 0x213de241fcf33835lu, 0xb0f2b8905a0976f6lu, 0xd8e3cf2bbd355417lu),
+    array<u64, 12>(0xe498691679d9330flu, 0x763b45d2a3821b28lu, 0x0908bf65eb0a1f0dlu, 0x7691eb2d194b24f4lu, 0x0e43551233ae13b2lu, 0x93c393dbfc2fe76flu, 0x98f607485d48cdealu, 0xe3d95f30309819c0lu, 0x1ef581a93eaf6acflu, 0x0b24c1b7a030fca4lu, 0x624370be5670b327lu, 0x5f1e28615a11e486lu),
+    array<u64, 12>(0xfe04051f909e042blu, 0x7257e5b147fd3803lu, 0xe6ae134bb82f2e78lu, 0x5711fd5cf4784511lu, 0xf83a42660c08c0bclu, 0x2cd8c96d9a3ce855lu, 0x7d2ffb1bb0e17271lu, 0x85ae1528caea3811lu, 0x52a345d5c7adb0b8lu, 0x504c4c51f3faee94lu, 0xbce34a649cfccaf9lu, 0xe0a3389266fb6dc9lu)
+);
+const MDS_DIAG: array<u64, 12> = array<u64, 12>(
+    0xc3b6c08e23ba9300lu, 0xd84b5de94a324fb6lu, 0x0d0c371c5b35b84flu, 0x7964f570e7188037lu, 0x5daf18bbd996604blu, 0x6743bc47b9595257lu, 0x5528b9362c59bb70lu, 0xac45e25b7127b68blu, 0xa2077d7dfbb606b5lu, 0xf3faac6faee378aelu, 0x0c6388b51545e883lu, 0xd27dbb6944917b60lu
+);
+
+// a + b mod P in lazy form. Wrapping carries fold back via 2^64 ≡ EPS64 (mod P).
+fn gf64_add(a: u64, b: u64) -> u64 {
+    let s0 = a + b;
+    let c1 = s0 < a;
+    let s1 = s0 + select(0lu, EPS64, c1);
+    let c2 = c1 && (s1 < s0);
+    return s1 + select(0lu, EPS64, c2);
+}
+
+// Reduce a 128-bit value (lo + hi*2^64) mod P using
+// 2^64 ≡ EPS64 and 2^96 ≡ -1 (mod P).
+fn gf64_reduce(lo: u64, hi: u64) -> u64 {
+    let hi_hi = hi >> 32u;
+    let hi_lo = hi & EPS64;
+    var t0 = lo - hi_hi;
+    t0 = t0 - select(0lu, EPS64, lo < hi_hi);
+    let t1 = hi_lo * EPS64;
+    let t2 = t0 + t1;
+    return t2 + select(0lu, EPS64, t2 < t0);
+}
+
+fn gf64_mul(a: u64, b: u64) -> u64 {
+    let a_lo = a & EPS64;
+    let a_hi = a >> 32u;
+    let b_lo = b & EPS64;
+    let b_hi = b >> 32u;
+    let ll = a_lo * b_lo;
+    let lh = a_lo * b_hi;
+    let hl = a_hi * b_lo;
+    let hh = a_hi * b_hi;
+    let mid = lh + hl;
+    let mid_c = select(0lu, 1lu, mid < lh);
+    let lo = ll + (mid << 32u);
+    let lo_c = select(0lu, 1lu, lo < ll);
+    let hi = hh + (mid >> 32u) + (mid_c << 32u) + lo_c;
+    return gf64_reduce(lo, hi);
+}
+
+fn gf64_sqr(a: u64) -> u64 {
+    let a_lo = a & EPS64;
+    let a_hi = a >> 32u;
+    let ll = a_lo * a_lo;
+    let lh = a_lo * a_hi;
+    let hh = a_hi * a_hi;
+    let mid = lh << 1u;
+    let mid_c = lh >> 63u;
+    let lo = ll + (mid << 32u);
+    let lo_c = select(0lu, 1lu, lo < ll);
+    let hi = hh + (mid >> 32u) + (mid_c << 32u) + lo_c;
+    return gf64_reduce(lo, hi);
+}
+
+fn gf64_sbox(x: u64) -> u64 {
+    let x2 = gf64_sqr(x);
+    let x4 = gf64_sqr(x2);
+    let x6 = gf64_mul(x4, x2);
+    return gf64_mul(x6, x);
+}
+
+fn gf64_canon(a: u64) -> u64 {
+    return a - select(0lu, P64, a >= P64);
+}
+
+// External linear layer: 4x4 MDS on each chunk, then circulant sums.
+fn ext_layer64(state: ptr<function, array<u64, 12>>) {
+    for (var chunk = 0u; chunk < 3u; chunk++) {
+        let o = chunk * 4u;
+        let x0 = (*state)[o];
+        let x1 = (*state)[o + 1u];
+        let x2 = (*state)[o + 2u];
+        let x3 = (*state)[o + 3u];
+        let t01 = gf64_add(x0, x1);
+        let t23 = gf64_add(x2, x3);
+        let t0123 = gf64_add(t01, t23);
+        let t01123 = gf64_add(t0123, x1);
+        let t01233 = gf64_add(t0123, x3);
+        (*state)[o + 3u] = gf64_add(t01233, gf64_add(x0, x0));
+        (*state)[o + 1u] = gf64_add(t01123, gf64_add(x2, x2));
+        (*state)[o] = gf64_add(t01123, t01);
+        (*state)[o + 2u] = gf64_add(t01233, t23);
+    }
+    var sums: array<u64, 4>;
+    for (var k = 0u; k < 4u; k++) {
+        sums[k] = gf64_add(gf64_add((*state)[k], (*state)[k + 4u]), (*state)[k + 8u]);
+    }
+    for (var i = 0u; i < 12u; i++) {
+        (*state)[i] = gf64_add((*state)[i], sums[i % 4u]);
+    }
+}
+
+// Internal linear layer: diagonal matrix plus full sum.
+fn int_layer64(state: ptr<function, array<u64, 12>>) {
+    var sum = (*state)[0];
+    for (var i = 1u; i < 12u; i++) {
+        sum = gf64_add(sum, (*state)[i]);
+    }
+    for (var i = 0u; i < 12u; i++) {
+        (*state)[i] = gf64_add(gf64_mul((*state)[i], MDS_DIAG[i]), sum);
+    }
+}
+
+fn permute64(state: ptr<function, array<u64, 12>>) {
+    ext_layer64(state);
+    for (var r = 0u; r < 4u; r++) {
+        for (var i = 0u; i < 12u; i++) {
+            (*state)[i] = gf64_add((*state)[i], RC_INITIAL[r][i]);
+        }
+        for (var i = 0u; i < 12u; i++) {
+            (*state)[i] = gf64_sbox((*state)[i]);
+        }
+        ext_layer64(state);
+    }
+    for (var r = 0u; r < 22u; r++) {
+        (*state)[0] = gf64_sbox(gf64_add((*state)[0], RC_INTERNAL[r]));
+        int_layer64(state);
+    }
+    for (var r = 0u; r < 4u; r++) {
+        for (var i = 0u; i < 12u; i++) {
+            (*state)[i] = gf64_add((*state)[i], RC_TERMINAL[r][i]);
+        }
+        for (var i = 0u; i < 12u; i++) {
+            (*state)[i] = gf64_sbox((*state)[i]);
+        }
+        ext_layer64(state);
+    }
+}
+
+fn bswap32(v: u32) -> u32 {
+    return ((v & 0xFFu) << 24u) | ((v & 0xFF00u) << 8u) | ((v >> 8u) & 0xFF00u) | (v >> 24u);
+}
+
+@compute @workgroup_size(256)
+fn mining_main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+    if (atomicLoad(&results[0]) != 0u) {
+        return;
+    }
+    let thread_id = global_id.x;
+    let total_threads = dispatch_config[0];
+    let nonces_per_thread = dispatch_config[1];
+    let total_nonces = dispatch_config[2];
+    if (thread_id >= total_threads) {
+        return;
+    }
+    let base_index = thread_id * nonces_per_thread;
+
+    // Hoist uniform storage reads out of the nonce loop
+    var mid: array<u64, 12>;
+    for (var i = 0u; i < 12u; i++) {
+        mid[i] = (u64(midstate[2u * i + 1u]) << 32u) | u64(midstate[2u * i]);
+    }
+    var tgt: array<u32, 16>;
+    for (var i = 0u; i < 16u; i++) {
+        tgt[i] = difficulty_target[i];
+    }
+    var nonce_base: array<u32, 16>;
+    for (var i = 0u; i < 16u; i++) {
+        nonce_base[i] = start_nonce[i];
+    }
+
+    for (var j = 0u; j < nonces_per_thread; j = j + 1u) {
+        let logical_index = base_index + j;
+        if (logical_index >= total_nonces) {
+            break;
+        }
+        if (j > 0u && atomicLoad(&results[0]) != 0u) {
+            return;
+        }
+
+        // The host guarantees a batch never carries into the high nonce half
+        // (limbs 8..15), so only the low 256 bits are incremented here.
+        var current_nonce: array<u32, 16>;
+        let val0 = nonce_base[0];
+        let sum0 = val0 + logical_index;
+        current_nonce[0] = sum0;
+        var carry = select(0u, 1u, sum0 < val0);
+        for (var i = 1u; i < 8u; i++) {
+            let val = nonce_base[i];
+            let sum = val + carry;
+            current_nonce[i] = sum;
+            carry = select(0u, 1u, sum < val);
+        }
+        for (var i = 8u; i < 16u; i++) {
+            current_nonce[i] = nonce_base[i];
+        }
+
+        // Resume the sponge from the precomputed midstate: absorb the low
+        // nonce half, pad, squeeze twice (3 permutations instead of 5).
+        var st: array<u64, 12>;
+        for (var i = 0u; i < 12u; i++) {
+            st[i] = mid[i];
+        }
+        for (var i = 0u; i < 8u; i++) {
+            st[i] = gf64_add(st[i], u64(bswap32(current_nonce[7u - i])));
+        }
+        permute64(&st);
+        st[0] = gf64_add(st[0], 1lu);
+        st[1] = gf64_add(st[1], 1lu);
+        permute64(&st);
+
+        // First squeeze yields the most significant 256 bits of the hash, which
+        // decide hash-vs-target on their own unless they exactly equal the
+        // target's high half. Only candidates pay for the second squeeze, and
+        // byte-swapped hash words are produced on demand during the compare.
+        var first: array<u32, 8>;
+        for (var i = 0u; i < 4u; i++) {
+            let c = gf64_canon(st[i]);
+            first[2u * i] = u32(c & EPS64);
+            first[2u * i + 1u] = u32(c >> 32u);
+        }
+        var cmp = 0u;
+        for (var i = 0u; i < 8u; i++) {
+            let h = bswap32(first[i]);
+            let t = tgt[15u - i];
+            if (h != t) {
+                cmp = select(2u, 1u, h > t);
+                break;
+            }
+        }
+        if (cmp == 1u) {
+            continue;
+        }
+
+        var hash_le: array<u32, 16>;
+        for (var i = 0u; i < 8u; i++) {
+            hash_le[15u - i] = bswap32(first[i]);
+        }
+        permute64(&st);
+        for (var i = 0u; i < 4u; i++) {
+            let c = gf64_canon(st[i]);
+            hash_le[7u - 2u * i] = bswap32(u32(c & EPS64));
+            hash_le[6u - 2u * i] = bswap32(u32(c >> 32u));
+        }
+        var below = cmp == 2u;
+        if (!below) {
+            for (var i = 0u; i < 8u; i++) {
+                let h = hash_le[7u - i];
+                let t = tgt[7u - i];
+                if (h != t) {
+                    below = h < t;
+                    break;
+                }
+            }
+        }
+
+        if (below) {
+            if (atomicExchange(&results[0], 1u) == 0u) {
+                for (var i = 0u; i < 16u; i++) {
+                    atomicStore(&results[1u + i], current_nonce[i]);
+                    atomicStore(&results[17u + i], hash_le[i]);
+                }
+            }
+            return;
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Compatibility layer: same API as mining.wgsl, backed by the u64 core above.
+// Only used by the component test harness; the mining kernel never calls it.
+// All outputs are canonical, matching the reference implementation.
+// ---------------------------------------------------------------------------
+
+struct GoldilocksField {
+    limb0: u32,
+    limb1: u32,
+}
+
+const INTERNAL_CONSTANTS: array<array<u32, 2>, 22> = array<array<u32, 2>, 22>(
+    array<u32, 2>(2018170979u, 2549578122u),
+    array<u32, 2>(794875120u, 3520249608u),
+    array<u32, 2>(2677723654u, 1772320679u),
+    array<u32, 2>(2743438884u, 2849007878u),
+    array<u32, 2>(518907317u, 693269760u),
+    array<u32, 2>(293328710u, 1484055617u),
+    array<u32, 2>(2834138828u, 2315799483u),
+    array<u32, 2>(1558078501u, 1039128420u),
+    array<u32, 2>(2266808631u, 966316006u),
+    array<u32, 2>(3393728842u, 1045622667u),
+    array<u32, 2>(2245828300u, 2521440415u),
+    array<u32, 2>(751064958u, 1070374632u),
+    array<u32, 2>(3490278765u, 2390340773u),
+    array<u32, 2>(3526960470u, 2224174634u),
+    array<u32, 2>(639988950u, 4000511088u),
+    array<u32, 2>(1839350858u, 504240201u),
+    array<u32, 2>(559852230u, 255489215u),
+    array<u32, 2>(2713771731u, 453385078u),
+    array<u32, 2>(1745082278u, 422331096u),
+    array<u32, 2>(2544763488u, 4141129721u),
+    array<u32, 2>(2700752774u, 1052996327u),
+    array<u32, 2>(4063512019u, 1429786100u)
+);
+
+const INITIAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2539329031u, 3221415792u),
+        array<u32, 2>(4262746426u, 3164936845u),
+        array<u32, 2>(3883202553u, 1922272763u),
+        array<u32, 2>(3761386668u, 3841130025u),
+        array<u32, 2>(1411081289u, 3588274735u),
+        array<u32, 2>(4090250945u, 3962812520u),
+        array<u32, 2>(1100826458u, 1215155029u),
+        array<u32, 2>(1489773809u, 1813820067u),
+        array<u32, 2>(2585015995u, 3824356688u),
+        array<u32, 2>(2378857513u, 3651555078u),
+        array<u32, 2>(2864423342u, 3852156759u),
+        array<u32, 2>(1531416540u, 708695120u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(1987505445u, 2913073372u),
+        array<u32, 2>(1426707734u, 655469195u),
+        array<u32, 2>(3385403543u, 1256631504u),
+        array<u32, 2>(1381422714u, 1458257259u),
+        array<u32, 2>(2453402910u, 528129365u),
+        array<u32, 2>(964329320u, 905986685u),
+        array<u32, 2>(1534247888u, 3842469367u),
+        array<u32, 2>(744525997u, 4241857185u),
+        array<u32, 2>(1756723870u, 3448331916u),
+        array<u32, 2>(3610291774u, 1105166073u),
+        array<u32, 2>(2596181885u, 3997051784u),
+        array<u32, 2>(3199845381u, 3533420525u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(4127777666u, 2103183598u),
+        array<u32, 2>(2867888172u, 2309916828u),
+        array<u32, 2>(1831532055u, 3009056407u),
+        array<u32, 2>(2947502451u, 3675530062u),
+        array<u32, 2>(3565886616u, 2029012066u),
+        array<u32, 2>(3833391242u, 642945968u),
+        array<u32, 2>(1773785903u, 2577032347u),
+        array<u32, 2>(1770914259u, 1689297286u),
+        array<u32, 2>(3752758200u, 3993707216u),
+        array<u32, 2>(3389302766u, 1339375184u),
+        array<u32, 2>(2180141127u, 1466089441u),
+        array<u32, 2>(3199591357u, 4111832034u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(1625498743u, 509487959u),
+        array<u32, 2>(4188712685u, 1646551713u),
+        array<u32, 2>(3451003566u, 2854767422u),
+        array<u32, 2>(1412166652u, 1674110767u),
+        array<u32, 2>(3410212320u, 1000704202u),
+        array<u32, 2>(3381743837u, 602777331u),
+        array<u32, 2>(3131873882u, 2866003620u),
+        array<u32, 2>(2610174026u, 3923414377u),
+        array<u32, 2>(3644719692u, 3450945356u),
+        array<u32, 2>(1458984419u, 2418851081u),
+        array<u32, 2>(3344519983u, 1531855103u),
+        array<u32, 2>(2721413879u, 3732495392u)
+    )
+);
+
+const TERMINAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(3773451374u, 2071119161u),
+        array<u32, 2>(3805190518u, 340095962u),
+        array<u32, 2>(2402679944u, 2149591222u),
+        array<u32, 2>(743434178u, 1832305922u),
+        array<u32, 2>(2847530739u, 2718290175u),
+        array<u32, 2>(514243119u, 4142392203u),
+        array<u32, 2>(3844443492u, 888639642u),
+        array<u32, 2>(2008645578u, 2957397405u),
+        array<u32, 2>(3732799654u, 1692252629u),
+        array<u32, 2>(664231319u, 248567644u),
+        array<u32, 2>(287781771u, 482031345u),
+        array<u32, 2>(3486561978u, 1718871301u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(814165505u, 1616875560u),
+        array<u32, 2>(2123759183u, 1070984082u),
+        array<u32, 2>(2722916813u, 3893372341u),
+        array<u32, 2>(3726899022u, 4157656693u),
+        array<u32, 2>(2824360073u, 4086907574u),
+        array<u32, 2>(4155973110u, 1837140488u),
+        array<u32, 2>(2297731723u, 4169165669u),
+        array<u32, 2>(707924090u, 1474243980u),
+        array<u32, 2>(1298483757u, 384287239u),
+        array<u32, 2>(4243798069u, 557703745u),
+        array<u32, 2>(1510569718u, 2968696976u),
+        array<u32, 2>(3174388759u, 3638808363u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2044277519u, 3835193622u),
+        array<u32, 2>(2743212840u, 1983595986u),
+        array<u32, 2>(3943309069u, 151568229u),
+        array<u32, 2>(424355060u, 1989274413u),
+        array<u32, 2>(867046322u, 239293714u),
+        array<u32, 2>(4230997871u, 2479068123u),
+        array<u32, 2>(1565052394u, 2566260552u),
+        array<u32, 2>(815274432u, 3822673712u),
+        array<u32, 2>(1051683535u, 519405993u),
+        array<u32, 2>(2687564964u, 186958263u),
+        array<u32, 2>(1450226471u, 1648586942u),
+        array<u32, 2>(1511122054u, 1595811937u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2426274859u, 4261676319u),
+        array<u32, 2>(1207777283u, 1918363057u),
+        array<u32, 2>(3090099832u, 3870167883u),
+        array<u32, 2>(4101522705u, 1460796764u),
+        array<u32, 2>(201900220u, 4164567654u),
+        array<u32, 2>(2587682901u, 752404845u),
+        array<u32, 2>(2967564913u, 2100296475u),
+        array<u32, 2>(3404347409u, 2242778408u),
+        array<u32, 2>(3350048952u, 1386431957u),
+        array<u32, 2>(4093308564u, 1347177553u),
+        array<u32, 2>(2633812729u, 3169012324u),
+        array<u32, 2>(1727753673u, 3768793234u)
+    )
+);
+
+fn gf_pack(g: GoldilocksField) -> u64 {
+    return (u64(g.limb1) << 32u) | u64(g.limb0);
+}
+
+fn gf_unpack(v: u64) -> GoldilocksField {
+    let c = gf64_canon(v);
+    return GoldilocksField(u32(c & EPS64), u32(c >> 32u));
+}
+
+fn gf_from_limbs(l0: u32, l1: u32) -> GoldilocksField {
+    return GoldilocksField(l0, l1);
+}
+
+fn gf_zero() -> GoldilocksField {
+    return GoldilocksField(0u, 0u);
+}
+
+fn gf_one() -> GoldilocksField {
+    return GoldilocksField(1u, 0u);
+}
+
+fn gf_from_u32(val: u32) -> GoldilocksField {
+    return GoldilocksField(val, 0u);
+}
+
+fn gf_from_u64_parts(low: u32, high: u32) -> GoldilocksField {
+    return gf_unpack((u64(high) << 32u) | u64(low));
+}
+
+fn gf_from_const(val: array<u32, 2>) -> GoldilocksField {
+    return gf_from_u64_parts(val[0], val[1]);
+}
+
+fn gf_add(a: GoldilocksField, b: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_add(gf_pack(a), gf_pack(b)));
+}
+
+fn gf_mul(a: GoldilocksField, b: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_mul(gf_pack(a), gf_pack(b)));
+}
+
+fn sbox(x: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_sbox(gf_pack(x)));
+}
+
+fn state_pack(state: ptr<function, array<GoldilocksField, 12>>, out: ptr<function, array<u64, 12>>) {
+    for (var i = 0u; i < 12u; i++) {
+        (*out)[i] = gf_pack((*state)[i]);
+    }
+}
+
+fn state_unpack(v: ptr<function, array<u64, 12>>, state: ptr<function, array<GoldilocksField, 12>>) {
+    for (var i = 0u; i < 12u; i++) {
+        (*state)[i] = gf_unpack((*v)[i]);
+    }
+}
+
+fn external_linear_layer(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    ext_layer64(&st);
+    state_unpack(&st, state);
+}
+
+fn internal_linear_layer(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    int_layer64(&st);
+    state_unpack(&st, state);
+}
+
+fn poseidon2_permute(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    permute64(&st);
+    state_unpack(&st, state);
+}
+
+fn bytes_to_field_elements(input: array<u32, 24>) -> array<GoldilocksField, 25> {
+    var felts: array<GoldilocksField, 25>;
+    for (var i = 0u; i < 24u; i++) {
+        felts[i] = gf_from_u32(input[i]);
+    }
+    felts[24] = gf_one();
+    return felts;
+}
+
+fn field_elements_to_bytes(felts: array<GoldilocksField, 4>) -> array<u32, 8> {
+    var result: array<u32, 8>;
+    for (var i = 0u; i < 4u; i++) {
+        result[i * 2u] = felts[i].limb0;
+        result[i * 2u + 1u] = felts[i].limb1;
+    }
+    return result;
+}
+
+fn poseidon2_hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
+    var st: array<u64, 12>;
+    for (var i = 0u; i < 12u; i++) {
+        st[i] = 0lu;
+    }
+    for (var chunk = 0u; chunk < 3u; chunk++) {
+        for (var i = 0u; i < 8u; i++) {
+            st[i] = gf64_add(st[i], u64(input[chunk * 8u + i]));
+        }
+        permute64(&st);
+    }
+    st[0] = gf64_add(st[0], 1lu);
+    st[1] = gf64_add(st[1], 1lu);
+    permute64(&st);
+
+    var result: array<u32, 16>;
+    for (var i = 0u; i < 4u; i++) {
+        let c = gf64_canon(st[i]);
+        result[2u * i] = u32(c & EPS64);
+        result[2u * i + 1u] = u32(c >> 32u);
+    }
+    permute64(&st);
+    for (var i = 0u; i < 4u; i++) {
+        let c = gf64_canon(st[i]);
+        result[8u + 2u * i] = u32(c & EPS64);
+        result[8u + 2u * i + 1u] = u32(c >> 32u);
+    }
+    return result;
+}
+
+fn hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
+    return poseidon2_hash_squeeze_twice(input);
+}
@@ -0,0 +1,650 @@
+// Native-u64 Poseidon2 mining kernel. Apple Metal path only.
+// Requires wgpu Features::SHADER_INT64. Bit-exact with mining.wgsl / pow_core.
+
+@group(0) @binding(0) var<storage, read_write> results: array<atomic<u32>>;
+// Sponge state after absorbing header + high nonce half (12 felts as LE u32 pairs),
+// precomputed on the host per batch. See pow_core::mining_midstate.
+@group(0) @binding(1) var<storage, read> midstate: array<u32, 24>;
+@group(0) @binding(2) var<storage, read> start_nonce: array<u32, 16>;
+@group(0) @binding(3) var<storage, read> difficulty_target: array<u32, 16>;
+@group(0) @binding(4) var<storage, read> dispatch_config: array<u32, 3>;
+
+const P64: u64 = 0xFFFFFFFF00000001lu;
+// EPS64 = 2^32 - 1 = 2^64 mod P
+const EPS64: u64 = 0xFFFFFFFFlu;
+
+// Padded with a trailing zero so the internal layer can add the next round's
+// constant unconditionally.
+const RC_INTERNAL: array<u64, 23> = array<u64, 23>(
+    0x97f7798a784ad863lu, 0xd1d2bf082f60d4f0lu, 0x69a377a79f9ad206lu, 0xa9d06906a3858e24lu, 0x295275001eede5b5lu, 0x5874e441117bd746lu, 0x8a084bbba8ed86cclu, 0x3defd7645cde6425lu, 0x3998cfe6871cc137lu, 0x3e52ef8bca48314alu, 0x964a209f85dc9ecclu, 0x3fcc9ee82cc4577elu, 0x8e79b4a5d0096d6dlu, 0x8492362ad2392556lu, 0xee72f470262574d6lu, 0x1e0e18496da2444alu, 0x0f3a74bf215eaac6lu, 0x1b061b76a1c0ded3lu, 0x192c42d86803d7a6lu, 0xf6d49ff997ae0260lu, 0x3ec372e7a0fa3786lu, 0x5538cdf4f23445d3lu, 0lu
+);
+const RC_INITIAL: array<array<u64, 12>, 4> = array<array<u64, 12>, 4>(
+    array<u64, 12>(0xc002e770975b1607lu, 0xbca51a8dfe14593alu, 0x72938dfbe774f7f9lu, 0xe4f2fe29e03234aclu, 0xd5e0ba2f541b6449lu, 0xec33b868f3cc46c1lu, 0x486dcb55419d475alu, 0x6c1cb2a358cc24f1lu, 0xe3f30d509a1436bblu, 0xd9a64f068dca7c29lu, 0xe59b3f57aabba1aelu, 0x2a3dd4505b478fdclu),
+    array<u64, 12>(0xada1f8dc7676ed25lu, 0x2711aa8b5509d516lu, 0x4ae6acd0c9c92897lu, 0x56eb3d6b5256d67alu, 0x1f7a9d55923bf51elu, 0x3600427d397a7f68lu, 0xe5076df75b72c3d0lu, 0xfcd59aa12c6090adlu, 0xcd895e8c68b57a9elu, 0x41df7ef9d730ae3elu, 0xee3e2b889abe977dlu, 0xd29bb7edbeb9c405lu),
+    array<u64, 12>(0x7d5c08eef608e382lu, 0x89ae889caaf0802clu, 0xb35a8e976d2af617lu, 0xdb14234eafaf5173lu, 0x78f04462d48b1c98lu, 0x265293b0e47ce88alu, 0x999a649b69b9d32flu, 0x64b0a186698e01d3lu, 0xee0b22d0dfae8bb8lu, 0x4fd53e50ca04a7eelu, 0x5762bfe181f25047lu, 0xf51593e2beb5e3bdlu),
+    array<u64, 12>(0x1e5e2b5760e32477lu, 0x622462a1f9aaaeedlu, 0xaa284b3ecdb222aelu, 0x63c8e72f542bf3fclu, 0x3ba588cacb43b5e0lu, 0x23eda6f3c99150ddlu, 0xaad3bea4baac9a5alu, 0xe9da8d699b94184alu, 0xcdb13f4cd93e024clu, 0x902cbd0956f655e3lu, 0x5b4e40ffc759532flu, 0xde795c20a2357af7lu)
+);
+const RC_TERMINAL: array<array<u64, 12>, 4> = array<array<u64, 12>, 4>(
+    array<u64, 12>(0x7b72c539e0ea4c6elu, 0x144573dae2ce9976lu, 0x802028b68f35fc88lu, 0x6d36c5022c4fe7c2lu, 0xa205d0ffa9b9def3lu, 0xf6e7e38b1ea6ba2flu, 0x34f7909ae5258d64lu, 0xb0464d9d77b97fcalu, 0x64ddb9d5de7e00a6lu, 0x0ed0d75c27975d97lu, 0x1cbb36f11127338blu, 0x6673e505cfd0b6balu),
+    array<u64, 12>(0x605f902830872e01lu, 0x3fd5eb927e95fe4flu, 0xe81025b5a24c69cdlu, 0xf7d0ce75de23f74elu, 0xf39942b6a8585089lu, 0x6d808a08f7b71df6lu, 0xf8806b6588f49a8blu, 0x57df2d8c2a32107alu, 0x16e7c2074d654a2dlu, 0x213de241fcf33835lu, 0xb0f2b8905a0976f6lu, 0xd8e3cf2bbd355417lu),
+    array<u64, 12>(0xe498691679d9330flu, 0x763b45d2a3821b28lu, 0x0908bf65eb0a1f0dlu, 0x7691eb2d194b24f4lu, 0x0e43551233ae13b2lu, 0x93c393dbfc2fe76flu, 0x98f607485d48cdealu, 0xe3d95f30309819c0lu, 0x1ef581a93eaf6acflu, 0x0b24c1b7a030fca4lu, 0x624370be5670b327lu, 0x5f1e28615a11e486lu),
+    array<u64, 12>(0xfe04051f909e042blu, 0x7257e5b147fd3803lu, 0xe6ae134bb82f2e78lu, 0x5711fd5cf4784511lu, 0xf83a42660c08c0bclu, 0x2cd8c96d9a3ce855lu, 0x7d2ffb1bb0e17271lu, 0x85ae1528caea3811lu, 0x52a345d5c7adb0b8lu, 0x504c4c51f3faee94lu, 0xbce34a649cfccaf9lu, 0xe0a3389266fb6dc9lu)
+);
+const MDS_DIAG: array<u64, 12> = array<u64, 12>(
+    0xc3b6c08e23ba9300lu, 0xd84b5de94a324fb6lu, 0x0d0c371c5b35b84flu, 0x7964f570e7188037lu, 0x5daf18bbd996604blu, 0x6743bc47b9595257lu, 0x5528b9362c59bb70lu, 0xac45e25b7127b68blu, 0xa2077d7dfbb606b5lu, 0xf3faac6faee378aelu, 0x0c6388b51545e883lu, 0xd27dbb6944917b60lu
+);
+
+// a + b mod P in lazy form. Wrapping carries fold back via 2^64 ≡ EPS64 (mod P).
+fn gf64_add(a: u64, b: u64) -> u64 {
+    let s0 = a + b;
+    let c1 = s0 < a;
+    let s1 = s0 + select(0lu, EPS64, c1);
+    let c2 = c1 && (s1 < s0);
+    return s1 + select(0lu, EPS64, c2);
+}
+
+// Sum accumulator: value = lo + carries * 2^64, folded once via 2^64 ≡ EPS64 (mod P).
+struct Acc {
+    lo: u64,
+    carries: u32,
+}
+
+fn acc_add(a: Acc, b: u64) -> Acc {
+    let s = a.lo + b;
+    return Acc(s, a.carries + select(0u, 1u, s < a.lo));
+}
+
+fn acc_add2(a: Acc, b: Acc) -> Acc {
+    let s = a.lo + b.lo;
+    return Acc(s, a.carries + b.carries + select(0u, 1u, s < a.lo));
+}
+
+fn acc_fold(a: Acc) -> u64 {
+    let c = u64(a.carries);
+    let t = a.lo + ((c << 32u) - c);
+    return t + select(0lu, EPS64, t < a.lo);
+}
+
+struct U128 {
+    lo: u64,
+    hi: u64,
+}
+
+// Reduce a 128-bit value (lo + hi*2^64) mod P using
+// 2^64 ≡ EPS64 and 2^96 ≡ -1 (mod P).
+fn gf64_reduce(v: U128) -> u64 {
+    let hi_hi = v.hi >> 32u;
+    let hi_lo = v.hi & EPS64;
+    var t0 = v.lo - hi_hi;
+    t0 = t0 - select(0lu, EPS64, v.lo < hi_hi);
+    let t1 = hi_lo * EPS64;
+    let t2 = t0 + t1;
+    return t2 + select(0lu, EPS64, t2 < t0);
+}
+
+fn mul_wide(a: u64, b: u64) -> U128 {
+    let a_lo = a & EPS64;
+    let a_hi = a >> 32u;
+    let b_lo = b & EPS64;
+    let b_hi = b >> 32u;
+    let ll = a_lo * b_lo;
+    let lh = a_lo * b_hi;
+    let hl = a_hi * b_lo;
+    let hh = a_hi * b_hi;
+    let mid = (ll >> 32u) + (lh & EPS64) + (hl & EPS64);
+    return U128((mid << 32u) | (ll & EPS64), hh + (lh >> 32u) + (hl >> 32u) + (mid >> 32u));
+}
+
+// (a*b + addend) mod P for b <= 2^64 - 2^32 (all MDS_DIAG entries): the addend's
+// value and carries are folded into the 128-bit product before reduction.
+fn gf64_mul_add(a: u64, b: u64, addend: Acc) -> u64 {
+    let v = mul_wide(a, b);
+    let lo = v.lo + addend.lo;
+    return gf64_reduce(U128(lo, v.hi + u64(addend.carries) + select(0lu, 1lu, lo < v.lo)));
+}
+
+fn gf64_mul(a: u64, b: u64) -> u64 {
+    return gf64_reduce(mul_wide(a, b));
+}
+
+fn gf64_sqr(a: u64) -> u64 {
+    let a_lo = a & EPS64;
+    let a_hi = a >> 32u;
+    let ll = a_lo * a_lo;
+    let lh = a_lo * a_hi;
+    let hh = a_hi * a_hi;
+    let mid = (ll >> 32u) + ((lh & EPS64) << 1u);
+    return gf64_reduce(U128((mid << 32u) | (ll & EPS64), hh + ((lh >> 32u) << 1u) + (mid >> 32u)));
+}
+
+fn gf64_sbox(x: u64) -> u64 {
+    let x2 = gf64_sqr(x);
+    let x4 = gf64_sqr(x2);
+    let x6 = gf64_mul(x4, x2);
+    return gf64_mul(x6, x);
+}
+
+fn gf64_canon(a: u64) -> u64 {
+    return a - select(0lu, P64, a >= P64);
+}
+
+// 4x4 MDS circ(2, 3, 1, 1) on one chunk, results left unreduced.
+fn mds4(x0: u64, x1: u64, x2: u64, x3: u64) -> array<Acc, 4> {
+    let t01 = acc_add(Acc(x0, 0u), x1);
+    let t23 = acc_add(Acc(x2, 0u), x3);
+    let t0123 = acc_add2(t01, t23);
+    let t01123 = acc_add(t0123, x1);
+    let t01233 = acc_add(t0123, x3);
+    return array<Acc, 4>(
+        acc_add2(t01123, t01),
+        acc_add2(t01123, acc_add(Acc(x2, 0u), x2)),
+        acc_add2(t01233, t23),
+        acc_add2(t01233, acc_add(Acc(x0, 0u), x0))
+    );
+}
+
+// External linear layer: 4x4 MDS on each chunk, then circulant sums, plus the
+// next round's constants. Additions are accumulated unreduced (at most 27
+// carries) and folded once per output.
+fn ext_layer64(state: ptr<function, array<u64, 12>>, rc: array<u64, 12>) {
+    var y: array<Acc, 12>;
+    for (var chunk = 0u; chunk < 3u; chunk++) {
+        let o = chunk * 4u;
+        let m = mds4((*state)[o], (*state)[o + 1u], (*state)[o + 2u], (*state)[o + 3u]);
+        y[o] = m[0];
+        y[o + 1u] = m[1];
+        y[o + 2u] = m[2];
+        y[o + 3u] = m[3];
+    }
+    for (var k = 0u; k < 4u; k++) {
+        let s = acc_add2(acc_add2(y[k], y[k + 4u]), y[k + 8u]);
+        (*state)[k] = acc_fold(acc_add(acc_add2(y[k], s), rc[k]));
+        (*state)[k + 4u] = acc_fold(acc_add(acc_add2(y[k + 4u], s), rc[k + 4u]));
+        (*state)[k + 8u] = acc_fold(acc_add(acc_add2(y[k + 8u], s), rc[k + 8u]));
+    }
+}
+
+// Internal linear layer: diagonal matrix plus full sum, plus the next round's
+// constant on element 0 (the S-box output, which is summed last).
+fn int_layer64(state: ptr<function, array<u64, 12>>, rc0: u64) {
+    let s12 = acc_add(Acc((*state)[1], 0u), (*state)[2]);
+    let s34 = acc_add(Acc((*state)[3], 0u), (*state)[4]);
+    let s56 = acc_add(Acc((*state)[5], 0u), (*state)[6]);
+    let s78 = acc_add(Acc((*state)[7], 0u), (*state)[8]);
+    let s910 = acc_add(Acc((*state)[9], 0u), (*state)[10]);
+    let s1234 = acc_add2(s12, s34);
+    let s5678 = acc_add2(s56, s78);
+    let s91011 = acc_add(s910, (*state)[11]);
+    let sum = acc_add(acc_add2(acc_add2(s1234, s5678), s91011), (*state)[0]);
+    (*state)[0] = gf64_mul_add((*state)[0], MDS_DIAG[0], acc_add(sum, rc0));
+    (*state)[1] = gf64_mul_add((*state)[1], MDS_DIAG[1], sum);
+    (*state)[2] = gf64_mul_add((*state)[2], MDS_DIAG[2], sum);
+    (*state)[3] = gf64_mul_add((*state)[3], MDS_DIAG[3], sum);
+    (*state)[4] = gf64_mul_add((*state)[4], MDS_DIAG[4], sum);
+    (*state)[5] = gf64_mul_add((*state)[5], MDS_DIAG[5], sum);
+    (*state)[6] = gf64_mul_add((*state)[6], MDS_DIAG[6], sum);
+    (*state)[7] = gf64_mul_add((*state)[7], MDS_DIAG[7], sum);
+    (*state)[8] = gf64_mul_add((*state)[8], MDS_DIAG[8], sum);
+    (*state)[9] = gf64_mul_add((*state)[9], MDS_DIAG[9], sum);
+    (*state)[10] = gf64_mul_add((*state)[10], MDS_DIAG[10], sum);
+    (*state)[11] = gf64_mul_add((*state)[11], MDS_DIAG[11], sum);
+}
+
+fn sbox_lanes(state: ptr<function, array<u64, 12>>, lanes: u32) {
+    for (var i = 0u; i < lanes; i++) {
+        (*state)[i] = gf64_sbox((*state)[i]);
+    }
+}
+
+fn add_rc(state: ptr<function, array<u64, 12>>, rc: array<u64, 12>) {
+    for (var i = 0u; i < 12u; i++) {
+        (*state)[i] = gf64_add((*state)[i], rc[i]);
+    }
+}
+
+const RC_ZERO: array<u64, 12> = array<u64, 12>(0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu);
+// Constants added by each external layer, which feeds the S-box that follows it:
+// entries 0..3 precede the initial rounds, entry 4 carries the first internal
+// round's constant on element 0, entries 5..7 precede terminal rounds 1..3.
+const RC_EXT: array<array<u64, 12>, 9> = array<array<u64, 12>, 9>(
+    RC_INITIAL[0], RC_INITIAL[1], RC_INITIAL[2], RC_INITIAL[3],
+    array<u64, 12>(RC_INTERNAL[0], 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu, 0lu),
+    RC_TERMINAL[1], RC_TERMINAL[2], RC_TERMINAL[3], RC_ZERO
+);
+
+// One loop drives all 30 rounds so each layer is emitted once: step 0 is the
+// initial external layer, steps 1..4 and 27..30 are external rounds, steps
+// 5..26 are internal rounds. Round constants are added by the linear layer
+// preceding each S-box (the last internal layer is followed by the first
+// terminal round's constants).
+fn permute64(state: ptr<function, array<u64, 12>>) {
+    for (var k = 0u; k < 31u; k++) {
+        let is_ext = k < 5u || k > 26u;
+        if (k > 0u) {
+            sbox_lanes(state, select(1u, 12u, is_ext));
+        }
+        if (is_ext) {
+            ext_layer64(state, RC_EXT[select(k, k - 22u, k > 26u)]);
+        } else {
+            int_layer64(state, RC_INTERNAL[k - 4u]);
+            if (k == 26u) {
+                add_rc(state, RC_TERMINAL[0]);
+            }
+        }
+    }
+}
+
+fn bswap32(v: u32) -> u32 {
+    return ((v & 0xFFu) << 24u) | ((v & 0xFF00u) << 8u) | ((v >> 8u) & 0xFF00u) | (v >> 24u);
+}
+
+@compute @workgroup_size(256)
+fn mining_main(@builtin(global_invocation_id) global_id: vec3<u32>) {
+    if (atomicLoad(&results[0]) != 0u) {
+        return;
+    }
+    let thread_id = global_id.x;
+    let total_threads = dispatch_config[0];
+    let nonces_per_thread = dispatch_config[1];
+    let total_nonces = dispatch_config[2];
+    if (thread_id >= total_threads) {
+        return;
+    }
+    let base_index = thread_id * nonces_per_thread;
+
+    // Hoist uniform storage reads out of the nonce loop
+    var mid: array<u64, 12>;
+    for (var i = 0u; i < 12u; i++) {
+        mid[i] = (u64(midstate[2u * i + 1u]) << 32u) | u64(midstate[2u * i]);
+    }
+    var tgt: array<u32, 16>;
+    for (var i = 0u; i < 16u; i++) {
+        tgt[i] = difficulty_target[i];
+    }
+    var nonce_base: array<u32, 16>;
+    for (var i = 0u; i < 16u; i++) {
+        nonce_base[i] = start_nonce[i];
+    }
+
+    for (var j = 0u; j < nonces_per_thread; j = j + 1u) {
+        let logical_index = base_index + j;
+        if (logical_index >= total_nonces) {
+            break;
+        }
+        if (j > 0u && atomicLoad(&results[0]) != 0u) {
+            return;
+        }
+
+        // The host guarantees a batch never carries into the high nonce half
+        // (limbs 8..15), so only the low 256 bits are incremented here.
+        var current_nonce: array<u32, 16>;
+        let val0 = nonce_base[0];
+        let sum0 = val0 + logical_index;
+        current_nonce[0] = sum0;
+        var carry = select(0u, 1u, sum0 < val0);
+        for (var i = 1u; i < 8u; i++) {
+            let val = nonce_base[i];
+            let sum = val + carry;
+            current_nonce[i] = sum;
+            carry = select(0u, 1u, sum < val);
+        }
+        for (var i = 8u; i < 16u; i++) {
+            current_nonce[i] = nonce_base[i];
+        }
+
+        // Resume the sponge from the precomputed midstate: absorb the low
+        // nonce half, pad, squeeze twice (3 permutations instead of 5).
+        var st = mid;
+        for (var i = 0u; i < 8u; i++) {
+            st[i] = gf64_add(st[i], u64(bswap32(current_nonce[7u - i])));
+        }
+        // Squeeze-and-compare phases share one inlined permutation: phase 0
+        // pads after absorbing, phase 1 yields the most significant 256 bits of
+        // the hash, which decide hash-vs-target on their own unless they exactly
+        // equal the target's high half, and only candidates run phase 2 for the
+        // low half. Byte-swapped hash words are produced on demand.
+        var hash_le: array<u32, 16>;
+        var cmp = 0u;
+        var below = false;
+        for (var phase = 0u; phase < 3u; phase++) {
+            permute64(&st);
+            if (phase == 0u) {
+                st[0] = gf64_add(st[0], 1lu);
+                st[1] = gf64_add(st[1], 1lu);
+                continue;
+            }
+            var words: array<u32, 8>;
+            for (var i = 0u; i < 4u; i++) {
+                let c = gf64_canon(st[i]);
+                words[2u * i] = bswap32(u32(c & EPS64));
+                words[2u * i + 1u] = bswap32(u32(c >> 32u));
+            }
+            let base = select(15u, 7u, phase == 2u);
+            for (var i = 0u; i < 8u; i++) {
+                hash_le[base - i] = words[i];
+            }
+            if (phase == 1u) {
+                for (var i = 0u; i < 8u; i++) {
+                    let h = words[i];
+                    let t = tgt[15u - i];
+                    if (h != t) {
+                        cmp = select(2u, 1u, h > t);
+                        break;
+                    }
+                }
+                if (cmp == 1u) {
+                    break;
+                }
+            } else {
+                below = cmp == 2u;
+                if (!below) {
+                    for (var i = 0u; i < 8u; i++) {
+                        let h = words[i];
+                        let t = tgt[7u - i];
+                        if (h != t) {
+                            below = h < t;
+                            break;
+                        }
+                    }
+                }
+            }
+        }
+        if (cmp == 1u) {
+            continue;
+        }
+
+        if (below) {
+            if (atomicExchange(&results[0], 1u) == 0u) {
+                for (var i = 0u; i < 16u; i++) {
+                    atomicStore(&results[1u + i], current_nonce[i]);
+                    atomicStore(&results[17u + i], hash_le[i]);
+                }
+            }
+            return;
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Compatibility layer: same API as mining.wgsl, backed by the u64 core above.
+// Only used by the component test harness; the mining kernel never calls it.
+// All outputs are canonical, matching the reference implementation.
+// ---------------------------------------------------------------------------
+
+struct GoldilocksField {
+    limb0: u32,
+    limb1: u32,
+}
+
+const INTERNAL_CONSTANTS: array<array<u32, 2>, 22> = array<array<u32, 2>, 22>(
+    array<u32, 2>(2018170979u, 2549578122u),
+    array<u32, 2>(794875120u, 3520249608u),
+    array<u32, 2>(2677723654u, 1772320679u),
+    array<u32, 2>(2743438884u, 2849007878u),
+    array<u32, 2>(518907317u, 693269760u),
+    array<u32, 2>(293328710u, 1484055617u),
+    array<u32, 2>(2834138828u, 2315799483u),
+    array<u32, 2>(1558078501u, 1039128420u),
+    array<u32, 2>(2266808631u, 966316006u),
+    array<u32, 2>(3393728842u, 1045622667u),
+    array<u32, 2>(2245828300u, 2521440415u),
+    array<u32, 2>(751064958u, 1070374632u),
+    array<u32, 2>(3490278765u, 2390340773u),
+    array<u32, 2>(3526960470u, 2224174634u),
+    array<u32, 2>(639988950u, 4000511088u),
+    array<u32, 2>(1839350858u, 504240201u),
+    array<u32, 2>(559852230u, 255489215u),
+    array<u32, 2>(2713771731u, 453385078u),
+    array<u32, 2>(1745082278u, 422331096u),
+    array<u32, 2>(2544763488u, 4141129721u),
+    array<u32, 2>(2700752774u, 1052996327u),
+    array<u32, 2>(4063512019u, 1429786100u)
+);
+
+const INITIAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2539329031u, 3221415792u),
+        array<u32, 2>(4262746426u, 3164936845u),
+        array<u32, 2>(3883202553u, 1922272763u),
+        array<u32, 2>(3761386668u, 3841130025u),
+        array<u32, 2>(1411081289u, 3588274735u),
+        array<u32, 2>(4090250945u, 3962812520u),
+        array<u32, 2>(1100826458u, 1215155029u),
+        array<u32, 2>(1489773809u, 1813820067u),
+        array<u32, 2>(2585015995u, 3824356688u),
+        array<u32, 2>(2378857513u, 3651555078u),
+        array<u32, 2>(2864423342u, 3852156759u),
+        array<u32, 2>(1531416540u, 708695120u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(1987505445u, 2913073372u),
+        array<u32, 2>(1426707734u, 655469195u),
+        array<u32, 2>(3385403543u, 1256631504u),
+        array<u32, 2>(1381422714u, 1458257259u),
+        array<u32, 2>(2453402910u, 528129365u),
+        array<u32, 2>(964329320u, 905986685u),
+        array<u32, 2>(1534247888u, 3842469367u),
+        array<u32, 2>(744525997u, 4241857185u),
+        array<u32, 2>(1756723870u, 3448331916u),
+        array<u32, 2>(3610291774u, 1105166073u),
+        array<u32, 2>(2596181885u, 3997051784u),
+        array<u32, 2>(3199845381u, 3533420525u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(4127777666u, 2103183598u),
+        array<u32, 2>(2867888172u, 2309916828u),
+        array<u32, 2>(1831532055u, 3009056407u),
+        array<u32, 2>(2947502451u, 3675530062u),
+        array<u32, 2>(3565886616u, 2029012066u),
+        array<u32, 2>(3833391242u, 642945968u),
+        array<u32, 2>(1773785903u, 2577032347u),
+        array<u32, 2>(1770914259u, 1689297286u),
+        array<u32, 2>(3752758200u, 3993707216u),
+        array<u32, 2>(3389302766u, 1339375184u),
+        array<u32, 2>(2180141127u, 1466089441u),
+        array<u32, 2>(3199591357u, 4111832034u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(1625498743u, 509487959u),
+        array<u32, 2>(4188712685u, 1646551713u),
+        array<u32, 2>(3451003566u, 2854767422u),
+        array<u32, 2>(1412166652u, 1674110767u),
+        array<u32, 2>(3410212320u, 1000704202u),
+        array<u32, 2>(3381743837u, 602777331u),
+        array<u32, 2>(3131873882u, 2866003620u),
+        array<u32, 2>(2610174026u, 3923414377u),
+        array<u32, 2>(3644719692u, 3450945356u),
+        array<u32, 2>(1458984419u, 2418851081u),
+        array<u32, 2>(3344519983u, 1531855103u),
+        array<u32, 2>(2721413879u, 3732495392u)
+    )
+);
+
+const TERMINAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(3773451374u, 2071119161u),
+        array<u32, 2>(3805190518u, 340095962u),
+        array<u32, 2>(2402679944u, 2149591222u),
+        array<u32, 2>(743434178u, 1832305922u),
+        array<u32, 2>(2847530739u, 2718290175u),
+        array<u32, 2>(514243119u, 4142392203u),
+        array<u32, 2>(3844443492u, 888639642u),
+        array<u32, 2>(2008645578u, 2957397405u),
+        array<u32, 2>(3732799654u, 1692252629u),
+        array<u32, 2>(664231319u, 248567644u),
+        array<u32, 2>(287781771u, 482031345u),
+        array<u32, 2>(3486561978u, 1718871301u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(814165505u, 1616875560u),
+        array<u32, 2>(2123759183u, 1070984082u),
+        array<u32, 2>(2722916813u, 3893372341u),
+        array<u32, 2>(3726899022u, 4157656693u),
+        array<u32, 2>(2824360073u, 4086907574u),
+        array<u32, 2>(4155973110u, 1837140488u),
+        array<u32, 2>(2297731723u, 4169165669u),
+        array<u32, 2>(707924090u, 1474243980u),
+        array<u32, 2>(1298483757u, 384287239u),
+        array<u32, 2>(4243798069u, 557703745u),
+        array<u32, 2>(1510569718u, 2968696976u),
+        array<u32, 2>(3174388759u, 3638808363u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2044277519u, 3835193622u),
+        array<u32, 2>(2743212840u, 1983595986u),
+        array<u32, 2>(3943309069u, 151568229u),
+        array<u32, 2>(424355060u, 1989274413u),
+        array<u32, 2>(867046322u, 239293714u),
+        array<u32, 2>(4230997871u, 2479068123u),
+        array<u32, 2>(1565052394u, 2566260552u),
+        array<u32, 2>(815274432u, 3822673712u),
+        array<u32, 2>(1051683535u, 519405993u),
+        array<u32, 2>(2687564964u, 186958263u),
+        array<u32, 2>(1450226471u, 1648586942u),
+        array<u32, 2>(1511122054u, 1595811937u)
+    ),
+    array<array<u32, 2>, 12>(
+        array<u32, 2>(2426274859u, 4261676319u),
+        array<u32, 2>(1207777283u, 1918363057u),
+        array<u32, 2>(3090099832u, 3870167883u),
+        array<u32, 2>(4101522705u, 1460796764u),
+        array<u32, 2>(201900220u, 4164567654u),
+        array<u32, 2>(2587682901u, 752404845u),
+        array<u32, 2>(2967564913u, 2100296475u),
+        array<u32, 2>(3404347409u, 2242778408u),
+        array<u32, 2>(3350048952u, 1386431957u),
+        array<u32, 2>(4093308564u, 1347177553u),
+        array<u32, 2>(2633812729u, 3169012324u),
+        array<u32, 2>(1727753673u, 3768793234u)
+    )
+);
+
+fn gf_pack(g: GoldilocksField) -> u64 {
+    return (u64(g.limb1) << 32u) | u64(g.limb0);
+}
+
+fn gf_unpack(v: u64) -> GoldilocksField {
+    let c = gf64_canon(v);
+    return GoldilocksField(u32(c & EPS64), u32(c >> 32u));
+}
+
+fn gf_from_limbs(l0: u32, l1: u32) -> GoldilocksField {
+    return GoldilocksField(l0, l1);
+}
+
+fn gf_zero() -> GoldilocksField {
+    return GoldilocksField(0u, 0u);
+}
+
+fn gf_one() -> GoldilocksField {
+    return GoldilocksField(1u, 0u);
+}
+
+fn gf_from_u32(val: u32) -> GoldilocksField {
+    return GoldilocksField(val, 0u);
+}
+
+fn gf_from_u64_parts(low: u32, high: u32) -> GoldilocksField {
+    return gf_unpack((u64(high) << 32u) | u64(low));
+}
+
+fn gf_from_const(val: array<u32, 2>) -> GoldilocksField {
+    return gf_from_u64_parts(val[0], val[1]);
+}
+
+fn gf_add(a: GoldilocksField, b: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_add(gf_pack(a), gf_pack(b)));
+}
+
+fn gf_mul(a: GoldilocksField, b: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_mul(gf_pack(a), gf_pack(b)));
+}
+
+fn sbox(x: GoldilocksField) -> GoldilocksField {
+    return gf_unpack(gf64_sbox(gf_pack(x)));
+}
+
+fn state_pack(state: ptr<function, array<GoldilocksField, 12>>, out: ptr<function, array<u64, 12>>) {
+    for (var i = 0u; i < 12u; i++) {
+        (*out)[i] = gf_pack((*state)[i]);
+    }
+}
+
+fn state_unpack(v: ptr<function, array<u64, 12>>, state: ptr<function, array<GoldilocksField, 12>>) {
+    for (var i = 0u; i < 12u; i++) {
+        (*state)[i] = gf_unpack((*v)[i]);
+    }
+}
+
+fn external_linear_layer(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    ext_layer64(&st, RC_ZERO);
+    state_unpack(&st, state);
+}
+
+fn internal_linear_layer(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    int_layer64(&st, 0lu);
+    state_unpack(&st, state);
+}
+
+fn poseidon2_permute(state: ptr<function, array<GoldilocksField, 12>>) {
+    var st: array<u64, 12>;
+    state_pack(state, &st);
+    permute64(&st);
+    state_unpack(&st, state);
+}
+
+fn bytes_to_field_elements(input: array<u32, 24>) -> array<GoldilocksField, 25> {
+    var felts: array<GoldilocksField, 25>;
+    for (var i = 0u; i < 24u; i++) {
+        felts[i] = gf_from_u32(input[i]);
+    }
+    felts[24] = gf_one();
+    return felts;
+}
+
+fn field_elements_to_bytes(felts: array<GoldilocksField, 4>) -> array<u32, 8> {
+    var result: array<u32, 8>;
+    for (var i = 0u; i < 4u; i++) {
+        result[i * 2u] = felts[i].limb0;
+        result[i * 2u + 1u] = felts[i].limb1;
+    }
+    return result;
+}
+
+fn poseidon2_hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
+    var st: array<u64, 12>;
+    for (var i = 0u; i < 12u; i++) {
+        st[i] = 0lu;
+    }
+    for (var chunk = 0u; chunk < 3u; chunk++) {
+        for (var i = 0u; i < 8u; i++) {
+            st[i] = gf64_add(st[i], u64(input[chunk * 8u + i]));
+        }
+        permute64(&st);
+    }
+    st[0] = gf64_add(st[0], 1lu);
+    st[1] = gf64_add(st[1], 1lu);
+    permute64(&st);
+
+    var result: array<u32, 16>;
+    for (var i = 0u; i < 4u; i++) {
+        let c = gf64_canon(st[i]);
+        result[2u * i] = u32(c & EPS64);
+        result[2u * i + 1u] = u32(c >> 32u);
+    }
+    permute64(&st);
+    for (var i = 0u; i < 4u; i++) {
+        let c = gf64_canon(st[i]);
+        result[8u + 2u * i] = u32(c & EPS64);
+        result[8u + 2u * i + 1u] = u32(c >> 32u);
+    }
+    return result;
+}
+
+fn hash_squeeze_twice(input: array<u32, 24>) -> array<u32, 16> {
+    return poseidon2_hash_squeeze_twice(input);
+}
@@ -0,0 +1,116 @@
+//! Poseidon2 mining kernels.
+//!
+//! Same `mining_main` bindings; must stay bit-exact with `pow_core`.
+//!
+//! - Apple Metal + `SHADER_INT64` → Apple Metal u64 (`mining_u64_apple.wgsl`)
+//! - other GPUs + `SHADER_INT64` → native u64 (`mining_u64.wgsl`)
+//! - no `SHADER_INT64` → 32-bit fallback (`mining.wgsl`)
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum Kernel {
+    U32,
+    Default,
+    Apple,
+}
+
+impl Kernel {
+    pub const fn needs_int64(self) -> bool {
+        !matches!(self, Self::U32)
+    }
+
+    pub const fn id(self) -> &'static str {
+        match self {
+            Self::U32 => "u32",
+            Self::Default => "u64",
+            Self::Apple => "u64-apple",
+        }
+    }
+
+    pub const fn label(self) -> &'static str {
+        match self {
+            Self::U32 => "32-bit",
+            Self::Default => "native-u64",
+            Self::Apple => "native-u64 Apple Metal",
+        }
+    }
+
+    pub const fn source(self) -> &'static str {
+        match self {
+            Self::U32 => include_str!("mining.wgsl"),
+            Self::Default => include_str!("mining_u64.wgsl"),
+            Self::Apple => include_str!("mining_u64_apple.wgsl"),
+        }
+    }
+
+    pub fn for_adapter(adapter: &wgpu::Adapter) -> Self {
+        Self::for_adapter_info(&adapter.get_info(), adapter.features())
+    }
+
+    pub fn for_adapter_info(info: &wgpu::AdapterInfo, features: wgpu::Features) -> Self {
+        if !features.contains(wgpu::Features::SHADER_INT64) {
+            return Self::U32;
+        }
+        if info.backend == wgpu::Backend::Metal {
+            Self::Apple
+        } else {
+            Self::Default
+        }
+    }
+
+    pub const fn all() -> &'static [Self] {
+        &[Self::U32, Self::Default, Self::Apple]
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn info(backend: wgpu::Backend) -> wgpu::AdapterInfo {
+        wgpu::AdapterInfo {
+            name: "test".into(),
+            vendor: 0,
+            device: 0,
+            device_type: wgpu::DeviceType::DiscreteGpu,
+            driver: String::new(),
+            driver_info: String::new(),
+            backend,
+        }
+    }
+
+    #[test]
+    fn metal_with_int64_selects_apple() {
+        assert_eq!(
+            Kernel::for_adapter_info(&info(wgpu::Backend::Metal), wgpu::Features::SHADER_INT64),
+            Kernel::Apple
+        );
+    }
+
+    #[test]
+    fn vulkan_with_int64_selects_default() {
+        assert_eq!(
+            Kernel::for_adapter_info(&info(wgpu::Backend::Vulkan), wgpu::Features::SHADER_INT64),
+            Kernel::Default
+        );
+    }
+
+    #[test]
+    fn dx12_with_int64_selects_default() {
+        assert_eq!(
+            Kernel::for_adapter_info(&info(wgpu::Backend::Dx12), wgpu::Features::SHADER_INT64),
+            Kernel::Default
+        );
+    }
+
+    #[test]
+    fn no_int64_falls_back_to_u32() {
+        assert_eq!(
+            Kernel::for_adapter_info(&info(wgpu::Backend::Metal), wgpu::Features::empty()),
+            Kernel::U32
+        );
+        assert_eq!(
+            Kernel::for_adapter_info(&info(wgpu::Backend::Vulkan), wgpu::Features::empty()),
+            Kernel::U32
+        );
+    }
+}
@@ -1,177 +1,132 @@
-use futures::executor::block_on;
-
 mod end_to_end_tests;
 mod tests;

-fn main() {
-    block_on(run()).unwrap();
-}
-
-async fn run() -> Result<(), Box<dyn std::error::Error>> {
-    // Setup GPU device and queue
+#[tokio::main]
+async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let instance = wgpu::Instance::new(&wgpu::InstanceDescriptor {
-        backends: wgpu::Backends::METAL, // Force Metal on Apple
+        backends: wgpu::Backends::PRIMARY,
        ..Default::default()
    });
    let adapter = instance
        .request_adapter(&wgpu::RequestAdapterOptions::default())
        .await
-        .unwrap();
-    let (device, queue) = adapter
-        .request_device(&wgpu::DeviceDescriptor::default())
-        .await?;
+        .expect("no GPU adapter");

-    println!("Running Poseidon2 GPU Component Tests...\n");
+    let mut failures = 0usize;
+    let has_int64 = adapter.features().contains(wgpu::Features::SHADER_INT64);

-    // Run all component tests
-    if let Err(e) = tests::test_gf_from_const(&device, &queue).await {
-        eprintln!("❌ gf_from_const tests failed: {}", e);
+    for kernel in engine_gpu::Kernel::all() {
+        if kernel.needs_int64() && !has_int64 {
+            println!(
+                "\nSHADER_INT64 not supported on this adapter; skipping {}",
+                kernel.label()
+            );
+            continue;
+        }
+        let (device, queue) = adapter
+            .request_device(&wgpu::DeviceDescriptor {
+                label: Some(kernel.label()),
+                required_features: if kernel.needs_int64() {
+                    wgpu::Features::SHADER_INT64
+                } else {
+                    wgpu::Features::empty()
+                },
+                ..Default::default()
+            })
+            .await?;
+        failures += run_suite(&device, &queue, kernel.source(), kernel.label()).await;
    }

-    if let Err(e) = tests::test_gf_mul(&device, &queue).await {
-        eprintln!("❌ gf_mul tests failed: {}", e);
+    if failures > 0 {
+        return Err(format!("{failures} test group(s) failed").into());
    }
-
-    if let Err(e) = tests::test_sbox(&device, &queue).await {
-        eprintln!("❌ S-box tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_mds_matrix(&device, &queue).await {
-        eprintln!("❌ MDS matrix tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_internal_linear_layer(&device, &queue).await {
-        eprintln!("❌ Internal linear layer tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_external_linear_layer(&device, &queue).await {
-        eprintln!("❌ External linear layer tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_initial_external_rounds(&device, &queue).await {
-        eprintln!("❌ Initial external rounds tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_terminal_external_rounds(&device, &queue).await {
-        eprintln!("❌ Terminal external rounds tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_constants_verification(&device, &queue).await {
-        eprintln!("❌ Constants verification test failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_internal_constants_verification(&device, &queue).await {
-        eprintln!("❌ Internal constants verification test failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_internal_rounds_only(&device, &queue).await {
-        eprintln!("❌ Internal rounds only test failed: {}", e);
-    }
-
-    if let Err(e) =
-        tests::test_poseidon2_terminal_external_constants_verification(&device, &queue).await
-    {
-        eprintln!(
-            "❌ Terminal external constants verification test failed: {}",
-            e
-        );
-    }
-
-    if let Err(e) = tests::test_poseidon2_permutation(&device, &queue).await {
-        eprintln!("❌ Poseidon2 permutation tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_bytes_to_field_elements(&device, &queue).await {
-        eprintln!("❌ Bytes to field elements tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_field_elements_to_bytes(&device, &queue).await {
-        eprintln!("❌ Field elements to bytes tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_poseidon2_squeeze_twice(&device, &queue).await {
-        eprintln!("❌ Poseidon2 squeeze-twice tests failed: {}", e);
-    }
-
-    if let Err(e) = tests::test_double_hash(&device, &queue).await {
-        eprintln!("❌ Double hash tests failed: {}", e);
-    }
-
-    if let Err(e) = end_to_end_tests::test_end_to_end_mining(&device, &queue).await {
-        eprintln!("❌ End-to-end mining test failed: {}", e);
-    }
-
    println!("\nAll tests completed!");
-    if let Err(e) = end_to_end_tests::test_end_to_end_mining(&device, &queue).await {
-        eprintln!("❌ End-to-end mining test failed: {}", e);
-    }
-
-    println!("\nAll tests completed!");
-
-    // generate_correct_wgsl_constants();
    Ok(())
 }

-#[allow(dead_code)]
-fn generate_correct_wgsl_constants() {
-    use qp_poseidon_constants::*;
-
-    println!("🔧 Generating correct WGSL constants...");
-
-    println!("// Initial external round constants (4 rounds x 12 elements)");
-    println!("const INITIAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(");
-
-    for (round_idx, round) in POSEIDON2_INITIAL_EXTERNAL_CONSTANTS_RAW.iter().enumerate() {
-        println!("    array<array<u32, 2>, 12>(");
-        for (elem_idx, &value) in round.iter().enumerate() {
-            let low = value as u32;
-            let high = (value >> 32) as u32;
-            if elem_idx == 11 {
-                println!("        array<u32, 2>({}u, {}u)", low, high);
-            } else {
-                println!("        array<u32, 2>({}u, {}u),", low, high);
+async fn run_suite(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    shader_src: &str,
+    label: &str,
+) -> usize {
+    println!("\n==== Running Poseidon2 GPU tests against {label} ====\n");
+    let mut failures = 0usize;
+    macro_rules! run {
+        ($name:literal, $fut:expr) => {
+            if let Err(e) = $fut.await {
+                eprintln!("❌ {} failed: {}", $name, e);
+                failures += 1;
            }
-        }
-        if round_idx == 3 {
-            println!("    )");
-        } else {
-            println!("    ),");
-        }
+        };
    }
-    println!(");");

-    println!("\n// Terminal external round constants (4 rounds x 12 elements)");
-    println!("const TERMINAL_EXTERNAL_CONSTANTS: array<array<array<u32, 2>, 12>, 4> = array<array<array<u32, 2>, 12>, 4>(");
+    run!(
+        "gf_from_const",
+        tests::test_gf_from_const(device, queue, shader_src)
+    );
+    run!("gf_mul", tests::test_gf_mul(device, queue, shader_src));
+    run!("sbox", tests::test_sbox(device, queue, shader_src));
+    run!(
+        "mds matrix",
+        tests::test_mds_matrix(device, queue, shader_src)
+    );
+    run!(
+        "internal linear layer",
+        tests::test_internal_linear_layer(device, queue, shader_src)
+    );
+    run!(
+        "external linear layer",
+        tests::test_external_linear_layer(device, queue, shader_src)
+    );
+    run!(
+        "initial external rounds",
+        tests::test_poseidon2_initial_external_rounds(device, queue, shader_src)
+    );
+    run!(
+        "terminal external rounds",
+        tests::test_poseidon2_terminal_external_rounds(device, queue, shader_src)
+    );
+    run!(
+        "constants verification",
+        tests::test_poseidon2_constants_verification(device, queue, shader_src)
+    );
+    run!(
+        "internal constants verification",
+        tests::test_poseidon2_internal_constants_verification(device, queue, shader_src)
+    );
+    run!(
+        "internal rounds only",
+        tests::test_poseidon2_internal_rounds_only(device, queue, shader_src)
+    );
+    run!(
+        "terminal external constants verification",
+        tests::test_poseidon2_terminal_external_constants_verification(device, queue, shader_src)
+    );
+    run!(
+        "poseidon2 permutation",
+        tests::test_poseidon2_permutation(device, queue, shader_src)
+    );
+    run!(
+        "bytes to field elements",
+        tests::test_bytes_to_field_elements(device, queue, shader_src)
+    );
+    run!(
+        "field elements to bytes",
+        tests::test_field_elements_to_bytes(device, queue, shader_src)
+    );
+    run!(
+        "poseidon2 squeeze-twice",
+        tests::test_poseidon2_squeeze_twice(device, queue, shader_src)
+    );
+    run!(
+        "hash squeeze twice",
+        tests::test_hash_squeeze_twice(device, queue, shader_src)
+    );
+    run!(
+        "end-to-end mining",
+        end_to_end_tests::test_end_to_end_mining(device, queue, shader_src)
+    );

-    for (round_idx, round) in POSEIDON2_TERMINAL_EXTERNAL_CONSTANTS_RAW.iter().enumerate() {
-        println!("    array<array<u32, 2>, 12>(");
-        for (elem_idx, &value) in round.iter().enumerate() {
-            let low = value as u32;
-            let high = (value >> 32) as u32;
-            if elem_idx == 11 {
-                println!("        array<u32, 2>({}u, {}u)", low, high);
-            } else {
-                println!("        array<u32, 2>({}u, {}u),", low, high);
-            }
-        }
-        if round_idx == 3 {
-            println!("    )");
-        } else {
-            println!("    ),");
-        }
-    }
-    println!(");");
-
-    println!("\n// Internal round constants (22 values)");
-    println!("const INTERNAL_CONSTANTS: array<array<u32, 2>, 22> = array<array<u32, 2>, 22>(");
-    for (idx, &value) in POSEIDON2_INTERNAL_CONSTANTS_RAW.iter().enumerate() {
-        let low = value as u32;
-        let high = (value >> 32) as u32;
-        if idx == 21 {
-            println!("    array<u32, 2>({}u, {}u)", low, high);
-        } else {
-            println!("    array<u32, 2>({}u, {}u),", low, high);
-        }
-    }
-    println!(");");
+    failures
 }
@@ -1,8 +1,7 @@
 use plonky2::hash::poseidon2::P2Permuter;
 use qp_plonky2_field::goldilocks_field::GoldilocksField;
 use qp_plonky2_field::types::{Field, Field64, PrimeField64};
-
-use qp_poseidon_core::serialization::{p2_backend::GF as P2Goldilocks, GoldiCompat};
+use qp_poseidon_core::Goldilocks as QpGoldilocks;
 use rand::{Rng, SeedableRng};
 use wgpu::util::DeviceExt;

@@ -204,8 +203,8 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {

    // Random small values (both operands < 2^32)
    for _ in 0..50 {
-        let a = GoldilocksField::from_canonical_u64(rng.gen::<u32>() as u64);
-        let b = GoldilocksField::from_canonical_u64(rng.gen::<u32>() as u64);
+        let a = GoldilocksField::from_canonical_u64(rng.random::<u32>() as u64);
+        let b = GoldilocksField::from_canonical_u64(rng.random::<u32>() as u64);
        vectors.push(GfMulTestCase {
            a,
            b,
@@ -215,8 +214,8 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {

    // Random mixed cases (one small, one large)
    for _ in 0..80 {
-        let a = GoldilocksField::from_canonical_u64(rng.gen::<u32>() as u64);
-        let b = GoldilocksField::from_canonical_u64(rng.gen::<u64>());
+        let a = GoldilocksField::from_canonical_u64(rng.random::<u32>() as u64);
+        let b = GoldilocksField::from_canonical_u64(rng.random::<u64>());
        vectors.push(GfMulTestCase {
            a,
            b,
@@ -226,8 +225,8 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {

    // Random large values (both operands >= 2^32)
    for _ in 0..100 {
-        let a = GoldilocksField::from_canonical_u64(rng.gen::<u64>() | 0x100000000u64);
-        let b = GoldilocksField::from_canonical_u64(rng.gen::<u64>() | 0x100000000u64);
+        let a = GoldilocksField::from_canonical_u64(rng.random::<u64>() | 0x100000000u64);
+        let b = GoldilocksField::from_canonical_u64(rng.random::<u64>() | 0x100000000u64);
        vectors.push(GfMulTestCase {
            a,
            b,
@@ -237,9 +236,9 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {

    // Random values near the field modulus
    for _ in 0..30 {
-        let offset = rng.gen::<u32>() as u64;
+        let offset = rng.random::<u32>() as u64;
        let a = GoldilocksField::from_canonical_u64(GoldilocksField::ORDER - offset);
-        let b = GoldilocksField::from_canonical_u64(rng.gen::<u64>());
+        let b = GoldilocksField::from_canonical_u64(rng.random::<u64>());
        vectors.push(GfMulTestCase {
            a,
            b,
@@ -250,7 +249,7 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {
    // Add more specific high-limb test cases to detect patterns
    for i in 20..40 {
        let val1 = 1u64 << i; // Powers of 2 from 2^20 to 2^39
-        let val2 = rng.gen::<u32>() as u64 | (1u64 << 32); // Random value with high bit set
+        let val2 = rng.random::<u32>() as u64 | (1u64 << 32); // Random value with high bit set
        let a = GoldilocksField::from_canonical_u64(val1);
        let b = GoldilocksField::from_canonical_u64(val2);
        vectors.push(GfMulTestCase {
@@ -262,9 +261,9 @@ fn generate_gf_mul_test_vectors() -> Vec<GfMulTestCase> {

    // Test values with specific limb patterns
    for _ in 0..20 {
-        let high_val = (rng.gen::<u32>() as u64) << 32;
+        let high_val = (rng.random::<u32>() as u64) << 32;
        let a = GoldilocksField::from_canonical_u64(high_val);
-        let b = GoldilocksField::from_canonical_u64(rng.gen::<u64>() | 0x100000000u64);
+        let b = GoldilocksField::from_canonical_u64(rng.random::<u64>() | 0x100000000u64);
        vectors.push(GfMulTestCase {
            a,
            b,
@@ -331,21 +330,21 @@ fn generate_sbox_test_vectors() -> Vec<SboxTestCase> {

    // Random small values (< 2^32)
    for _ in 0..20 {
-        let input = GoldilocksField::from_canonical_u64(rng.gen::<u32>() as u64);
+        let input = GoldilocksField::from_canonical_u64(rng.random::<u32>() as u64);
        let expected = input.exp_u64(7);
        vectors.push(SboxTestCase { input, expected });
    }

    // Random large values (>= 2^32)
    for _ in 0..30 {
-        let input = GoldilocksField::from_canonical_u64(rng.gen::<u64>() | 0x100000000u64);
+        let input = GoldilocksField::from_canonical_u64(rng.random::<u64>() | 0x100000000u64);
        let expected = input.exp_u64(7);
        vectors.push(SboxTestCase { input, expected });
    }

    // Values near the field modulus
    for _ in 0..10 {
-        let offset = rng.gen::<u32>() as u64;
+        let offset = rng.random::<u32>() as u64;
        let input = GoldilocksField::from_canonical_u64(GoldilocksField::ORDER - offset);
        let expected = input.exp_u64(7);
        vectors.push(SboxTestCase { input, expected });
@@ -387,7 +386,7 @@ fn generate_external_linear_layer_test_vectors() -> Vec<ExternalLinearLayerTestC
    for _ in 0..15 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen::<u16>() as u64;
+            let val = rng.random::<u16>() as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = apply_full_external_linear_layer_to_state(&input);
@@ -398,7 +397,7 @@ fn generate_external_linear_layer_test_vectors() -> Vec<ExternalLinearLayerTestC
    for _ in 0..10 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen::<u32>() as u64;
+            let val = rng.random::<u32>() as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = apply_full_external_linear_layer_to_state(&input);
@@ -432,7 +431,7 @@ fn generate_internal_linear_layer_test_vectors() -> Vec<InternalLinearLayerTestC
    for _ in 0..10 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            *item = GoldilocksField::from_canonical_u64(rng.gen_range(0..100));
+            *item = GoldilocksField::from_canonical_u64(rng.random_range(0..100));
        }
        let expected = apply_internal_linear_layer_to_state(&input);
        vectors.push(InternalLinearLayerTestCase { input, expected });
@@ -442,7 +441,7 @@ fn generate_internal_linear_layer_test_vectors() -> Vec<InternalLinearLayerTestC
    for _ in 0..20 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            *item = GoldilocksField::from_noncanonical_u64(rng.gen());
+            *item = GoldilocksField::from_noncanonical_u64(rng.random());
        }
        let expected = apply_internal_linear_layer_to_state(&input);
        vectors.push(InternalLinearLayerTestCase { input, expected });
@@ -487,7 +486,7 @@ fn generate_mds_test_vectors() -> Vec<MdsTestCase> {
    for _ in 0..50 {
        let mut input = [GoldilocksField::ZERO; 4];
        for item in &mut input {
-            *item = GoldilocksField::from_canonical_u64(rng.gen::<u16>() as u64);
+            *item = GoldilocksField::from_canonical_u64(rng.random::<u16>() as u64);
        }
        let expected = apply_external_linear_layer_to_chunk(&input);
        vectors.push(MdsTestCase { input, expected });
@@ -497,7 +496,7 @@ fn generate_mds_test_vectors() -> Vec<MdsTestCase> {
    for _ in 0..30 {
        let mut input = [GoldilocksField::ZERO; 4];
        for item in &mut input {
-            *item = GoldilocksField::from_canonical_u64(rng.gen::<u32>() as u64);
+            *item = GoldilocksField::from_canonical_u64(rng.random::<u32>() as u64);
        }
        let expected = apply_external_linear_layer_to_chunk(&input);
        vectors.push(MdsTestCase { input, expected });
@@ -507,7 +506,7 @@ fn generate_mds_test_vectors() -> Vec<MdsTestCase> {
    for _ in 0..30 {
        let mut input = [GoldilocksField::ZERO; 4];
        for item in &mut input {
-            *item = GoldilocksField::from_canonical_u64(rng.gen::<u64>());
+            *item = GoldilocksField::from_canonical_u64(rng.random::<u64>());
        }
        let expected = apply_external_linear_layer_to_chunk(&input);
        vectors.push(MdsTestCase { input, expected });
@@ -517,7 +516,7 @@ fn generate_mds_test_vectors() -> Vec<MdsTestCase> {
    for _ in 0..10 {
        let mut input = [GoldilocksField::ZERO; 4];
        for item in &mut input {
-            let offset = rng.gen::<u32>() as u64;
+            let offset = rng.random::<u32>() as u64;
            *item = GoldilocksField::from_canonical_u64(GoldilocksField::ORDER - offset);
        }
        let expected = apply_external_linear_layer_to_chunk(&input);
@@ -710,7 +709,7 @@ fn generate_poseidon2_test_vectors() -> Vec<Poseidon2TestCase> {
    for _ in 0..15 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen::<u16>() as u64;
+            let val = rng.random::<u16>() as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = <GoldilocksField as P2Permuter>::permute(input);
@@ -721,7 +720,7 @@ fn generate_poseidon2_test_vectors() -> Vec<Poseidon2TestCase> {
    for _ in 0..20 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen::<u16>() as u64;
+            let val = rng.random::<u16>() as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = <GoldilocksField as P2Permuter>::permute(input);
@@ -732,7 +731,7 @@ fn generate_poseidon2_test_vectors() -> Vec<Poseidon2TestCase> {
    for _ in 0..15 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen::<u32>() as u64;
+            let val = rng.random::<u32>() as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = <GoldilocksField as P2Permuter>::permute(input);
@@ -744,7 +743,7 @@ fn generate_poseidon2_test_vectors() -> Vec<Poseidon2TestCase> {
    for _ in 0..15 {
        let mut input = [GoldilocksField::ZERO; 12];
        for item in &mut input {
-            let val = rng.gen_range(65536..u32::MAX) as u64;
+            let val = rng.random_range(65536..u32::MAX) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let expected = <GoldilocksField as P2Permuter>::permute(input);
@@ -802,10 +801,10 @@ fn generate_field_to_bytes_test_vectors() -> Vec<FieldToBytesTestCase> {
    // Test 4: Random values
    for _ in 0..10 {
        let input = [
-            GoldilocksField::from_canonical_u64(rng.gen::<u64>() % GoldilocksField::ORDER),
-            GoldilocksField::from_canonical_u64(rng.gen::<u64>() % GoldilocksField::ORDER),
-            GoldilocksField::from_canonical_u64(rng.gen::<u64>() % GoldilocksField::ORDER),
-            GoldilocksField::from_canonical_u64(rng.gen::<u64>() % GoldilocksField::ORDER),
+            GoldilocksField::from_canonical_u64(rng.random::<u64>() % GoldilocksField::ORDER),
+            GoldilocksField::from_canonical_u64(rng.random::<u64>() % GoldilocksField::ORDER),
+            GoldilocksField::from_canonical_u64(rng.random::<u64>() % GoldilocksField::ORDER),
+            GoldilocksField::from_canonical_u64(rng.random::<u64>() % GoldilocksField::ORDER),
        ];
        let mut expected = [0u8; 32];
        for (i, &felt) in input.iter().enumerate() {
@@ -888,13 +887,17 @@ fn generate_bytes_to_field_test_vectors() -> Vec<BytesToFieldTestCase> {
    let mut vectors = Vec::new();
    let mut rng = rand_chacha::ChaCha8Rng::seed_from_u64(0x98765432);

+    // Helper to convert qp_poseidon_core::Goldilocks to Plonky2 GoldilocksField
+    fn qp_to_p2(felts: Vec<QpGoldilocks>) -> Vec<GoldilocksField> {
+        felts
+            .into_iter()
+            .map(|g| GoldilocksField::from_noncanonical_u64(g.as_canonical_u64()))
+            .collect()
+    }
+
    // Test 1: All zeros
    let zero_input = [0u8; 96];
-    let zero_expected =
-        qp_poseidon_core::serialization::injective_bytes_to_felts::<P2Goldilocks>(&zero_input)
-            .into_iter()
-            .map(|g| GoldilocksField::from_noncanonical_u64(g.to_u64()))
-            .collect();
+    let zero_expected = qp_to_p2(qp_poseidon_core::serialization::bytes_to_felts(&zero_input));
    vectors.push(BytesToFieldTestCase {
        input_bytes: zero_input,
        expected_felts: zero_expected,
@@ -902,11 +905,7 @@ fn generate_bytes_to_field_test_vectors() -> Vec<BytesToFieldTestCase> {

    // Test 2: All 0xFF
    let ff_input = [0xFFu8; 96];
-    let ff_expected =
-        qp_poseidon_core::serialization::injective_bytes_to_felts::<P2Goldilocks>(&ff_input)
-            .into_iter()
-            .map(|g| GoldilocksField::from_noncanonical_u64(g.to_u64()))
-            .collect();
+    let ff_expected = qp_to_p2(qp_poseidon_core::serialization::bytes_to_felts(&ff_input));
    vectors.push(BytesToFieldTestCase {
        input_bytes: ff_input,
        expected_felts: ff_expected,
@@ -917,11 +916,7 @@ fn generate_bytes_to_field_test_vectors() -> Vec<BytesToFieldTestCase> {
    for (i, byte) in seq_input.iter_mut().enumerate() {
        *byte = (i % 256) as u8;
    }
-    let seq_expected =
-        qp_poseidon_core::serialization::injective_bytes_to_felts::<P2Goldilocks>(&seq_input)
-            .into_iter()
-            .map(|g| GoldilocksField::from_noncanonical_u64(g.to_u64()))
-            .collect();
+    let seq_expected = qp_to_p2(qp_poseidon_core::serialization::bytes_to_felts(&seq_input));
    vectors.push(BytesToFieldTestCase {
        input_bytes: seq_input,
        expected_felts: seq_expected,
@@ -931,11 +926,7 @@ fn generate_bytes_to_field_test_vectors() -> Vec<BytesToFieldTestCase> {
    for _ in 0..10 {
        let mut input = [0u8; 96];
        rng.fill(&mut input[..]);
-        let expected =
-            qp_poseidon_core::serialization::injective_bytes_to_felts::<P2Goldilocks>(&input)
-                .into_iter()
-                .map(|g| GoldilocksField::from_noncanonical_u64(g.to_u64()))
-                .collect();
+        let expected = qp_to_p2(qp_poseidon_core::serialization::bytes_to_felts(&input));
        vectors.push(BytesToFieldTestCase {
            input_bytes: input,
            expected_felts: expected,
@@ -1040,20 +1031,18 @@ fn print_test_failure_hex(
    }
 }

-fn generate_double_hash_test_vectors() -> Vec<DoubleHashTestCase> {
+fn generate_hash_squeeze_twice_test_vectors() -> Vec<DoubleHashTestCase> {
    let mut vectors = Vec::new();
    let mut rng = rand_chacha::ChaCha8Rng::seed_from_u64(0xDEADBEEF);

-    // Test 1-250: Random header and nonce combinations (25 tests)
+    // Test 1-250: Random header and nonce combinations
    for i in 0..250 {
        let mut input = [0u8; 96];

        rng.fill(&mut input[..]);

-        let expected = {
-            let first_hash = qp_poseidon_core::hash_squeeze_twice(&input);
-            qp_poseidon_core::hash_squeeze_twice(&first_hash)
-        };
+        // Single hash with double squeeze (no longer double hash)
+        let expected = qp_poseidon_core::hash_squeeze_twice(&input);

        vectors.push(DoubleHashTestCase {
            input_96_bytes: input,
@@ -1145,7 +1134,7 @@ fn generate_true_internal_only_test_vectors() -> Vec<InternalRoundsTestCase> {
    for _ in 0..15 {
        let mut random_input = [GoldilocksField::ZERO; 12];
        for item in &mut random_input {
-            let val = (rng.gen::<u8>()) as u64;
+            let val = (rng.random::<u8>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let after_initial_external = apply_initial_external(random_input);
@@ -1160,7 +1149,7 @@ fn generate_true_internal_only_test_vectors() -> Vec<InternalRoundsTestCase> {
    for _ in 0..10 {
        let mut random_input = [GoldilocksField::ZERO; 12];
        for item in &mut random_input {
-            let val = (rng.gen::<u32>()) as u64;
+            let val = (rng.random::<u32>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        let after_initial_external = apply_initial_external(random_input);
@@ -1177,6 +1166,7 @@ fn generate_true_internal_only_test_vectors() -> Vec<InternalRoundsTestCase> {
 pub async fn test_poseidon2_permutation(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_poseidon2_test_vectors();
    let total_tests = test_vectors.len();
@@ -1213,7 +1203,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -1381,6 +1371,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_initial_external_rounds(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    // Test just the initial 4 external rounds
    let test_vectors = generate_gf_from_const_test_vectors();
@@ -1426,7 +1417,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -1471,7 +1462,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    for _ in 0..15 {
        let mut state = [GoldilocksField::ZERO; 12];
        for item in &mut state {
-            let val = (rng.gen::<u32>()) as u64;
+            let val = (rng.random::<u32>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        test_cases.push(state);
@@ -1481,7 +1472,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    for _ in 0..10 {
        let mut state = [GoldilocksField::ZERO; 12];
        for item in &mut state {
-            let val = (rng.gen::<u16>()) as u64;
+            let val = (rng.random::<u16>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        test_cases.push(state);
@@ -1641,6 +1632,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_terminal_external_rounds(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    // Test just the terminal 4 external rounds
    let _total_tests = 2;
@@ -1685,7 +1677,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -1730,7 +1722,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    for _ in 0..15 {
        let mut state = [GoldilocksField::ZERO; 12];
        for item in &mut state {
-            let val = (rng.gen::<u8>()) as u64;
+            let val = (rng.random::<u8>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        test_cases.push(state);
@@ -1740,7 +1732,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    for _ in 0..10 {
        let mut state = [GoldilocksField::ZERO; 12];
        for item in &mut state {
-            let val = (rng.gen::<u16>()) as u64;
+            let val = (rng.random::<u16>()) as u64;
            *item = GoldilocksField::from_canonical_u64(val);
        }
        test_cases.push(state);
@@ -1899,6 +1891,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_gf_from_const(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    // Test the gf_from_const function that converts constant arrays to GoldilocksField
    let test_vectors = generate_gf_from_const_test_vectors();
@@ -1919,7 +1912,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    output_fields[index] = gf_from_const(input_constants[index]);
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2034,6 +2027,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_constants_verification(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let mut passed_tests = 0;
    let mut failed_tests = Vec::new();
@@ -2060,7 +2054,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2171,6 +2165,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_internal_rounds_only(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    // Test ONLY the internal rounds (22 rounds) without initial/terminal external rounds
    let test_vectors = generate_true_internal_only_test_vectors();
@@ -2220,7 +2215,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-            include_str!("mining.wgsl")
+            shader_src
        );

        let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2354,6 +2349,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_internal_constants_verification(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let mut passed_tests = 0;
    let mut failed_tests = Vec::new();
@@ -2375,7 +2371,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2473,6 +2469,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_terminal_external_rounds_issue(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    println!("🔍 Testing terminal external rounds issue for sequential input...");

@@ -2513,7 +2510,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2641,6 +2638,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_terminal_external_constants_verification(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let mut passed_tests = 0;
    let mut failed_tests = Vec::new();
@@ -2667,7 +2665,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2775,6 +2773,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_gf_mul(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_gf_mul_test_vectors();
    let total_tests = test_vectors.len();
@@ -2797,7 +2796,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    output[i] = gf_mul(input_a[i], input_b[i]);
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -2958,6 +2957,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_mds_matrix(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_mds_test_vectors();
    let total_tests = test_vectors.len();
@@ -2966,7 +2966,7 @@ pub async fn test_mds_matrix(
    let mut failure_details = Vec::new();

    // Read the mining shader source and create a test wrapper
-    let mining_shader_source = include_str!("mining.wgsl");
+    let mining_shader_source = shader_src;
    let shader_source = format!(
        "
 {}
@@ -3226,6 +3226,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_internal_linear_layer(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_internal_linear_layer_test_vectors();
    let total_tests = test_vectors.len();
@@ -3234,7 +3235,7 @@ pub async fn test_internal_linear_layer(
    let mut failure_details = Vec::new();

    // Read the mining shader source and create a test wrapper
-    let mining_shader_source = include_str!("mining.wgsl");
+    let mining_shader_source = shader_src;
    let shader_source = format!(
        "
 {}
@@ -3459,6 +3460,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_external_linear_layer(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_external_linear_layer_test_vectors();
    let total_tests = test_vectors.len();
@@ -3496,7 +3498,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -3636,6 +3638,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_sbox(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_sbox_test_vectors();
    let total_tests = test_vectors.len();
@@ -3673,7 +3676,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    output_data[index] = result;
 }}
 ",
-            include_str!("mining.wgsl"),
+            shader_src,
            batch.len()
        );

@@ -3818,6 +3821,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_poseidon2_squeeze_twice(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_poseidon2_squeeze_twice_test_vectors();
    let total_tests = test_vectors.len();
@@ -3853,7 +3857,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -3996,6 +4000,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_bytes_to_field_elements(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_bytes_to_field_test_vectors();
    let total_tests = test_vectors.len();
@@ -4031,7 +4036,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -4179,6 +4184,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
 pub async fn test_field_elements_to_bytes(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
    let test_vectors = generate_field_to_bytes_test_vectors();
    let total_tests = test_vectors.len();
@@ -4214,7 +4220,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -4344,11 +4350,12 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    results.print_summary()
 }

-pub async fn test_double_hash(
+pub async fn test_hash_squeeze_twice(
    device: &wgpu::Device,
    queue: &wgpu::Queue,
+    shader_src: &str,
 ) -> Result<(), Box<dyn std::error::Error>> {
-    let test_vectors = generate_double_hash_test_vectors();
+    let test_vectors = generate_hash_squeeze_twice_test_vectors();
    let total_tests = test_vectors.len();
    let mut passed_tests = 0;
    let mut failed_tests = Vec::new();
@@ -4372,8 +4379,8 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
        input_array[i] = input_data[input_offset + i];
    }}

-    // Hash the input using double hash (96 bytes -> 64 bytes)
-    let hash_result = double_hash(input_array);
+    // Hash the input using single hash with double squeeze (96 bytes -> 64 bytes)
+    let hash_result = hash_squeeze_twice(input_array);

    // Write full 64 bytes (16 u32s) of result as the hash
    let output_offset = test_id * 16u;
@@ -4382,16 +4389,16 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {{
    }}
 }}
 ",
-        include_str!("mining.wgsl")
+        shader_src
    );

    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
-        label: Some("Double Hash Test Shader"),
+        label: Some("Hash Squeeze Twice Test Shader"),
        source: wgpu::ShaderSource::Wgsl(shader_source.into()),
    });

    let pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
-        label: Some("Double Hash Test Pipeline"),
+        label: Some("Hash Squeeze Twice Test Pipeline"),
        layout: None,
        module: &shader,
        entry_point: Some("main"),
@@ -6,11 +6,9 @@ publish = false
 description = "Metrics and Prometheus exporter utilities for the Quantus External Miner"

 [features]
-# Provide a Registry by default; HTTP exporter is opt-in.
-default = ["registry"]
-registry = []
+default = []
 # Enable a Warp-based HTTP exporter (used when --metrics-port is provided).
-http-exporter = ["dep:serde", "dep:serde_json", "dep:tokio", "dep:warp"]
+http-exporter = ["dep:tokio", "dep:warp"]

 [dependencies]
 anyhow = { workspace = true }
@@ -21,5 +19,3 @@ prometheus = { version = "0.13", default-features = false }
 # Optional HTTP exporter stack (enabled via `http-exporter` feature)
 warp = { workspace = true, optional = true }
 tokio = { workspace = true, features = ["full"], optional = true }
-serde = { workspace = true, features = ["derive"], optional = true }
-serde_json = { workspace = true, optional = true }
@@ -1,55 +1,91 @@
 #![deny(rust_2018_idioms)]
 #![forbid(unsafe_code)]

-//! Minimal metrics scaffolding for the Quantus External Miner.
+//! Minimal metrics for the Quantus External Miner.
 //!
-//! - Provides a global Prometheus registry and a handful of default metrics.
-//! - Exposes helper functions to update metrics from the service layer.
-//! - Optionally runs a Warp-based HTTP endpoint (/metrics) when the
-//!   `http-exporter` feature is enabled. This is gated at runtime by
-//!   the presence of a metrics port in the CLI (`--metrics-port`).
+//! Exposes a small set of Prometheus metrics for monitoring mining performance:
+//! - `miner_hash_rate`: Total hash rate (CPU + GPU combined)
+//! - `miner_cpu_hash_rate`: CPU-only hash rate
+//! - `miner_gpu_hash_rate`: GPU-only hash rate
+//! - `miner_hashes_total`: Total hashes computed (all time)
+//! - `miner_active_jobs`: Currently running jobs (0 or 1)
+//! - `miner_workers`: Total worker count
+//! - `miner_cpu_workers`: Number of CPU workers
+//! - `miner_gpu_devices`: Number of GPU devices
+//! - `miner_effective_cpus`: Detected CPU cores
 //!
-//! Default metrics:
-//! - miner_jobs_total{status}        : number of jobs by terminal state
-//! - miner_hashes_total              : total nonces tested
-//! - miner_hash_rate                 : current estimated hash rate (nonces/sec)
-//! - miner_http_requests_total{code,endpoint} : HTTP request counts for miner API
-//!
-//! Notes:
-//! - The HTTP exporter is spawned as a background task and does not block.
-//! - If the `http-exporter` feature is disabled, `start_http_exporter` becomes
-//!   a no-op that returns immediately.
-//! - When jobs/threads end, prefer removing gauge label children (series) rather than
-//!   writing a zero; this avoids scrape-timing artifacts and produces cleaner rollups.
-//! - Service-level observability: expose an `active_jobs` gauge and a `mine_requests_total`
-//!   counter (labeled by result) to disambiguate "idle because no jobs" vs "actively mining".
+//! Optionally runs a Warp-based HTTP endpoint (`/metrics`) when the
+//! `http-exporter` feature is enabled.

 use once_cell::sync::Lazy;
-use prometheus::{
-    opts, Encoder, Gauge, GaugeVec, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, Registry,
-    TextEncoder,
-};
+use prometheus::{IntCounter, IntGauge, Registry};
+use std::sync::Mutex;
+use std::time::Instant;

 #[cfg(feature = "http-exporter")]
-use {anyhow::Result, std::net::SocketAddr, warp::Filter};
+use {
+    anyhow::Result,
+    prometheus::{Encoder, TextEncoder},
+    std::net::SocketAddr,
+    warp::Filter,
+};

 #[cfg(not(feature = "http-exporter"))]
 use anyhow::Result;

-use std::collections::{HashMap, HashSet};
-use std::sync::Mutex;
-use std::thread;
-use std::time::{Duration, Instant};
-
-// -------------------------------------------------------------------------------------
-// Global Registry and Default Metrics
-// -------------------------------------------------------------------------------------
+// ---------------------------------------------------------------------------
+// Global Registry
+// ---------------------------------------------------------------------------

 static REGISTRY: Lazy<Registry> = Lazy::new(Registry::new);

-// -------------------------------------------------------------------------------------
-// High-level service metrics
-// -------------------------------------------------------------------------------------
+// ---------------------------------------------------------------------------
+// Hash Rate Metrics
+// ---------------------------------------------------------------------------
+
+static HASH_RATE: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_hash_rate", "Total hash rate in hashes per second")
+        .expect("create miner_hash_rate");
+    REGISTRY
+        .register(Box::new(g.clone()))
+        .expect("register miner_hash_rate");
+    g
+});
+
+static CPU_HASH_RATE: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_cpu_hash_rate", "CPU hash rate in hashes per second")
+        .expect("create miner_cpu_hash_rate");
+    REGISTRY
+        .register(Box::new(g.clone()))
+        .expect("register miner_cpu_hash_rate");
+    g
+});
+
+static GPU_HASH_RATE: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_gpu_hash_rate", "GPU hash rate in hashes per second")
+        .expect("create miner_gpu_hash_rate");
+    REGISTRY
+        .register(Box::new(g.clone()))
+        .expect("register miner_gpu_hash_rate");
+    g
+});
+
+// ---------------------------------------------------------------------------
+// Counter Metrics
+// ---------------------------------------------------------------------------
+
+static HASHES_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
+    let c = IntCounter::new("miner_hashes_total", "Total hashes computed")
+        .expect("create miner_hashes_total");
+    REGISTRY
+        .register(Box::new(c.clone()))
+        .expect("register miner_hashes_total");
+    c
+});
+
+// ---------------------------------------------------------------------------
+// Gauge Metrics
+// ---------------------------------------------------------------------------

 static ACTIVE_JOBS: Lazy<IntGauge> = Lazy::new(|| {
    let g = IntGauge::new(
@@ -63,25 +99,37 @@ static ACTIVE_JOBS: Lazy<IntGauge> = Lazy::new(|| {
    g
 });

-static ENGINE_BACKEND_INFO: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_engine_backend",
-            "Engine backend info (label-only gauge set to 1). Labels: engine, backend"
-        ),
-        &["engine", "backend"],
-    )
-    .expect("create miner_engine_backend");
+static WORKERS: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_workers", "Total number of worker threads")
+        .expect("create miner_workers");
    REGISTRY
        .register(Box::new(g.clone()))
-        .expect("register miner_engine_backend");
+        .expect("register miner_workers");
+    g
+});
+
+static CPU_WORKERS: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_cpu_workers", "Number of CPU worker threads")
+        .expect("create miner_cpu_workers");
+    REGISTRY
+        .register(Box::new(g.clone()))
+        .expect("register miner_cpu_workers");
+    g
+});
+
+static GPU_DEVICES: Lazy<IntGauge> = Lazy::new(|| {
+    let g = IntGauge::new("miner_gpu_devices", "Number of GPU devices")
+        .expect("create miner_gpu_devices");
+    REGISTRY
+        .register(Box::new(g.clone()))
+        .expect("register miner_gpu_devices");
    g
 });

 static EFFECTIVE_CPUS: Lazy<IntGauge> = Lazy::new(|| {
    let g = IntGauge::new(
        "miner_effective_cpus",
-        "Detected logical CPU capacity available to this process (cpuset-aware)",
+        "Detected logical CPU cores available to this process",
    )
    .expect("create miner_effective_cpus");
    REGISTRY
@@ -90,676 +138,167 @@ static EFFECTIVE_CPUS: Lazy<IntGauge> = Lazy::new(|| {
    g
 });

-static MINE_REQUESTS_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_mine_requests_total",
-            "Count of /mine requests by result"
-        ),
-        &["result"], // accepted, duplicate, invalid, error
-    )
-    .expect("create miner_mine_requests_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_mine_requests_total");
-    c
-});
+// ---------------------------------------------------------------------------
+// Hash Rate Tracking
+// ---------------------------------------------------------------------------

-static JOBS_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!("miner_jobs_total", "Number of jobs by status"),
-        &["status"],
-    )
-    .expect("create miner_jobs_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_jobs_total");
-    c
-});
-
-static HASHES_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
-    let c = IntCounter::new("miner_hashes_total", "Total nonces tested")
-        .expect("create miner_hashes_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_hashes_total");
-    c
-});
-
-static HASH_RATE: Lazy<Gauge> = Lazy::new(|| {
-    let g =
-        Gauge::new("miner_hash_rate", "Estimated hash rate (nonces per second)").expect("create");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_hash_rate");
-    g
-});
-
-static HTTP_REQUESTS_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_http_requests_total",
-            "HTTP requests count by endpoint and status code"
-        ),
-        &["endpoint", "code"],
-    )
-    .expect("create miner_http_requests_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_http_requests_total");
-    c
-});
-
-/// Access the global Prometheus registry used by this crate.
-// Engine-aware, per-job, and per-thread labeled metrics
-pub fn set_effective_cpus(n: i64) {
-    EFFECTIVE_CPUS.set(n);
+/// Tracks cumulative hashes to compute rolling hash rates.
+///
+/// The hash rate is computed as total_hashes / elapsed_time since mining started.
+/// Call `reset()` when a new mining session begins to get accurate rates.
+struct HashRateTracker {
+    /// Cumulative CPU hashes since last reset
+    cpu_total: u64,
+    /// Cumulative GPU hashes since last reset
+    gpu_total: u64,
+    /// When tracking started - set via reset() when mining begins
+    start_time: Option<Instant>,
 }

-static JOB_HASHES_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_job_hashes_total",
-            "Total nonces tested per job and engine"
-        ),
-        &["engine", "job_id"],
-    )
-    .expect("create miner_job_hashes_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_job_hashes_total");
-    c
-});
+impl HashRateTracker {
+    fn new() -> Self {
+        Self {
+            cpu_total: 0,
+            gpu_total: 0,
+            start_time: None,
+        }
+    }

-static JOB_HASH_RATE: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_job_hash_rate",
-            "Estimated hash rate (nonces per second) per job and engine"
-        ),
-        &["engine", "job_id"],
-    )
-    .expect("create miner_job_hash_rate");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_job_hash_rate");
-    g
-});
+    /// Reset the tracker - call this when a new mining session starts.
+    fn reset(&mut self) {
+        self.cpu_total = 0;
+        self.gpu_total = 0;
+        self.start_time = Some(Instant::now());
+    }

-static JOB_ESTIMATED_RATE: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_job_estimated_rate",
-            "Estimated work rate (nonces per second) per engine and backend"
-        ),
-        &["engine", "backend"],
-    )
-    .expect("create miner_job_estimated_rate");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_job_estimated_rate");
-    g
-});
+    fn record_cpu(&mut self, hashes: u64) {
+        self.cpu_total += hashes;
+        self.update_rates();
+    }

-static CANDIDATES_FOUND_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_candidates_found_total",
-            "Total candidates found per job and engine"
-        ),
-        &["engine", "job_id"],
-    )
-    .expect("create miner_candidates_found_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_candidates_found_total");
-    c
-});
+    fn record_gpu(&mut self, hashes: u64) {
+        self.gpu_total += hashes;
+        self.update_rates();
+    }

-static CANDIDATES_FALSE_POSITIVE_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_candidates_false_positive_total",
-            "Total false-positive candidates rejected by host re-verification per engine"
-        ),
-        &["engine"],
-    )
-    .expect("create miner_candidates_false_positive_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_candidates_false_positive_total");
-    c
-});
-
-static SAMPLE_MISMATCH_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_sample_mismatch_total",
-            "Total decision parity mismatches between engine and host per engine"
-        ),
-        &["engine"],
-    )
-    .expect("create miner_sample_mismatch_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_sample_mismatch_total");
-    c
-});
-
-static MISSED_WINNER_TOTAL: Lazy<IntCounter> = Lazy::new(|| {
-    let c = IntCounter::new(
-        "miner_gpu_g2_missed_winner_total",
-        "Host detected a winner in G2 batch sample but device did not flag early-exit",
-    )
-    .expect("create miner_gpu_g2_missed_winner_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_gpu_g2_missed_winner_total");
-    c
-});
-
-static FOUND_BY_ORIGIN_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_found_by_origin_total",
-            "Count of candidates found by origin per engine"
-        ),
-        &["engine", "origin"],
-    )
-    .expect("create miner_found_by_origin_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_found_by_origin_total");
-    c
-});
-
-static THREAD_HASHES_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_thread_hashes_total",
-            "Total nonces tested per thread, job, and engine"
-        ),
-        &["engine", "job_id", "thread_id"],
-    )
-    .expect("create miner_thread_hashes_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_thread_hashes_total");
-    c
-});
-
-static THREAD_HASH_RATE: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_thread_hash_rate",
-            "Estimated hash rate (nonces per second) per thread, job, and engine"
-        ),
-        &["engine", "job_id", "thread_id"],
-    )
-    .expect("create miner_thread_hash_rate");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_thread_hash_rate");
-    g
-});
-
-static JOBS_BY_ENGINE_TOTAL: Lazy<IntCounterVec> = Lazy::new(|| {
-    let c = IntCounterVec::new(
-        opts!(
-            "miner_jobs_by_engine_total",
-            "Number of jobs by engine and terminal status"
-        ),
-        &["engine", "status"],
-    )
-    .expect("create miner_jobs_by_engine_total");
-    REGISTRY
-        .register(Box::new(c.clone()))
-        .expect("register miner_jobs_by_engine_total");
-    c
-});
-
-static JOB_STATUS_GAUGE: Lazy<IntGaugeVec> = Lazy::new(|| {
-    let g = IntGaugeVec::new(
-        opts!(
-            "miner_job_status",
-            "Job status gauge (set to 1 for current status)"
-        ),
-        &["engine", "job_id", "status"],
-    )
-    .expect("create miner_job_status");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_job_status");
-    g
-});
-
-static JOB_FOUND_ORIGIN: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_job_found_origin",
-            "Per-job found candidate origin (0=unknown, 1=cpu, 2=gpu-g1, 3=gpu-g2)"
-        ),
-        &["engine", "job_id", "origin"],
-    )
-    .expect("create miner_job_found_origin");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_job_found_origin");
-    g
-});
-
-// -------------------------------------------------------------------------------------
-// GPU-specific metrics
-// -------------------------------------------------------------------------------------
-
-static GPU_DEVICE_COUNT: Lazy<IntGauge> = Lazy::new(|| {
-    let g = IntGauge::new(
-        "miner_gpu_devices_total",
-        "Number of GPU devices available for mining",
-    )
-    .expect("create miner_gpu_devices_total");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_gpu_devices_total");
-    g
-});
-
-static GPU_DEVICE_INFO: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_gpu_device_info",
-            "GPU device information (label-only gauge set to 1). Labels: device_id, name, backend, vendor, device_type"
-        ),
-        &["device_id", "name", "backend", "vendor", "device_type"],
-    )
-    .expect("create miner_gpu_device_info");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_gpu_device_info");
-    g
-});
-
-static GPU_BATCH_SIZE: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_gpu_batch_size",
-            "Current GPU batch size (hashes per batch) per device"
-        ),
-        &["device_id"],
-    )
-    .expect("create miner_gpu_batch_size");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_gpu_batch_size");
-    g
-});
-
-static GPU_WORKGROUPS: Lazy<GaugeVec> = Lazy::new(|| {
-    let g = GaugeVec::new(
-        opts!(
-            "miner_gpu_workgroups",
-            "Number of GPU workgroups dispatched per device"
-        ),
-        &["device_id"],
-    )
-    .expect("create miner_gpu_workgroups");
-    REGISTRY
-        .register(Box::new(g.clone()))
-        .expect("register miner_gpu_workgroups");
-    g
-});
-
-pub fn default_registry() -> &'static Registry {
-    &REGISTRY
+    fn update_rates(&self) {
+        if let Some(start) = self.start_time {
+            let elapsed = start.elapsed().as_secs_f64();
+            // Require at least 0.1 seconds of data for meaningful rate
+            if elapsed >= 0.1 {
+                let cpu_rate = (self.cpu_total as f64 / elapsed) as i64;
+                let gpu_rate = (self.gpu_total as f64 / elapsed) as i64;
+                CPU_HASH_RATE.set(cpu_rate);
+                GPU_HASH_RATE.set(gpu_rate);
+                HASH_RATE.set(cpu_rate + gpu_rate);
+            }
+        }
+    }
 }

-// -------------------------------------------------------------------------------------
-// Update Helpers (to be called from the service layer)
-// -------------------------------------------------------------------------------------
+static HASH_TRACKER: Lazy<Mutex<HashRateTracker>> =
+    Lazy::new(|| Mutex::new(HashRateTracker::new()));

-/// Increment the total hashes counter by `n` (number of nonces tested).
-// Labeled helpers for engine/job/thread metrics
-//
-// Service-level helpers
-pub fn set_active_jobs(n: i64) {
-    ACTIVE_JOBS.set(n);
+// ---------------------------------------------------------------------------
+// Public API - Hash Recording
+// ---------------------------------------------------------------------------
+
+/// Reset the hash rate tracker - call this when mining starts.
+/// This resets the cumulative hash counts and starts a fresh timing window.
+pub fn reset_hash_tracker() {
+    if let Ok(mut tracker) = HASH_TRACKER.lock() {
+        tracker.reset();
+    }
 }

-pub fn set_engine_backend(engine: &str, backend: &str) {
-    ENGINE_BACKEND_INFO
-        .with_label_values(&[engine, backend])
-        .set(1.0);
+/// Record CPU hashes and update hash rate metrics.
+pub fn record_cpu_hashes(n: u64) {
+    HASHES_TOTAL.inc_by(n);
+    if let Ok(mut tracker) = HASH_TRACKER.lock() {
+        tracker.record_cpu(n);
+    }
 }

-pub fn inc_mine_requests(result: &str) {
-    MINE_REQUESTS_TOTAL.with_label_values(&[result]).inc();
+/// Record GPU hashes and update hash rate metrics.
+pub fn record_gpu_hashes(n: u64) {
+    HASHES_TOTAL.inc_by(n);
+    if let Ok(mut tracker) = HASH_TRACKER.lock() {
+        tracker.record_gpu(n);
+    }
 }

-pub fn inc_job_hashes(engine: &str, job_id: &str, n: u64) {
-    touch_job(engine, job_id);
-    JOB_HASHES_TOTAL
-        .with_label_values(&[engine, job_id])
-        .inc_by(n);
+// ---------------------------------------------------------------------------
+// Public API - Hash Rates (for direct setting, kept for compatibility)
+// ---------------------------------------------------------------------------
+
+/// Set the total hash rate (CPU + GPU combined).
+pub fn set_hash_rate(rate: i64) {
+    HASH_RATE.set(rate);
 }

-pub fn set_job_hash_rate(engine: &str, job_id: &str, rate: f64) {
-    touch_job(engine, job_id);
-    JOB_HASH_RATE.with_label_values(&[engine, job_id]).set(rate);
+/// Set the CPU-only hash rate.
+pub fn set_cpu_hash_rate(rate: i64) {
+    CPU_HASH_RATE.set(rate);
 }

-pub fn set_job_estimated_rate_backend(engine: &str, backend: &str, rate: f64) {
-    JOB_ESTIMATED_RATE
-        .with_label_values(&[engine, backend])
-        .set(rate);
+/// Set the GPU-only hash rate.
+pub fn set_gpu_hash_rate(rate: i64) {
+    GPU_HASH_RATE.set(rate);
 }

-pub fn set_job_estimated_rate(engine: &str, _job_id: &str, rate: f64) {
-    // Compatibility shim: map per-job series to backend="unknown" to avoid cardinality growth
-    JOB_ESTIMATED_RATE
-        .with_label_values(&[engine, "unknown"])
-        .set(rate);
-}
-
-pub fn inc_candidates_found(engine: &str, job_id: &str) {
-    touch_job(engine, job_id);
-    CANDIDATES_FOUND_TOTAL
-        .with_label_values(&[engine, job_id])
-        .inc();
-}
-
-pub fn inc_candidates_false_positive(engine: &str) {
-    CANDIDATES_FALSE_POSITIVE_TOTAL
-        .with_label_values(&[engine])
-        .inc();
-}
-
-pub fn inc_sample_mismatch(engine: &str) {
-    SAMPLE_MISMATCH_TOTAL.with_label_values(&[engine]).inc();
-}
-
-pub fn inc_found_by_origin(engine: &str, origin: &str) {
-    FOUND_BY_ORIGIN_TOTAL
-        .with_label_values(&[engine, origin])
-        .inc();
-}
-
-pub fn inc_gpu_g2_missed_winner() {
-    MISSED_WINNER_TOTAL.inc();
-}
-
-pub fn set_job_found_origin(engine: &str, job_id: &str, origin: &str) {
-    touch_job(engine, job_id);
-    JOB_FOUND_ORIGIN
-        .with_label_values(&[engine, job_id, origin])
-        .set(1.0);
-}
-
-pub fn job_estimated_rate_backend(engine: &str, backend: &str, rate: f64) {
-    JOB_ESTIMATED_RATE
-        .with_label_values(&[engine, backend])
-        .set(rate);
-}
-
-pub fn job_estimated_rate(engine: &str, _job_id: &str, rate: f64) {
-    // Compatibility shim: prefer new API using backend; fallback to backend="unknown"
-    JOB_ESTIMATED_RATE
-        .with_label_values(&[engine, "unknown"])
-        .set(rate);
-}
-
-pub fn inc_thread_hashes(engine: &str, job_id: &str, thread_id: &str, n: u64) {
-    THREAD_HASHES_TOTAL
-        .with_label_values(&[engine, job_id, thread_id])
-        .inc_by(n);
-}
-
-pub fn set_thread_hash_rate(engine: &str, job_id: &str, thread_id: &str, rate: f64) {
-    touch_thread(engine, job_id, thread_id);
-    THREAD_HASH_RATE
-        .with_label_values(&[engine, job_id, thread_id])
-        .set(rate);
-}
-
-pub fn inc_jobs_by_engine(engine: &str, status: &str) {
-    JOBS_BY_ENGINE_TOTAL
-        .with_label_values(&[engine, status])
-        .inc();
-}
-
-pub fn set_job_status_gauge(engine: &str, job_id: &str, status: &str, value: i64) {
-    touch_job(engine, job_id);
-    JOB_STATUS_GAUGE
-        .with_label_values(&[engine, job_id, status])
-        .set(value);
-}
+// ---------------------------------------------------------------------------
+// Public API - Counters (kept for compatibility)
+// ---------------------------------------------------------------------------

+/// Increment the total hashes counter (without updating rates).
+/// Prefer `record_cpu_hashes` or `record_gpu_hashes` instead.
 pub fn inc_hashes(n: u64) {
    HASHES_TOTAL.inc_by(n);
 }

-/// Set the current estimated hash rate (nonces per second).
-pub fn set_hash_rate(rate: f64) {
-    HASH_RATE.set(rate);
+// ---------------------------------------------------------------------------
+// Public API - Gauges
+// ---------------------------------------------------------------------------
+
+/// Set the number of active jobs (0 or 1).
+pub fn set_active_jobs(n: i64) {
+    ACTIVE_JOBS.set(n);
 }

-/// Increment the jobs counter for a terminal status: completed | failed | cancelled.
-pub fn inc_job_status(status: &str) {
-    JOBS_TOTAL.with_label_values(&[status]).inc();
+/// Set the total number of workers.
+pub fn set_workers(n: i64) {
+    WORKERS.set(n);
 }

-/// Increment HTTP request counters for an endpoint with a status code.
-pub fn inc_http_request(endpoint: &str, code: u16) {
-    HTTP_REQUESTS_TOTAL
-        .with_label_values(&[endpoint, &code.to_string()])
-        .inc();
+/// Set the number of CPU workers.
+pub fn set_cpu_workers(n: i64) {
+    CPU_WORKERS.set(n);
 }

-// -------------------------------------------------------------------------------------
-// Removal helpers for end-of-life series
-// -------------------------------------------------------------------------------------
-
-const METRICS_TTL_SECS: u64 = 300;
-const JANITOR_INTERVAL_SECS: u64 = 60;
-
-type JobKey = (String, String); // (engine, job_id)
-type ThreadKey = (String, String, String); // (engine, job_id, thread_id)
-
-static JOB_KEYS: Lazy<Mutex<HashSet<JobKey>>> = Lazy::new(|| Mutex::new(HashSet::new()));
-static JOB_LAST: Lazy<Mutex<HashMap<JobKey, Instant>>> = Lazy::new(|| Mutex::new(HashMap::new()));
-static THREAD_KEYS: Lazy<Mutex<HashSet<ThreadKey>>> = Lazy::new(|| Mutex::new(HashSet::new()));
-static THREAD_LAST: Lazy<Mutex<HashMap<ThreadKey, Instant>>> =
-    Lazy::new(|| Mutex::new(HashMap::new()));
-
-static JANITOR_INIT: Lazy<()> = Lazy::new(|| {
-    thread::spawn(|| loop {
-        thread::sleep(Duration::from_secs(JANITOR_INTERVAL_SECS));
-        prune_stale();
-    });
-});
-
-fn prune_stale() {
-    let now = Instant::now();
-    // Jobs
-    let mut to_remove_jobs: Vec<JobKey> = Vec::new();
-    {
-        let last = JOB_LAST.lock().unwrap();
-        for (k, ts) in last.iter() {
-            if now.duration_since(*ts).as_secs() > METRICS_TTL_SECS {
-                to_remove_jobs.push(k.clone());
-            }
-        }
-    }
-    for (engine, job_id) in to_remove_jobs {
-        let _ = JOB_HASHES_TOTAL.remove_label_values(&[&engine, &job_id]);
-        let _ = JOB_HASH_RATE.remove_label_values(&[&engine, &job_id]);
-        let _ = CANDIDATES_FOUND_TOTAL.remove_label_values(&[&engine, &job_id]);
-        let _ = JOB_STATUS_GAUGE.remove_label_values(&[&engine, &job_id, "completed"]);
-        let _ = JOB_STATUS_GAUGE.remove_label_values(&[&engine, &job_id, "failed"]);
-        let _ = JOB_STATUS_GAUGE.remove_label_values(&[&engine, &job_id, "cancelled"]);
-        // Remove per-job origin gauge variants (known origins)
-        let _ = JOB_FOUND_ORIGIN.remove_label_values(&[&engine, &job_id, "cpu"]);
-        let _ = JOB_FOUND_ORIGIN.remove_label_values(&[&engine, &job_id, "gpu-g1"]);
-        let _ = JOB_FOUND_ORIGIN.remove_label_values(&[&engine, &job_id, "gpu-g2"]);
-        JOB_KEYS
-            .lock()
-            .unwrap()
-            .remove(&(engine.clone(), job_id.clone()));
-        JOB_LAST.lock().unwrap().remove(&(engine, job_id));
-    }
-    // Threads
-    let mut to_remove_threads: Vec<ThreadKey> = Vec::new();
-    {
-        let last = THREAD_LAST.lock().unwrap();
-        for (k, ts) in last.iter() {
-            if now.duration_since(*ts).as_secs() > METRICS_TTL_SECS {
-                to_remove_threads.push(k.clone());
-            }
-        }
-    }
-    for (engine, job_id, thread_id) in to_remove_threads {
-        let _ = THREAD_HASHES_TOTAL.remove_label_values(&[&engine, &job_id, &thread_id]);
-        let _ = THREAD_HASH_RATE.remove_label_values(&[&engine, &job_id, &thread_id]);
-        THREAD_KEYS
-            .lock()
-            .unwrap()
-            .remove(&(engine.clone(), job_id.clone(), thread_id.clone()));
-        THREAD_LAST
-            .lock()
-            .unwrap()
-            .remove(&(engine, job_id, thread_id));
-    }
+/// Set the number of GPU devices.
+pub fn set_gpu_devices(n: i64) {
+    GPU_DEVICES.set(n);
 }

-fn touch_job(engine: &str, job_id: &str) {
-    *JANITOR_INIT; // ensure janitor starts
-    let key = (engine.to_string(), job_id.to_string());
-    JOB_KEYS.lock().unwrap().insert(key.clone());
-    JOB_LAST.lock().unwrap().insert(key, Instant::now());
+/// Set the detected CPU core count.
+pub fn set_effective_cpus(n: i64) {
+    EFFECTIVE_CPUS.set(n);
 }

-fn touch_thread(engine: &str, job_id: &str, thread_id: &str) {
-    *JANITOR_INIT; // ensure janitor starts
-    let key = (
-        engine.to_string(),
-        job_id.to_string(),
-        thread_id.to_string(),
-    );
-    THREAD_KEYS.lock().unwrap().insert(key.clone());
-    THREAD_LAST.lock().unwrap().insert(key, Instant::now());
-}
+// ---------------------------------------------------------------------------
+// HTTP Exporter
+// ---------------------------------------------------------------------------

-/// Explicitly remove all job-scoped metrics for a job (call on job completion)
-pub fn remove_job_metrics(engine: &str, job_id: &str) {
-    let _ = JOB_HASHES_TOTAL.remove_label_values(&[engine, job_id]);
-    let _ = JOB_HASH_RATE.remove_label_values(&[engine, job_id]);
-    let _ = CANDIDATES_FOUND_TOTAL.remove_label_values(&[engine, job_id]);
-    let _ = JOB_STATUS_GAUGE.remove_label_values(&[engine, job_id, "completed"]);
-    let _ = JOB_STATUS_GAUGE.remove_label_values(&[engine, job_id, "failed"]);
-    let _ = JOB_STATUS_GAUGE.remove_label_values(&[engine, job_id, "cancelled"]);
-    // Remove per-job origin gauge variants (known origins)
-    let _ = JOB_FOUND_ORIGIN.remove_label_values(&[engine, job_id, "cpu"]);
-    let _ = JOB_FOUND_ORIGIN.remove_label_values(&[engine, job_id, "gpu-g1"]);
-    let _ = JOB_FOUND_ORIGIN.remove_label_values(&[engine, job_id, "gpu-g2"]);
-    JOB_KEYS
-        .lock()
-        .unwrap()
-        .remove(&(engine.to_string(), job_id.to_string()));
-    JOB_LAST
-        .lock()
-        .unwrap()
-        .remove(&(engine.to_string(), job_id.to_string()));
-}
-
-/// Explicitly remove all thread-scoped metrics for a job
-pub fn remove_thread_metrics_for_job(engine: &str, job_id: &str) {
-    // Collect matching thread keys
-    let keys: Vec<ThreadKey> = {
-        let set = THREAD_KEYS.lock().unwrap();
-        set.iter()
-            .filter(|(e, j, _)| e == engine && j == job_id)
-            .cloned()
-            .collect()
-    };
-    for (e, j, t) in keys {
-        let _ = THREAD_HASHES_TOTAL.remove_label_values(&[&e, &j, &t]);
-        let _ = THREAD_HASH_RATE.remove_label_values(&[&e, &j, &t]);
-        THREAD_KEYS
-            .lock()
-            .unwrap()
-            .remove(&(e.clone(), j.clone(), t.clone()));
-        THREAD_LAST
-            .lock()
-            .unwrap()
-            .remove(&(e.clone(), j.clone(), t.clone()));
-    }
-}
-
-/// Remove the per-job hash rate series for a finished job.
-pub fn remove_job_hash_rate(engine: &str, job_id: &str) {
-    let _ = JOB_HASH_RATE.remove_label_values(&[engine, job_id]);
-}
-
-/// Remove the per-thread hash rate series for a finished thread.
-pub fn remove_thread_hash_rate(engine: &str, job_id: &str, thread_id: &str) {
-    let _ = THREAD_HASH_RATE.remove_label_values(&[engine, job_id, thread_id]);
-}
-
-// -------------------------------------------------------------------------------------
-// GPU-specific metric helpers
-// -------------------------------------------------------------------------------------
-
-/// Set the total number of GPU devices available for mining
-pub fn set_gpu_device_count(count: i64) {
-    GPU_DEVICE_COUNT.set(count);
-}
-
-/// Set GPU device information (call once per device during initialization)
-pub fn set_gpu_device_info(
-    device_id: &str,
-    name: &str,
-    backend: &str,
-    vendor: &str,
-    device_type: &str,
-) {
-    GPU_DEVICE_INFO
-        .with_label_values(&[device_id, name, backend, vendor, device_type])
-        .set(1.0);
-}
-
-/// Set GPU batch size for a specific device
-pub fn set_gpu_batch_size(device_id: &str, batch_size: f64) {
-    GPU_BATCH_SIZE
-        .with_label_values(&[device_id])
-        .set(batch_size);
-}
-
-/// Set GPU workgroup count for a specific device
-pub fn set_gpu_workgroups(device_id: &str, workgroups: f64) {
-    GPU_WORKGROUPS
-        .with_label_values(&[device_id])
-        .set(workgroups);
-}
-
-// -------------------------------------------------------------------------------------
-// HTTP Exporter (feature: http-exporter)
-// -------------------------------------------------------------------------------------
-
-/// Start the Prometheus HTTP exporter on 0.0.0.0:`port`.
+/// Start the Prometheus HTTP exporter on `0.0.0.0:port`.
 ///
-/// Behavior:
-/// - Spawns the exporter as a background task and returns immediately.
-/// - Serves plaintext metrics at GET /metrics.
-/// - If called multiple times, multiple servers may be created (call once).
+/// Spawns the exporter as a background task and returns immediately.
+/// Serves plaintext metrics at `GET /metrics`.
 #[cfg(feature = "http-exporter")]
 pub async fn start_http_exporter(port: u16) -> Result<()> {
-    // Encoder is created inside the handler to avoid capturing non-Clone state
-    // Use REGISTRY.gather() directly in the handler
-
-    // GET /metrics -> plaintext Prometheus format
    let metrics_route = warp::path("metrics").and(warp::get()).map(|| {
        let encoder = TextEncoder::new();
        let metric_families = REGISTRY.gather();
-        let mut buffer = Vec::with_capacity(16 * 1024);
+        let mut buffer = Vec::with_capacity(4096);
        encoder
            .encode(&metric_families, &mut buffer)
            .unwrap_or_default();
@@ -777,7 +316,7 @@ pub async fn start_http_exporter(port: u16) -> Result<()> {
    Ok(())
 }

-// If the exporter feature is not enabled, expose a no-op to keep call sites simple.
+/// No-op when HTTP exporter feature is disabled.
 #[cfg(not(feature = "http-exporter"))]
 pub async fn start_http_exporter(_port: u16) -> Result<()> {
    log::warn!(
@@ -5,12 +5,10 @@ edition.workspace = true
 publish = false
 description = "CLI binary to run the Quantus External Miner service"

-[features]
-default = []
-gpu = ["miner-service/gpu"]
-
 [dependencies]
 miner-service = { path = "../miner-service" }
+metrics = { path = "../metrics", features = ["http-exporter"] }
+quic-transport = { path = "../quic-transport" }
 clap = { workspace = true, features = ["derive", "env"] }
 tokio = { workspace = true, features = ["full"] }
 env_logger = { workspace = true }
@@ -1,134 +0,0 @@
-use std::env;
-use std::fs;
-use std::io::{self, Read};
-use std::path::{Path, PathBuf};
-
-fn main() {
-    // Always re-run if version env or git state might change
-    println!("cargo:rerun-if-env-changed=CARGO_PKG_VERSION");
-    println!("cargo:rerun-if-env-changed=GITHUB_SHA");
-    println!("cargo:rerun-if-changed=.git/HEAD");
-    println!("cargo:rerun-if-changed=.git/refs");
-    println!("cargo:rerun-if-changed=.git/packed-refs");
-
-    let pkg_version = env::var("CARGO_PKG_VERSION").unwrap_or_else(|_| "0.0.0".to_string());
-    let short_sha = match env::var("GITHUB_SHA") {
-        Ok(s) => s.get(0..8).map(|t| t.to_string()),
-        Err(_) => discover_short_git_sha().ok(),
-    };
-
-    let version = match short_sha {
-        Some(sha) if !sha.is_empty() => format!("{pkg_version}+{sha}"),
-        _ => pkg_version.clone(),
-    };
-
-    // Expose to the crate as an env! binding
-    println!("cargo:rustc-env=MINER_VERSION={version}");
-
-    // Optional extra info for diagnostics
-    if let Ok(sha) = env::var("GITHUB_SHA") {
-        println!("cargo:rustc-env=MINER_BUILD_SHA={}", sha);
-    }
-
-    // Log the computed version for debugging
-    println!("miner-cli: computed MINER_VERSION={version}");
-}
-
-/// Attempts to determine the short git SHA (first 8 hex chars) without requiring the `git` binary.
-/// Handles both normal repositories and worktrees where `.git` is a file pointing to the real gitdir.
-fn discover_short_git_sha() -> io::Result<String> {
-    let git_dir = find_git_dir()?;
-    let head_path = git_dir.join("HEAD");
-    let head = fs::read_to_string(&head_path)?.trim().to_string();
-
-    // HEAD can be either a direct SHA or a ref: refs/heads/branch
-    if let Some(ref_line) = head.strip_prefix("ref: ") {
-        let ref_path = git_dir.join(ref_line.trim());
-        if ref_path.is_file() {
-            let mut contents = String::new();
-            fs::File::open(ref_path)?.read_to_string(&mut contents)?;
-            let sha = contents.trim();
-            return Ok(shorten_sha(sha));
-        }
-        // Not a loose ref; try packed-refs
-        let packed = git_dir.join("packed-refs");
-        if packed.is_file() {
-            let contents = fs::read_to_string(packed)?;
-            for line in contents.lines() {
-                // Skip comments and peeled lines
-                if line.starts_with('#') || line.starts_with('^') || line.trim().is_empty() {
-                    continue;
-                }
-                // Format: <sha> <ref>
-                if let Some((sha, r)) = line.split_once(' ') {
-                    if r.trim() == ref_line.trim() {
-                        return Ok(shorten_sha(sha.trim()));
-                    }
-                }
-            }
-        }
-        Err(io::Error::new(
-            io::ErrorKind::NotFound,
-            format!(
-                "Could not resolve ref '{}' in {:?}",
-                ref_line.trim(),
-                git_dir
-            ),
-        ))
-    } else {
-        // HEAD contained a SHA directly (e.g., detached HEAD)
-        Ok(shorten_sha(head.trim()))
-    }
-}
-
-/// Finds the actual .git directory, handling both:
-/// - Standard repo: a ".git" directory exists, or
-/// - Worktree/submodule: ".git" is a file containing "gitdir: <path>"
-fn find_git_dir() -> io::Result<PathBuf> {
-    let mut dir = env::current_dir()?;
-    loop {
-        let dot_git = dir.join(".git");
-        if dot_git.is_dir() {
-            return Ok(dot_git);
-        } else if dot_git.is_file() {
-            // Read the pointer file
-            let contents = fs::read_to_string(&dot_git)?;
-            if let Some(p) = contents.strip_prefix("gitdir:") {
-                let p = p.trim();
-                let path = Path::new(p);
-                // If relative, it's relative to the location of the .git file's parent
-                let resolved = if path.is_absolute() {
-                    path.to_path_buf()
-                } else {
-                    dir.join(path)
-                };
-                if resolved.is_dir() {
-                    return Ok(resolved);
-                } else {
-                    return Err(io::Error::new(
-                        io::ErrorKind::NotFound,
-                        format!("Resolved gitdir does not exist: {}", resolved.display()),
-                    ));
-                }
-            }
-        }
-        // Ascend until root
-        if !dir.pop() {
-            break;
-        }
-    }
-
-    Err(io::Error::new(
-        io::ErrorKind::NotFound,
-        "Could not locate .git directory",
-    ))
-}
-
-fn shorten_sha(full: &str) -> String {
-    let s = full.trim();
-    if s.len() >= 8 {
-        s[..8].to_string()
-    } else {
-        s.to_string()
-    }
-}
@@ -1,123 +1,124 @@
 use clap::{Parser, Subcommand};
-use engine_cpu::{EngineRange, MinerEngine};
+use engine_cpu::{AtomicBoolCancelCheck, EngineRange, MinerEngine};
 use miner_service::{run, ServiceConfig};
 use primitive_types::U512;
 use rand::RngCore;
+use std::path::PathBuf;
 use std::sync::atomic::AtomicBool;
 use std::sync::Arc;
 use std::thread;
 use std::time::{Duration, Instant};

-#[derive(Subcommand, Debug)]
-#[allow(clippy::large_enum_variant)]
-enum Command {
-    /// Run the mining service (default behavior)
-    Serve {
-        /// Port number to listen on for the miner HTTP API
-        #[arg(short, long, env = "MINER_PORT", default_value_t = 9833)]
-        port: u16,
+// CLI defaults
+const DEFAULT_GPU_BATCH_SIZE: u32 = 1_000_000;
+const DEFAULT_CPU_BATCH_SIZE: u64 = 10_000;

-        /// Number of CPU worker threads to use for mining (None = auto-detect)
+#[derive(Subcommand, Debug)]
+enum Command {
+    /// Run the mining service
+    Serve {
+        /// Address of the node to connect to
+        #[arg(long, env = "MINER_NODE_ADDR", default_value = "127.0.0.1:9833")]
+        node_addr: std::net::SocketAddr,
+
+        /// Shared auth token from the node's `miner-auth-token` file
+        /// (`<base-path>/chains/<chain>/miner-auth-token`). Prefer `--auth-token-file`.
+        #[arg(long, env = "MINER_AUTH_TOKEN", conflicts_with = "auth_token_file")]
+        auth_token: Option<String>,
+
+        /// Path to the node's `miner-auth-token` file (trimmed). Preferred over
+        /// `--auth-token` so the secret is not placed on the command line.
+        #[arg(long, env = "MINER_AUTH_TOKEN_FILE", conflicts_with = "auth_token")]
+        auth_token_file: Option<PathBuf>,
+
+        /// SHA-256 fingerprint of the node's miner TLS certificate (64 hex chars).
+        /// Prefer `--tls-cert-sha256-file` pointing at `miner-tls-cert-sha256`
+        /// (the node also logs this fingerprint).
+        #[arg(
+            long,
+            env = "MINER_TLS_CERT_SHA256",
+            conflicts_with = "tls_cert_sha256_file"
+        )]
+        tls_cert_sha256: Option<String>,
+
+        /// Path to the node's `miner-tls-cert-sha256` file.
+        #[arg(
+            long,
+            env = "MINER_TLS_CERT_SHA256_FILE",
+            conflicts_with = "tls_cert_sha256"
+        )]
+        tls_cert_sha256_file: Option<PathBuf>,
+
+        /// Number of CPU worker threads to use for mining (default: auto-detect)
        #[arg(long = "cpu-workers", env = "MINER_CPU_WORKERS")]
        cpu_workers: Option<usize>,

-        /// Number of GPU worker threads to use for mining (None = auto-detect)
-        #[arg(long = "gpu-workers", env = "MINER_GPU_WORKERS")]
-        gpu_workers: Option<usize>,
+        /// Number of GPU devices to use for mining (default: auto-detect)
+        #[arg(long = "gpu-devices", env = "MINER_GPU_DEVICES")]
+        gpu_devices: Option<usize>,

-        /// Optional Prometheus metrics exporter port; if omitted, metrics are disabled
-        #[arg(long, env = "MINER_METRICS_PORT")]
-        metrics_port: Option<u16>,
+        /// GPU batch size in nonces - controls how often GPU checks for cancellation
+        #[arg(long = "gpu-batch-size", env = "MINER_GPU_BATCH_SIZE", default_value_t = DEFAULT_GPU_BATCH_SIZE, value_parser = clap::value_parser!(u32).range(1..))]
+        gpu_batch_size: u32,

-        /// Enable verbose logging (shows debug info, progress details, etc.)
+        /// CPU batch size in hashes - controls how often CPU checks for cancellation
+        #[arg(long = "cpu-batch-size", env = "MINER_CPU_BATCH_SIZE", default_value_t = DEFAULT_CPU_BATCH_SIZE, value_parser = clap::value_parser!(u64).range(1..))]
+        cpu_batch_size: u64,
+
+        /// Port for Prometheus metrics HTTP endpoint (default: 9900)
+        #[arg(
+            long = "metrics-port",
+            env = "MINER_METRICS_PORT",
+            default_value_t = 9900
+        )]
+        metrics_port: u16,
+
+        /// GPU throttle delay in milliseconds between batches (0 = no throttle)
+        #[arg(
+            long = "gpu-throttle-ms",
+            env = "MINER_GPU_THROTTLE_MS",
+            default_value_t = 0
+        )]
+        gpu_throttle_ms: u64,
+
+        /// Allow integrated GPUs (APUs) even when discrete GPUs are available.
+        /// By default, integrated GPUs are skipped when a discrete GPU is present
+        /// to avoid resource contention and driver instability.
+        #[arg(long = "allow-integrated", env = "MINER_ALLOW_INTEGRATED")]
+        allow_integrated: bool,
+
+        /// Enable verbose logging
        #[arg(short, long, env = "MINER_VERBOSE")]
        verbose: bool,
-
-        /// How often to report mining progress (in milliseconds).
-        /// Smaller values give more frequent updates but slightly reduce performance.
-        #[arg(long = "progress-interval-ms", env = "MINER_PROGRESS_INTERVAL_MS")]
-        progress_interval_ms: Option<u64>,
-
-        /// Size of work chunks to process before reporting progress (number of hashes).
-        /// If omitted, uses engine-specific defaults (200K for CPU, 100M for GPU).
-        #[arg(long = "chunk-size", env = "MINER_CHUNK_SIZE")]
-        chunk_size: Option<u64>,
-
-        /// For cpu-chain-manipulator: start throttle index at this many solved blocks
-        /// to "pick up where we left off" after restarts.
-        #[arg(long = "manip-solved-blocks", env = "MINER_MANIP_SOLVED_BLOCKS")]
-        manip_solved_blocks: Option<u64>,
-
-        /// For cpu-chain-manipulator: base sleep per batch in nanoseconds (default 500_000 ns)
-        #[arg(long = "manip-base-delay-ns", env = "MINER_MANIP_BASE_DELAY_NS")]
-        manip_base_delay_ns: Option<u64>,
-
-        /// For cpu-chain-manipulator: number of nonce attempts between sleeps (default 10_000)
-        #[arg(long = "manip-step-batch", env = "MINER_MANIP_STEP_BATCH")]
-        manip_step_batch: Option<u64>,
-
-        /// For cpu-chain-manipulator: optional cap on solved-blocks throttle index
-        #[arg(long = "manip-throttle-cap", env = "MINER_MANIP_THROTTLE_CAP")]
-        manip_throttle_cap: Option<u64>,
-
-        /// Telemetry endpoints (repeat --telemetry-endpoint or comma-separated)
-        #[arg(long = "telemetry-endpoint", env = "MINER_TELEMETRY_ENDPOINTS", value_delimiter = ',', num_args = 0.., value_name = "URL")]
-        telemetry_endpoints: Option<Vec<String>>,
-
-        /// Enable or disable telemetry explicitly
-        #[arg(long = "telemetry-enabled", env = "MINER_TELEMETRY_ENABLED")]
-        telemetry_enabled: Option<bool>,
-
-        /// Telemetry verbosity level (0..=4 typical)
-        #[arg(long = "telemetry-verbosity", env = "MINER_TELEMETRY_VERBOSITY")]
-        telemetry_verbosity: Option<u8>,
-
-        /// Interval seconds for system.interval messages
-        #[arg(
-            long = "telemetry-interval-secs",
-            env = "MINER_TELEMETRY_INTERVAL_SECS"
-        )]
-        telemetry_interval_secs: Option<u64>,
-
-        /// Default association: chain name
-        #[arg(long = "telemetry-chain", env = "MINER_TELEMETRY_CHAIN")]
-        telemetry_chain: Option<String>,
-
-        /// Default association: genesis hash (hex)
-        #[arg(long = "telemetry-genesis", env = "MINER_TELEMETRY_GENESIS")]
-        telemetry_genesis: Option<String>,
-
-        /// Default association: node telemetry id
-        #[arg(long = "telemetry-node-id", env = "MINER_TELEMETRY_NODE_ID")]
-        telemetry_node_id: Option<String>,
-
-        /// Default association: node libp2p peer id
-        #[arg(long = "telemetry-node-peer-id", env = "MINER_TELEMETRY_NODE_PEER_ID")]
-        telemetry_node_peer_id: Option<String>,
-
-        /// Default association: node name
-        #[arg(long = "telemetry-node-name", env = "MINER_TELEMETRY_NODE_NAME")]
-        telemetry_node_name: Option<String>,
-
-        /// Default association: node version
-        #[arg(long = "telemetry-node-version", env = "MINER_TELEMETRY_NODE_VERSION")]
-        telemetry_node_version: Option<String>,
    },
+
    /// Run a quick benchmark of the mining engines
    Benchmark {
        /// Number of CPU workers to use for benchmark
        #[arg(long = "cpu-workers", env = "MINER_CPU_WORKERS")]
        cpu_workers: Option<usize>,

-        /// Number of GPU workers to use for benchmark
-        #[arg(long = "gpu-workers", env = "MINER_GPU_WORKERS")]
-        gpu_workers: Option<usize>,
+        /// Number of GPU devices to use for benchmark
+        #[arg(long = "gpu-devices", env = "MINER_GPU_DEVICES")]
+        gpu_devices: Option<usize>,
+
+        /// GPU batch size in nonces - controls how often GPU checks for cancellation
+        #[arg(long = "gpu-batch-size", env = "MINER_GPU_BATCH_SIZE", default_value_t = DEFAULT_GPU_BATCH_SIZE, value_parser = clap::value_parser!(u32).range(1..))]
+        gpu_batch_size: u32,
+
+        /// CPU batch size in hashes - controls how often CPU checks for cancellation
+        #[arg(long = "cpu-batch-size", env = "MINER_CPU_BATCH_SIZE", default_value_t = DEFAULT_CPU_BATCH_SIZE, value_parser = clap::value_parser!(u64).range(1..))]
+        cpu_batch_size: u64,

        /// Benchmark duration in seconds (default: 10)
        #[arg(short, long, default_value_t = 10)]
        duration: u64,

-        /// Enable verbose logging during benchmark
+        /// Allow integrated GPUs (APUs) even when discrete GPUs are available
+        #[arg(long = "allow-integrated", env = "MINER_ALLOW_INTEGRATED")]
+        allow_integrated: bool,
+
+        /// Enable verbose logging
        #[arg(short, long, env = "MINER_VERBOSE")]
        verbose: bool,
    },
@@ -125,336 +126,320 @@ enum Command {

 /// Quantus External Miner CLI
 #[derive(Parser, Debug)]
-#[command(author, version = option_env!("MINER_VERSION").unwrap_or(env!("CARGO_PKG_VERSION")), about, long_about = None)]
+#[command(author, version, about, long_about = None)]
 struct Args {
    #[command(subcommand)]
    command: Option<Command>,
 }

-#[allow(clippy::enum_variant_names)]
 #[tokio::main]
 async fn main() {
    let args = Args::parse();

-    match args.command.unwrap_or(Command::Serve {
-        port: 9833,
-        cpu_workers: None,
-        gpu_workers: None,
-        metrics_port: None,
-        verbose: false,
-        progress_interval_ms: None,
-        chunk_size: None,
-        manip_solved_blocks: None,
-        manip_base_delay_ns: None,
-        manip_step_batch: None,
-        manip_throttle_cap: None,
-        telemetry_endpoints: None,
-        telemetry_enabled: None,
-        telemetry_verbosity: None,
-        telemetry_interval_secs: None,
-        telemetry_chain: None,
-        telemetry_genesis: None,
-        telemetry_node_id: None,
-        telemetry_node_peer_id: None,
-        telemetry_node_name: None,
-        telemetry_node_version: None,
-    }) {
+    let Some(command) = args.command else {
+        eprintln!("Error: No command provided. Use 'serve' to start mining (defaults to local node at 127.0.0.1:9833).");
+        eprintln!(
+            "Example: quantus-miner serve --node-addr 127.0.0.1:9833 \
+             --auth-token-file /path/to/miner-auth-token \
+             --tls-cert-sha256-file /path/to/miner-tls-cert-sha256"
+        );
+        std::process::exit(1);
+    };
+
+    match command {
        Command::Serve {
-            port,
+            node_addr,
+            auth_token,
+            auth_token_file,
+            tls_cert_sha256,
+            tls_cert_sha256_file,
            cpu_workers,
-            gpu_workers,
+            gpu_devices,
+            gpu_batch_size,
+            cpu_batch_size,
+            gpu_throttle_ms,
            metrics_port,
+            allow_integrated,
            verbose,
-            progress_interval_ms,
-            chunk_size,
-            manip_solved_blocks,
-            manip_base_delay_ns,
-            manip_step_batch,
-            manip_throttle_cap,
-            telemetry_endpoints,
-            telemetry_enabled,
-            telemetry_verbosity,
-            telemetry_interval_secs,
-            telemetry_chain,
-            telemetry_genesis,
-            telemetry_node_id,
-            telemetry_node_peer_id,
-            telemetry_node_name,
-            telemetry_node_version,
        } => {
-            run_serve_command(
-                port,
+            init_logger(verbose);
+
+            let auth_token = match resolve_auth_token(auth_token, auth_token_file) {
+                Ok(token) => token,
+                Err(e) => {
+                    eprintln!("Error: {e}");
+                    std::process::exit(1);
+                }
+            };
+            let tls_cert_sha256 =
+                match resolve_tls_cert_sha256(tls_cert_sha256, tls_cert_sha256_file) {
+                    Ok(fp) => fp,
+                    Err(e) => {
+                        eprintln!("Error: {e}");
+                        std::process::exit(1);
+                    }
+                };
+            // Fail closed on permanent misconfig before metrics/workers start.
+            if let Err(e) = quic_transport::validate_auth_config(&auth_token, &tls_cert_sha256) {
+                eprintln!("Error: {e}");
+                std::process::exit(1);
+            }
+
+            log::info!("Starting external miner service...");
+
+            // Start metrics HTTP server
+            if let Err(e) = metrics::start_http_exporter(metrics_port).await {
+                log::error!("Failed to start metrics exporter: {e:?}");
+                std::process::exit(1);
+            }
+            log::info!(
+                "Metrics available at http://0.0.0.0:{}/metrics",
+                metrics_port
+            );
+
+            let config = ServiceConfig {
+                node_addr,
+                auth_token,
+                tls_cert_sha256,
                cpu_workers,
-                gpu_workers,
-                metrics_port,
-                verbose,
-                progress_interval_ms,
-                chunk_size,
-                manip_solved_blocks,
-                manip_base_delay_ns,
-                manip_step_batch,
-                manip_throttle_cap,
-                telemetry_endpoints,
-                telemetry_enabled,
-                telemetry_verbosity,
-                telemetry_interval_secs,
-                telemetry_chain,
-                telemetry_genesis,
-                telemetry_node_id,
-                telemetry_node_peer_id,
-                telemetry_node_name,
-                telemetry_node_version,
+                gpu_devices,
+                gpu_batch_size,
+                cpu_batch_size,
+                gpu_throttle_ms,
+                allow_integrated,
+            };
+
+            if let Err(e) = run(config).await {
+                log::error!("Miner service terminated with error: {e:?}");
+                std::process::exit(1);
+            }
+        }
+
+        Command::Benchmark {
+            cpu_workers,
+            gpu_devices,
+            gpu_batch_size,
+            cpu_batch_size,
+            duration,
+            allow_integrated,
+            verbose,
+        } => {
+            init_logger(verbose);
+            run_benchmark(
+                cpu_workers,
+                gpu_devices,
+                gpu_batch_size,
+                cpu_batch_size,
+                duration,
+                allow_integrated,
            )
            .await;
        }
-        Command::Benchmark {
-            cpu_workers,
-            gpu_workers,
-            duration,
-            verbose,
-        } => {
-            run_benchmark_command(cpu_workers, gpu_workers, duration, verbose).await;
-        }
    }
 }

-#[allow(clippy::too_many_arguments)]
-async fn run_serve_command(
-    port: u16,
-    cpu_workers: Option<usize>,
-    gpu_workers: Option<usize>,
-    metrics_port: Option<u16>,
-    verbose: bool,
-    progress_interval_ms: Option<u64>,
-    chunk_size: Option<u64>,
-    manip_solved_blocks: Option<u64>,
-    manip_base_delay_ns: Option<u64>,
-    manip_step_batch: Option<u64>,
-    manip_throttle_cap: Option<u64>,
-    telemetry_endpoints: Option<Vec<String>>,
-    telemetry_enabled: Option<bool>,
-    telemetry_verbosity: Option<u8>,
-    telemetry_interval_secs: Option<u64>,
-    telemetry_chain: Option<String>,
-    telemetry_genesis: Option<String>,
-    telemetry_node_id: Option<String>,
-    telemetry_node_peer_id: Option<String>,
-    telemetry_node_name: Option<String>,
-    telemetry_node_version: Option<String>,
-) {
-    // Initialize logger early to capture startup messages.
-    // If RUST_LOG is not set, default to appropriate level based on verbose flag
+fn resolve_auth_token(
+    auth_token: Option<String>,
+    auth_token_file: Option<PathBuf>,
+) -> Result<String, String> {
+    resolve_required_secret(
+        auth_token,
+        auth_token_file,
+        "--auth-token",
+        "--auth-token-file",
+        "miner auth token required: pass --auth-token-file <PATH> to the node's \
+         miner-auth-token file (or --auth-token <TOKEN>)",
+    )
+}
+
+fn resolve_tls_cert_sha256(value: Option<String>, file: Option<PathBuf>) -> Result<String, String> {
+    resolve_required_secret(
+        value,
+        file,
+        "--tls-cert-sha256",
+        "--tls-cert-sha256-file",
+        "TLS cert fingerprint required: pass --tls-cert-sha256-file <PATH> to the node's \
+         miner-tls-cert-sha256 file (or --tls-cert-sha256 <HEX>; also printed in node logs)",
+    )
+}
+
+fn resolve_required_secret(
+    value: Option<String>,
+    file: Option<PathBuf>,
+    value_flag: &str,
+    file_flag: &str,
+    missing_msg: &str,
+) -> Result<String, String> {
+    if let Some(value) = value {
+        let value = value.trim().to_string();
+        if value.is_empty() {
+            return Err(format!("{value_flag} is empty"));
+        }
+        return Ok(value);
+    }
+    if let Some(path) = file {
+        let contents = std::fs::read_to_string(&path)
+            .map_err(|e| format!("failed to read {file_flag} {}: {}", path.display(), e))?;
+        let value = contents.trim().to_string();
+        if value.is_empty() {
+            return Err(format!("{file_flag} {} is empty", path.display()));
+        }
+        return Ok(value);
+    }
+    Err(missing_msg.into())
+}
+
+fn init_logger(verbose: bool) {
    if std::env::var("RUST_LOG").is_err() {
+        // Filter out noisy wgpu/naga shader compilation logs
        let log_level = if verbose {
-            "debug,miner=debug,gpu_engine=debug,engine_cpu=debug"
+            "debug,miner=debug,gpu_engine=debug,engine_cpu=debug,wgpu=warn,wgpu_core=warn,wgpu_hal=warn,naga=warn"
        } else {
-            "info,miner=info,gpu_engine=info"
+            "info,miner=info,gpu_engine=info,wgpu=error,wgpu_core=error,wgpu_hal=error,naga=error"
        };
        std::env::set_var("RUST_LOG", log_level);
    }
    env_logger::init();
-
-    // Telemetry CLI passthrough to env for miner-service bootstrap
-    if let Some(eps) = telemetry_endpoints.as_ref() {
-        if !eps.is_empty() {
-            std::env::set_var("MINER_TELEMETRY_ENDPOINTS", eps.join(","));
-        }
-    }
-    if let Some(v) = telemetry_enabled {
-        std::env::set_var("MINER_TELEMETRY_ENABLED", if v { "1" } else { "0" });
-    }
-    if let Some(v) = telemetry_verbosity {
-        std::env::set_var("MINER_TELEMETRY_VERBOSITY", v.to_string());
-    }
-    if let Some(v) = telemetry_interval_secs {
-        std::env::set_var("MINER_TELEMETRY_INTERVAL_SECS", v.to_string());
-    }
-    if let Some(v) = telemetry_chain.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_CHAIN", v);
-    }
-    if let Some(v) = telemetry_genesis.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_GENESIS", v);
-    }
-    if let Some(v) = telemetry_node_id.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_NODE_ID", v);
-    }
-    if let Some(v) = telemetry_node_peer_id.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_NODE_PEER_ID", v);
-    }
-    if let Some(v) = telemetry_node_name.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_NODE_NAME", v);
-    }
-    if let Some(v) = telemetry_node_version.as_ref() {
-        std::env::set_var("MINER_TELEMETRY_NODE_VERSION", v);
-    }
-
-    // Log effective configuration (concise; see ServiceConfig Display)
-    log::info!("Starting external miner service...");
-
-    let config = ServiceConfig {
-        port,
-        cpu_workers,
-        gpu_workers,
-        metrics_port,
-        progress_interval_ms,
-        chunk_size,
-        manip_solved_blocks,
-        manip_base_delay_ns,
-        manip_step_batch,
-        manip_throttle_cap,
-    };
-    log::info!("Effective config: {config}");
-
-    if let Err(e) = run(config).await {
-        log::error!("Miner service terminated with error: {e:?}");
-        std::process::exit(1);
-    }
 }

-async fn run_benchmark_command(
+async fn run_benchmark(
    cpu_workers: Option<usize>,
-    gpu_workers: Option<usize>,
+    gpu_devices: Option<usize>,
+    gpu_batch_size: u32,
+    cpu_batch_size: u64,
    duration: u64,
-    verbose: bool,
+    allow_integrated: bool,
 ) {
-    // Initialize logger early to capture startup messages.
-    if std::env::var("RUST_LOG").is_err() {
-        let log_level = if verbose {
-            "debug,miner=debug,gpu_engine=debug,engine_cpu=debug"
-        } else {
-            "info,miner=info,gpu_engine=warn,engine_cpu=info"
-        };
-        std::env::set_var("RUST_LOG", log_level);
-    }
-    env_logger::init();
-
    let effective_cpu_workers = cpu_workers.unwrap_or_else(num_cpus::get);
-    let effective_gpu_workers = gpu_workers.unwrap_or(0);
-    let total_workers = effective_cpu_workers + effective_gpu_workers;
+
+    // Initialize GPU engine (no throttle for benchmark)
+    let (gpu_engine, effective_gpu_devices) = match miner_service::resolve_gpu_configuration(
+        gpu_devices,
+        gpu_batch_size,
+        0,
+        allow_integrated,
+    ) {
+        Ok((engine, count)) => (engine, count),
+        Err(e) => {
+            eprintln!("❌ ERROR: {}", e);
+            std::process::exit(1);
+        }
+    };
+
+    let total_workers = effective_cpu_workers + effective_gpu_devices;

    println!("🚀 Quantus Miner Benchmark");
    println!("==========================");
    println!(
-        "Configuration: {} CPU workers, {} GPU workers",
-        effective_cpu_workers, effective_gpu_workers
+        "CPU Workers: {} (Available: {})",
+        effective_cpu_workers,
+        num_cpus::get()
    );
-    println!("Duration: {} seconds", duration);
-    println!("Total Workers: {}", total_workers);
-    println!("Available CPUs: {}", num_cpus::get());
-
-    // Create engines based on configuration
-    let cpu_engine = if effective_cpu_workers > 0 {
-        Some(Arc::new(engine_cpu::FastCpuEngine::new()) as Arc<dyn MinerEngine>)
-    } else {
-        None
-    };
-
-    let gpu_engine = if effective_gpu_workers > 0 {
-        #[cfg(feature = "gpu")]
-        {
-            Some(Arc::new(engine_gpu::GpuEngine::new()) as Arc<dyn MinerEngine>)
-        }
-        #[cfg(not(feature = "gpu"))]
-        {
-            eprintln!("Error: GPU workers requested but this binary was built without GPU support");
-            std::process::exit(1);
-        }
-    } else {
-        None
-    };
-
-    // For benchmark, we'll use CPU engine as primary for simplicity
-    let engine = cpu_engine
-        .or(gpu_engine)
-        .expect("At least one engine must be available");
+    println!("GPU Devices: {}", effective_gpu_devices);
+    if effective_cpu_workers > 0 {
+        println!("CPU batch size: {cpu_batch_size} hashes");
+    }
+    if effective_gpu_devices > 0 {
+        println!("GPU batch size: {gpu_batch_size} nonces");
+    }
+    println!("Duration: {duration} seconds");
+    println!();

    if total_workers == 0 {
        eprintln!("Error: No workers specified");
        std::process::exit(1);
    }
-    println!("Workers: {}", total_workers);
-    println!();

-    // Run benchmark
+    // Create CPU engine
+    let cpu_engine: Option<Arc<dyn MinerEngine>> = if effective_cpu_workers > 0 {
+        Some(Arc::new(engine_cpu::FastCpuEngine::new(cpu_batch_size)))
+    } else {
+        None
+    };
+
    let cancel_flag = Arc::new(AtomicBool::new(false));
    let benchmark_start = Instant::now();

-    // Create a large range that should take the full duration
-    let benchmark_range = EngineRange {
-        start: U512::from(0u64),
-        end: U512::from(100_000_000u64), // 100M nonces - should be plenty
-    };
-
+    // Random header hash for benchmark
    let mut header = [0u8; 32];
-    rand::thread_rng().fill_bytes(&mut header);
+    rand::rng().fill_bytes(&mut header);
    let difficulty = U512::MAX; // High difficulty - no solutions expected
-    let ctx = engine.prepare_context(header, difficulty);
+
+    let ref_engine = cpu_engine.as_ref().or(gpu_engine.as_ref()).unwrap();
+    let ctx = ref_engine.prepare_context(header, difficulty);

    println!("⛏️  Starting benchmark...");

    // Spawn worker threads
    let mut handles = Vec::new();
-    let total_hashes_arc = Arc::new(std::sync::Mutex::new(0u64));
+    let total_hashes = Arc::new(std::sync::Mutex::new(0u64));

-    // Use larger ranges for GPU (1M) vs CPU (10K)
-    let nonces_per_worker = if effective_gpu_workers > 0 {
-        1_000_000u64 // 1M nonces per GPU worker
-    } else {
-        10_000u64 // 10K nonces per CPU worker
-    };
+    // Floor range widths at the old constants: engines still batch at the flag's
+    // size internally, but tiny flags don't turn per-call harness overhead into
+    // the measured quantity.
+    let cpu_chunk = cpu_batch_size.max(10_000);
+    let gpu_chunk = (gpu_batch_size as u64).max(1_000_000);

    for worker_id in 0..total_workers {
-        let engine = engine.clone();
+        let (engine, nonces_per_batch) = if worker_id < effective_cpu_workers {
+            (cpu_engine.as_ref().unwrap().clone(), cpu_chunk)
+        } else {
+            (gpu_engine.as_ref().unwrap().clone(), gpu_chunk)
+        };
+
        let ctx = ctx.clone();
-        let cancel_flag = cancel_flag.clone();
-        let total_hashes = total_hashes_arc.clone();
+        let cancel = cancel_flag.clone();
+        let hashes = total_hashes.clone();
+        let start = benchmark_start;

        let handle = thread::spawn(move || {
-            let worker_range = EngineRange {
-                start: benchmark_range.start.saturating_add(
-                    U512::from(worker_id as u64).saturating_mul(U512::from(nonces_per_worker)),
-                ),
-                end: benchmark_range
-                    .start
-                    .saturating_add(
-                        U512::from((worker_id + 1) as u64)
-                            .saturating_mul(U512::from(nonces_per_worker)),
-                    )
-                    .saturating_sub(U512::from(1u64)),
-            };
-
-            let mut worker_hashes = 0u64;
+            let stride = U512::from(1_000_000_000_000u64);
+            let mut current = U512::from(worker_id as u64).saturating_mul(stride);
+            let step = U512::from(nonces_per_batch);

            loop {
-                if cancel_flag.load(std::sync::atomic::Ordering::Relaxed) {
+                if cancel.load(std::sync::atomic::Ordering::Relaxed) {
                    break;
                }

-                let result = engine.search_range(&ctx, worker_range.clone(), &cancel_flag);
+                let worker_range = EngineRange {
+                    start: current,
+                    end: current
+                        .saturating_add(step)
+                        .saturating_sub(U512::from(1u64)),
+                };
+
+                let cancel_check = AtomicBoolCancelCheck(&cancel);
+                let result = engine.search_range(&ctx, worker_range, &cancel_check);

                match result {
                    engine_cpu::EngineStatus::Found { hash_count, .. }
                    | engine_cpu::EngineStatus::Exhausted { hash_count }
-                    | engine_cpu::EngineStatus::Cancelled { hash_count } => {
-                        worker_hashes += hash_count;
-                        *total_hashes.lock().unwrap() += hash_count;
+                    | engine_cpu::EngineStatus::Cancelled { hash_count }
+                    | engine_cpu::EngineStatus::DeviceLost { hash_count } => {
+                        *hashes.lock().unwrap() += hash_count;
                    }
                    engine_cpu::EngineStatus::Running { .. } => {}
                }

-                // Check if we've exceeded the time limit
-                if benchmark_start.elapsed() >= Duration::from_secs(duration) {
+                if matches!(result, engine_cpu::EngineStatus::DeviceLost { .. }) {
+                    break;
+                }
+
+                current = current.saturating_add(step);
+
+                if start.elapsed() >= Duration::from_secs(duration) {
                    break;
                }
            }

-            worker_hashes
+            engine_gpu::GpuEngine::clear_worker_resources();
        });

        handles.push(handle);
    }

-    // Wait for duration or interrupt
+    // Progress updates
    let mut last_update = Instant::now();

    loop {
@@ -465,70 +450,47 @@ async fn run_benchmark_command(
            break;
        }

-        // Update progress every second
        if last_update.elapsed() >= Duration::from_secs(1) {
-            let current_hashes = *total_hashes_arc.lock().unwrap();
+            let current = *total_hashes.lock().unwrap();
            let elapsed = benchmark_start.elapsed().as_secs_f64();
-
-            if current_hashes > 0 {
-                let hash_rate = current_hashes as f64 / elapsed;
-                let hash_rate_str = if hash_rate >= 1_000_000.0 {
-                    format!("{:.1}M", hash_rate / 1_000_000.0)
-                } else if hash_rate >= 1_000.0 {
-                    format!("{:.1}K", hash_rate / 1_000.0)
-                } else {
-                    format!("{:.0}", hash_rate)
-                };
-                println!("⏱️  {:.1}s - {} H/s", elapsed, hash_rate_str);
-            } else if effective_gpu_workers > 0 {
-                println!("⏱️  {:.1}s - processing...", elapsed);
-            } else {
-                println!("⏱️  {:.1}s - starting...", elapsed);
+            if current > 0 {
+                let rate = current as f64 / elapsed;
+                println!("⏱️  {:.1}s - {} H/s", elapsed, format_hash_rate(rate));
            }
-
            last_update = Instant::now();
        }
    }

-    // Wait for all threads to finish
+    // Wait for threads
    for handle in handles {
        let _ = handle.join();
    }

    let total_elapsed = benchmark_start.elapsed();
-    let final_hashes = *total_hashes_arc.lock().unwrap();
-    let avg_hash_rate = final_hashes as f64 / total_elapsed.as_secs_f64();
+    let final_hashes = *total_hashes.lock().unwrap();
+    let avg_rate = final_hashes as f64 / total_elapsed.as_secs_f64();

    println!();
    println!("📊 Benchmark Results");
    println!("===================");
-    println!("Total time: {:.2} seconds", total_elapsed.as_secs_f64());
+    println!("Total time: {:.2}s", total_elapsed.as_secs_f64());
    println!("Total hashes: {}", final_hashes);
-
-    let hash_rate_str = if avg_hash_rate >= 1_000_000.0 {
-        format!("{:.2}M H/s", avg_hash_rate / 1_000_000.0)
-    } else if avg_hash_rate >= 1_000.0 {
-        format!("{:.2}K H/s", avg_hash_rate / 1_000.0)
-    } else {
-        format!("{:.0} H/s", avg_hash_rate)
-    };
-
-    println!("Average hash rate: {}", hash_rate_str);
+    println!("Average rate: {} H/s", format_hash_rate(avg_rate));

    if total_workers > 1 {
-        let per_worker_rate = avg_hash_rate / total_workers as f64;
-        let per_worker_str = if per_worker_rate >= 1_000_000.0 {
-            format!("{:.2}M H/s", per_worker_rate / 1_000_000.0)
-        } else if per_worker_rate >= 1_000.0 {
-            format!("{:.2}K H/s", per_worker_rate / 1_000.0)
-        } else {
-            format!("{:.0} H/s", per_worker_rate)
-        };
-        println!(
-            "Per-worker rate: {} (across {} workers)",
-            per_worker_str, total_workers
-        );
+        let per_worker = avg_rate / total_workers as f64;
+        println!("Per-worker: {} H/s", format_hash_rate(per_worker));
    }

-    println!("✅ Benchmark completed successfully!");
+    println!("✅ Benchmark completed!");
+}
+
+fn format_hash_rate(rate: f64) -> String {
+    if rate >= 1_000_000.0 {
+        format!("{:.2}M", rate / 1_000_000.0)
+    } else if rate >= 1_000.0 {
+        format!("{:.2}K", rate / 1_000.0)
+    } else {
+        format!("{:.0}", rate)
+    }
 }
@@ -3,43 +3,32 @@ name = "miner-service"
 version.workspace = true
 edition.workspace = true
 publish = false
-description = "Service layer: HTTP API, job orchestration, and engine abstraction for the Quantus External Miner"
+description = "QUIC-based mining service for the Quantus External Miner"

 [features]
-default = ["cpu", "metrics"]
-# Enable CPU engine by default.
+default = ["cpu"]
 cpu = ["engine-cpu"]
-# Enable GPU engine.
-gpu = ["engine-gpu"]
-# Optional metrics/observability (Prometheus endpoint).
-metrics = [
-	"dep:metrics",
-	"metrics/http-exporter",
-]

 [dependencies]
 # Workspace-shared deps
-warp = { workspace = true }
 tokio = { workspace = true, features = ["full"] }
-serde = { workspace = true, features = ["derive"] }
-serde_json = { workspace = true }
 hex = { workspace = true }
 primitive-types = { workspace = true }
 log = { workspace = true }
 crossbeam-channel = { workspace = true }
 num_cpus = { workspace = true }
-thiserror = { workspace = true }
 anyhow = { workspace = true }
-reqwest = { version = "0.11", features = ["json"] }
-quinn = "0.10"
-rustls = { version = "0.21", default-features = false, features = ["dangerous_configuration", "quic"] }

-# Protocol types from the node repo (keep API compatibility)
+# QUIC transport
+quinn = "0.10"
+quic-transport = { path = "../quic-transport" }
+getrandom = "0.2"
+
+# Protocol types from the node repo
 quantus-miner-api = { workspace = true }

 # Local crates
 pow-core = { path = "../pow-core" }
 engine-cpu = { path = "../engine-cpu", optional = true }
-engine-gpu = { path = "../engine-gpu", optional = true }
-metrics = { path = "../metrics", optional = true }
-miner-telemetry = { path = "../miner-telemetry" }
+engine-gpu = { path = "../engine-gpu" }
+metrics = { path = "../metrics" }
@@ -0,0 +1,337 @@
+//! QUIC client for connecting to blockchain nodes.
+//!
+//! This module provides a QUIC client that connects to a blockchain node
+//! and handles bidirectional streaming for receiving mining jobs and
+//! sending results.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use engine_cpu::MinerEngine;
+use primitive_types::U512;
+
+use quantus_miner_api::{
+    read_message, write_message, ApiResponseStatus, MinerMessage, MiningResult,
+};
+
+use crate::{EngineType, WorkerPool};
+use pow_core::format_hashrate;
+
+/// Connect to a node and start mining.
+///
+/// This function connects to the node, receives mining jobs, and sends results.
+/// It automatically reconnects if the connection is lost.
+///
+/// Uses a persistent worker pool to avoid thread creation overhead between jobs.
+pub async fn connect_and_mine(
+    node_addr: SocketAddr,
+    auth_token: &str,
+    tls_cert_sha256: &str,
+    cpu_engine: Option<Arc<dyn MinerEngine>>,
+    gpu_engine: Option<Arc<dyn MinerEngine>>,
+    cpu_workers: usize,
+    gpu_devices: usize,
+) -> anyhow::Result<()> {
+    // Create persistent worker pool once - it lives for the entire miner lifetime
+    let worker_pool = WorkerPool::new(cpu_engine, gpu_engine, cpu_workers, gpu_devices);
+
+    let mut reconnect_delay = Duration::from_secs(1);
+    const MAX_RECONNECT_DELAY: Duration = Duration::from_secs(30);
+
+    loop {
+        log::info!("⛏️ Connecting to node at {}...", node_addr);
+
+        match establish_connection(node_addr, auth_token, tls_cert_sha256).await {
+            Ok((connection, send, recv)) => {
+                log::info!("⛏️ Connected to node at {}", node_addr);
+
+                let mut authenticated = false;
+                if let Err(e) =
+                    handle_connection(connection, send, recv, &worker_pool, &mut authenticated)
+                        .await
+                {
+                    // Cancel any running job when connection drops
+                    worker_pool.cancel();
+                    if e.downcast_ref::<quic_transport::PermanentConnectError>()
+                        .is_some()
+                    {
+                        log::error!("⛏️ Permanent connection error (not retrying): {e}");
+                        return Err(e);
+                    }
+                    log::info!("⛏️ Connection lost: {}", e);
+                }
+                // Only clear backoff after the node accepted auth (first NewJob).
+                // connect() already treats explicit "auth failed" as permanent;
+                // this covers any other post-Ready close that looked like success.
+                if authenticated {
+                    reconnect_delay = Duration::from_secs(1);
+                }
+            }
+            Err(e) => {
+                if e.downcast_ref::<quic_transport::PermanentConnectError>()
+                    .is_some()
+                {
+                    log::error!("⛏️ Permanent connection error (not retrying): {e}");
+                    return Err(e);
+                }
+                log::warn!("⛏️ Failed to connect to node: {}", e);
+            }
+        }
+
+        log::info!("⛏️ Reconnecting in {:?}...", reconnect_delay);
+        tokio::time::sleep(reconnect_delay).await;
+        reconnect_delay = (reconnect_delay * 2).min(MAX_RECONNECT_DELAY);
+    }
+}
+
+/// Establish a QUIC connection to the node (shared transport crate).
+async fn establish_connection(
+    addr: SocketAddr,
+    auth_token: &str,
+    tls_cert_sha256: &str,
+) -> anyhow::Result<(quinn::Connection, quinn::SendStream, quinn::RecvStream)> {
+    let result = quic_transport::connect(addr, auth_token, tls_cert_sha256).await?;
+    log::info!(
+        "⛏️ QUIC connection and bidirectional stream established to {}",
+        addr
+    );
+    Ok(result)
+}
+
+/// Helper to send a message while monitoring connection health.
+async fn send_message_checked(
+    connection: &quinn::Connection,
+    send: &mut quinn::SendStream,
+    msg: &MinerMessage,
+) -> anyhow::Result<()> {
+    tokio::select! {
+        biased;
+        reason = connection.closed() => {
+            Err(anyhow::anyhow!("Connection closed: {}", reason))
+        }
+        result = write_message(send, msg) => {
+            result.map_err(|e| anyhow::anyhow!("Failed to send message: {}", e))
+        }
+    }
+}
+
+/// Handle an established connection, receiving jobs and sending results.
+///
+/// Sets `authenticated` when the first `NewJob` arrives (proof the node
+/// accepted our Ready token).
+async fn handle_connection(
+    connection: quinn::Connection,
+    mut send: quinn::SendStream,
+    mut recv: quinn::RecvStream,
+    worker_pool: &WorkerPool,
+    authenticated: &mut bool,
+) -> anyhow::Result<()> {
+    use crossbeam_channel::RecvTimeoutError;
+
+    // Set static metrics once per connection
+    metrics::set_effective_cpus(num_cpus::get() as i64);
+    metrics::set_workers(worker_pool.worker_count() as i64);
+    metrics::set_cpu_workers(worker_pool.cpu_worker_count() as i64);
+    metrics::set_gpu_devices(worker_pool.gpu_worker_count() as i64);
+    metrics::reset_hash_tracker();
+
+    // Current job state
+    // - node_job_id: The string ID from the node (e.g., "27") - used in protocol messages
+    // - internal_job_id: Our internal numeric ID from WorkerPool - used to detect stale results
+    let mut node_job_id: Option<String> = None;
+    let mut internal_job_id: u64 = 0;
+    let mut job_start_time: Option<Instant> = None;
+    let mut cpu_hashes: u64 = 0;
+    let mut gpu_hashes: u64 = 0;
+    let mut result_sent_for_current_job = false;
+
+    log::info!("⛏️ Waiting for mining jobs from node...");
+
+    loop {
+        // Poll for worker results (non-blocking via spawn_blocking)
+        let poll_result = if node_job_id.is_some() && !result_sent_for_current_job {
+            let rx = worker_pool.result_receiver().clone();
+            tokio::task::spawn_blocking(move || rx.recv_timeout(Duration::from_millis(10)))
+                .await
+                .unwrap_or(Err(RecvTimeoutError::Disconnected))
+        } else {
+            Err(RecvTimeoutError::Timeout)
+        };
+
+        // Handle worker result if any
+        if let Ok(worker_result) = poll_result {
+            // Track hashes by engine type (always, even for stale results)
+            match worker_result.engine_type {
+                EngineType::Cpu => {
+                    cpu_hashes += worker_result.hash_count;
+                    metrics::record_cpu_hashes(worker_result.hash_count);
+                }
+                EngineType::Gpu => {
+                    gpu_hashes += worker_result.hash_count;
+                    metrics::record_gpu_hashes(worker_result.hash_count);
+                }
+            }
+
+            // Check if this result is for the current job (not stale)
+            if worker_result.job_id != internal_job_id {
+                log::debug!(
+                    "⏰ Discarding stale result from worker {} (result job_id {} != current {})",
+                    worker_result.thread_id,
+                    worker_result.job_id,
+                    internal_job_id
+                );
+                continue;
+            }
+
+            // Only send result for the FIRST solution found for THIS job
+            if let Some(candidate) = worker_result.candidate {
+                if !result_sent_for_current_job {
+                    if let Some(ref job_id) = node_job_id {
+                        let total_hashes = cpu_hashes + gpu_hashes;
+                        let elapsed = job_start_time
+                            .map(|t| t.elapsed().as_secs_f64())
+                            .unwrap_or(0.0);
+
+                        log::info!(
+                            "⛏️ Job {job_id} completed: {total_hashes} hashes in {elapsed:.2}s ({})",
+                            format_hashrate(total_hashes as f64 / elapsed.max(0.001))
+                        );
+
+                        // Mark as sent BEFORE sending to prevent duplicates
+                        result_sent_for_current_job = true;
+                        worker_pool.cancel();
+                        metrics::set_active_jobs(0);
+
+                        let result = MiningResult {
+                            status: ApiResponseStatus::Completed,
+                            job_id: job_id.clone(),
+                            nonce: Some(format!("{:x}", candidate.nonce)),
+                            work: Some(hex::encode(candidate.work)),
+                            hash_count: total_hashes,
+                            elapsed_time: elapsed,
+                            miner_id: None,
+                        };
+
+                        let msg = MinerMessage::JobResult(result);
+                        send_message_checked(&connection, &mut send, &msg).await?;
+                    }
+                }
+            }
+        }
+
+        // Check for incoming messages and connection health
+        tokio::select! {
+            biased;
+
+            reason = connection.closed() => {
+                let msg = reason.to_string();
+                if !*authenticated && msg.to_ascii_lowercase().contains("auth") {
+                    return Err(quic_transport::PermanentConnectError(format!(
+                        "node rejected miner auth ({msg}); check --auth-token / miner-auth-token"
+                    ))
+                    .into());
+                }
+                return Err(anyhow::anyhow!("Connection closed: {}", msg));
+            }
+
+            msg_result = read_message(&mut recv) => {
+                match msg_result {
+                    Ok(MinerMessage::NewJob(request)) => {
+                        *authenticated = true;
+                        log::info!(
+                            "⛏️ Received job: id={}, hash=0x{}",
+                            request.job_id,
+                            request.mining_hash
+                        );
+
+                        // Parse header hash
+                        let header_hash: [u8; 32] = match hex::decode(&request.mining_hash) {
+                            Ok(bytes) if bytes.len() == 32 => bytes.try_into().unwrap(),
+                            _ => {
+                                log::warn!("Invalid mining_hash in request");
+                                let result = MiningResult {
+                                    status: ApiResponseStatus::Failed,
+                                    job_id: request.job_id,
+                                    nonce: None,
+                                    work: None,
+                                    hash_count: 0,
+                                    elapsed_time: 0.0,
+                                    miner_id: None,
+                                };
+                                let msg = MinerMessage::JobResult(result);
+                                send_message_checked(&connection, &mut send, &msg).await?;
+                                continue;
+                            }
+                        };
+
+                        // Parse difficulty
+                        let difficulty = match U512::from_dec_str(&request.difficulty) {
+                            Ok(d) => d,
+                            Err(_) => {
+                                log::warn!("Invalid difficulty in request: parse error");
+                                let result = MiningResult {
+                                    status: ApiResponseStatus::Failed,
+                                    job_id: request.job_id,
+                                    nonce: None,
+                                    work: None,
+                                    hash_count: 0,
+                                    elapsed_time: 0.0,
+                                    miner_id: None,
+                                };
+                                let msg = MinerMessage::JobResult(result);
+                                send_message_checked(&connection, &mut send, &msg).await?;
+                                continue;
+                            }
+                        };
+
+                        // Reject zero difficulty to prevent division-by-zero panic in JobContext::new
+                        if difficulty.is_zero() {
+                            log::warn!("Invalid difficulty in request: zero is not allowed");
+                            let result = MiningResult {
+                                status: ApiResponseStatus::Failed,
+                                job_id: request.job_id,
+                                nonce: None,
+                                work: None,
+                                hash_count: 0,
+                                elapsed_time: 0.0,
+                                miner_id: None,
+                            };
+                            let msg = MinerMessage::JobResult(result);
+                            send_message_checked(&connection, &mut send, &msg).await?;
+                            continue;
+                        }
+
+                        // Reset state for new job
+                        cpu_hashes = 0;
+                        gpu_hashes = 0;
+                        job_start_time = Some(Instant::now());
+                        node_job_id = Some(request.job_id.clone());
+                        result_sent_for_current_job = false;
+
+                        log::debug!("Starting job {}", request.job_id);
+                        metrics::set_active_jobs(1);
+
+                        // start_job returns the internal job ID used to detect stale results
+                        internal_job_id = worker_pool.start_job(header_hash, difficulty);
+                    }
+                    Ok(MinerMessage::JobResult(_)) => {
+                        log::warn!("Received unexpected JobResult from node");
+                    }
+                    Ok(MinerMessage::Ready { .. }) => {
+                        log::warn!("Received unexpected Ready from node");
+                    }
+                    Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => {
+                        return Err(anyhow::anyhow!("Node disconnected"));
+                    }
+                    Err(e) => {
+                        return Err(anyhow::anyhow!("Read error: {}", e));
+                    }
+                }
+            }
+
+            // Short sleep to yield when no messages
+            _ = tokio::time::sleep(Duration::from_millis(1)) => {}
+        }
+    }
+}
@@ -21,5 +21,4 @@ serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true }
 tokio = { workspace = true, features = ["full"] }
 tokio-tungstenite = { version = "0.23", features = ["rustls-tls-webpki-roots"] }
-url = "2"
 uuid = { version = "1", features = ["fast-rng", "v4"] }
@@ -106,33 +106,28 @@ impl TelemetryHandle {
    }

    /// Emit a fully custom payload (must include `"msg"`).
-    pub async fn emit_payload(&self, payload: Value, verbosity: Option<u8>) {
+    ///
+    /// This is intentionally lossy: if the internal channel is full, the message is dropped
+    /// silently to avoid blocking mining paths or causing unbounded memory growth.
+    pub fn emit_payload(&self, payload: Value, verbosity: Option<u8>) {
        if self.senders.is_empty() {
            return;
        }
        let msg = EnvelopeMsg { payload, verbosity };
        for tx in self.senders.iter() {
-            // Drop on full; do not block mining paths. This is intentionally lossy under backpressure.
-            if tx.try_send(msg.clone()).is_err() {
-                // Best-effort: if try_send failed due to full buffer, attempt bounded send with small timeout.
-                // Avoid blocking indefinitely.
-                let tx = tx.clone();
-                let msg = msg.clone();
-                tokio::spawn(async move {
-                    let _ = tx.send(msg).await;
-                });
-            }
+            // Drop on full; do not block mining paths.
+            let _ = tx.try_send(msg.clone());
        }
    }

    /// Emit a `system.connected` message using configuration defaults and an optional link.
-    pub async fn emit_system_connected(&self, link: Option<&TelemetryNodeLink>) {
+    pub fn emit_system_connected(&self, link: Option<&TelemetryNodeLink>) {
        let payload = build_system_connected_payload(&self.config, link);
-        self.emit_payload(payload, Some(self.verbosity)).await;
+        self.emit_payload(payload, Some(self.verbosity));
    }

    /// Emit a `system.interval` message with lightweight health stats and an optional link hint.
-    pub async fn emit_system_interval(&self, interval: &SystemInterval, link_hint: Option<&str>) {
+    pub fn emit_system_interval(&self, interval: &SystemInterval, link_hint: Option<&str>) {
        let mut payload = json!({
            "msg": "system.interval",
            "uptime_ms": interval.uptime_ms,
@@ -169,7 +164,7 @@ impl TelemetryHandle {
                .insert("linked_node_hint".to_string(), json!(hint));
        }

-        self.emit_payload(payload, Some(self.verbosity)).await;
+        self.emit_payload(payload, Some(self.verbosity));
    }
 }

@@ -463,9 +458,7 @@ mod tests {
        let cfg = TelemetryConfig::default();
        let handle = start(cfg);
        assert!(!handle.is_enabled());
-        handle
-            .emit_system_interval(&SystemInterval::default(), None)
-            .await;
-        handle.emit_system_connected(None).await;
+        handle.emit_system_interval(&SystemInterval::default(), None);
+        handle.emit_system_connected(None);
    }
 }
@@ -0,0 +1,32 @@
+[package]
+name = "pool-service"
+version.workspace = true
+edition.workspace = true
+publish = false
+description = "Captcha share pool: turns browser proof-of-work captcha solves into Quantus mining shares"
+
+[dependencies]
+anyhow = { workspace = true }
+clap = { workspace = true }
+env_logger = { workspace = true }
+hex = { workspace = true }
+log = { workspace = true }
+primitive-types = { workspace = true }
+rand = { workspace = true }
+serde = { workspace = true }
+serde_json = { workspace = true, features = ["std"] }
+tokio = { workspace = true }
+warp = { workspace = true }
+
+# QUIC transport (shared with miner-service)
+quinn = "0.10"
+quic-transport = { path = "../quic-transport" }
+
+# Constant-time comparison for the site secret
+constant_time_eq = "0.3"
+
+# Protocol types from the node repo
+quantus-miner-api = { workspace = true }
+
+# Local crates
+pow-core = { path = "../pow-core" }
@@ -0,0 +1,219 @@
+//! HTTP API for the captcha widget (session/share) and for protected sites
+//! (siteverify), plus optional static file serving for the demo page.
+
+use std::net::{IpAddr, SocketAddr};
+use std::path::PathBuf;
+use std::sync::Arc;
+
+use primitive_types::U512;
+use serde::{Deserialize, Serialize};
+use warp::Filter;
+
+use crate::rate_limit::SessionIssuerLimiter;
+use crate::state::{PoolState, ShareOutcome};
+
+#[derive(Serialize)]
+struct SessionResponse {
+    session_id: String,
+    /// Hex, 32 bytes, no 0x prefix.
+    header_hash: String,
+    /// Hex U512, no 0x prefix. Solver iterates nonce_start..nonce_end.
+    nonce_start: String,
+    nonce_end: String,
+    /// Share target as hex U512: a share is valid iff hash < share_target.
+    /// (Target form avoids the solver needing U512 division.)
+    share_target: String,
+    /// Expected number of hashes to find a share (= share difficulty), FYI/UX.
+    expected_hashes: u64,
+    expires_in_secs: u64,
+}
+
+#[derive(Deserialize)]
+struct ShareRequest {
+    session_id: String,
+    /// Hex U512 nonce, no 0x prefix.
+    nonce: String,
+}
+
+#[derive(Serialize)]
+struct ShareResponse {
+    success: bool,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    token: Option<String>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    error: Option<&'static str>,
+    block_found: bool,
+}
+
+/// Request shape mirrors reCAPTCHA/Turnstile: `secret` + `response`.
+#[derive(Deserialize)]
+struct SiteVerifyRequest {
+    secret: String,
+    response: String,
+}
+
+#[derive(Serialize)]
+struct SiteVerifyResponse {
+    success: bool,
+    #[serde(rename = "challenge_ts", skip_serializing_if = "Option::is_none")]
+    challenge_ts: Option<u64>,
+    #[serde(rename = "error-codes", skip_serializing_if = "Vec::is_empty")]
+    error_codes: Vec<&'static str>,
+}
+
+pub async fn serve(
+    state: Arc<PoolState>,
+    limiter: Arc<SessionIssuerLimiter>,
+    addr: SocketAddr,
+    serve_dir: Option<PathBuf>,
+) {
+    let with_state = {
+        let state = state.clone();
+        warp::any().map(move || state.clone())
+    };
+    let with_limiter = {
+        let limiter = limiter.clone();
+        warp::any().map(move || limiter.clone())
+    };
+
+    // POST /api/session -> new captcha challenge
+    let session = warp::path!("api" / "session")
+        .and(warp::post())
+        .and(warp::addr::remote())
+        .and(with_limiter.clone())
+        .and(with_state.clone())
+        .map(
+            |remote: Option<SocketAddr>,
+             limiter: Arc<SessionIssuerLimiter>,
+             state: Arc<PoolState>| {
+                let ip = remote
+                    .map(|a| a.ip())
+                    .unwrap_or(IpAddr::from([127, 0, 0, 1]));
+                if let Err(reason) = limiter.check(ip) {
+                    return reply_error(warp::http::StatusCode::TOO_MANY_REQUESTS, reason);
+                }
+                match state.issue_session() {
+                    Err(reason) => reply_error(warp::http::StatusCode::SERVICE_UNAVAILABLE, reason),
+                    Ok(s) => {
+                        let target = U512::MAX / s.share_difficulty;
+                        let expected = s.share_difficulty.min(U512::from(u64::MAX)).as_u64();
+                        warp::reply::with_status(
+                            warp::reply::json(&SessionResponse {
+                                session_id: s.id,
+                                header_hash: hex::encode(s.job.header),
+                                nonce_start: format!("{:x}", s.nonce_start),
+                                nonce_end: format!("{:x}", s.nonce_end),
+                                share_target: format!("{:x}", target),
+                                expected_hashes: expected,
+                                expires_in_secs: crate::state::SESSION_TTL.as_secs(),
+                            }),
+                            warp::http::StatusCode::OK,
+                        )
+                    }
+                }
+            },
+        );
+
+    // POST /api/share -> verify a solved nonce, mint a token
+    let share = warp::path!("api" / "share")
+        .and(warp::post())
+        .and(warp::body::content_length_limit(4096))
+        .and(warp::body::json())
+        .and(with_state.clone())
+        .map(|req: ShareRequest, state: Arc<PoolState>| {
+            let nonce = match U512::from_str_radix(req.nonce.trim_start_matches("0x"), 16) {
+                Ok(n) => n,
+                Err(_) => {
+                    return warp::reply::with_status(
+                        warp::reply::json(&ShareResponse {
+                            success: false,
+                            token: None,
+                            error: Some("malformed_nonce"),
+                            block_found: false,
+                        }),
+                        warp::http::StatusCode::BAD_REQUEST,
+                    )
+                }
+            };
+            match state.submit_share(&req.session_id, nonce) {
+                ShareOutcome::Accepted { token, block_found } => warp::reply::with_status(
+                    warp::reply::json(&ShareResponse {
+                        success: true,
+                        token: Some(token),
+                        error: None,
+                        block_found,
+                    }),
+                    warp::http::StatusCode::OK,
+                ),
+                ShareOutcome::Rejected(reason) => warp::reply::with_status(
+                    warp::reply::json(&ShareResponse {
+                        success: false,
+                        token: None,
+                        error: Some(reason),
+                        block_found: false,
+                    }),
+                    warp::http::StatusCode::FORBIDDEN,
+                ),
+            }
+        });
+
+    // POST /siteverify -> protected site's backend redeems a token
+    let siteverify = warp::path!("siteverify")
+        .and(warp::post())
+        .and(warp::body::content_length_limit(4096))
+        // Accept both JSON and form encoding, like the incumbents do.
+        .and(
+            warp::body::json()
+                .or(warp::body::form())
+                .unify()
+                .map(|req: SiteVerifyRequest| req),
+        )
+        .and(with_state.clone())
+        .map(|req: SiteVerifyRequest, state: Arc<PoolState>| {
+            match state.verify_token(&req.secret, &req.response) {
+                Ok(ts) => warp::reply::json(&SiteVerifyResponse {
+                    success: true,
+                    challenge_ts: Some(ts),
+                    error_codes: vec![],
+                }),
+                Err(code) => warp::reply::json(&SiteVerifyResponse {
+                    success: false,
+                    challenge_ts: None,
+                    error_codes: vec![code],
+                }),
+            }
+        });
+
+    // GET /api/stats -> pool counters
+    let stats = warp::path!("api" / "stats")
+        .and(warp::get())
+        .and(with_state.clone())
+        .map(|state: Arc<PoolState>| warp::reply::json(&state.stats()));
+
+    let cors = warp::cors()
+        .allow_any_origin()
+        .allow_methods(vec!["GET", "POST", "OPTIONS"])
+        .allow_headers(vec!["content-type"]);
+
+    let api = session.or(share).or(siteverify).or(stats);
+
+    log::info!("HTTP API listening on {}", addr);
+    if let Some(dir) = serve_dir {
+        log::info!("Serving demo statics from {}", dir.display());
+        let statics = warp::get().and(warp::fs::dir(dir));
+        warp::serve(api.or(statics).with(cors)).run(addr).await;
+    } else {
+        warp::serve(api.with(cors)).run(addr).await;
+    }
+}
+
+fn reply_error(
+    status: warp::http::StatusCode,
+    error: &'static str,
+) -> warp::reply::WithStatus<warp::reply::Json> {
+    #[derive(Serialize)]
+    struct Err<'a> {
+        error: &'a str,
+    }
+    warp::reply::with_status(warp::reply::json(&Err { error }), status)
+}
@@ -0,0 +1,250 @@
+//! Quantus captcha share pool.
+//!
+//! Sits between a quantus-node (external miner protocol) and browser captcha
+//! solvers: hands out low-difficulty share challenges over the real block
+//! header, verifies solves, mints single-use tokens for site backends, and
+//! submits any share that happens to meet full network difficulty as a block.
+
+mod http;
+mod rate_limit;
+mod state;
+mod upstream;
+
+use std::net::SocketAddr;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::Duration;
+
+use clap::Parser;
+use primitive_types::U512;
+
+use crate::rate_limit::SessionRateLimit;
+
+#[derive(Parser, Debug)]
+#[command(name = "pool-service", about = "Quantus captcha share pool")]
+struct Args {
+    /// HTTP listen address for the captcha/siteverify API.
+    #[arg(long, default_value = "127.0.0.1:8787", env = "POOL_HTTP_ADDR")]
+    http_addr: SocketAddr,
+
+    /// Address of the quantus-node external-miner QUIC endpoint.
+    /// If omitted, runs in standalone mode with synthetic jobs (demo only).
+    #[arg(long, env = "POOL_NODE_ADDR")]
+    node_addr: Option<SocketAddr>,
+
+    /// Shared auth token from the node's `miner-auth-token` file.
+    /// Required when `--node-addr` is set. Prefer `--auth-token-file`.
+    #[arg(long, env = "POOL_AUTH_TOKEN", conflicts_with = "auth_token_file")]
+    auth_token: Option<String>,
+
+    /// Path to the node's `miner-auth-token` file (trimmed).
+    #[arg(long, env = "POOL_AUTH_TOKEN_FILE", conflicts_with = "auth_token")]
+    auth_token_file: Option<PathBuf>,
+
+    /// SHA-256 fingerprint of the node's miner TLS certificate (64 hex chars).
+    /// Required when `--node-addr` is set. Prefer `--tls-cert-sha256-file`.
+    #[arg(
+        long,
+        env = "POOL_TLS_CERT_SHA256",
+        conflicts_with = "tls_cert_sha256_file"
+    )]
+    tls_cert_sha256: Option<String>,
+
+    /// Path to the node's `miner-tls-cert-sha256` file.
+    #[arg(
+        long,
+        env = "POOL_TLS_CERT_SHA256_FILE",
+        conflicts_with = "tls_cert_sha256"
+    )]
+    tls_cert_sha256_file: Option<PathBuf>,
+
+    /// Share difficulty: expected number of hashes per captcha solve.
+    /// Measured browser WASM rate ≈ 120 kH/s on an M-series laptop, so
+    /// 50000 ≈ 0.4 s desktop / ~2 s phone. Raise for stronger rate limiting.
+    #[arg(long, default_value = "50000", env = "POOL_SHARE_DIFFICULTY")]
+    share_difficulty: u64,
+
+    /// Secret that protected-site backends must present to /siteverify.
+    #[arg(long, default_value = "dev-secret", env = "POOL_SITE_SECRET")]
+    site_secret: String,
+
+    /// Optional directory of static files to serve (demo page / widget).
+    #[arg(long, env = "POOL_SERVE_DIR")]
+    serve_dir: Option<PathBuf>,
+
+    /// Job rotation interval in standalone mode, seconds.
+    #[arg(long, default_value = "20")]
+    standalone_job_secs: u64,
+
+    /// Maximum live captcha sessions; /api/session returns 503 when full.
+    #[arg(long, default_value = "100000", env = "POOL_MAX_SESSIONS")]
+    max_sessions: usize,
+
+    /// Maximum live (unredeemed) share tokens; shares are refused when full.
+    #[arg(long, default_value = "100000", env = "POOL_MAX_TOKENS")]
+    max_tokens: usize,
+
+    /// Max /api/session issuances per client IP per minute.
+    #[arg(long, default_value = "60", env = "POOL_SESSIONS_PER_IP_PER_MIN")]
+    sessions_per_ip_per_min: u32,
+}
+
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    env_logger::Builder::from_env(env_logger::Env::default().default_filter_or("info")).init();
+    let args = Args::parse();
+
+    anyhow::ensure!(args.share_difficulty > 0, "--share-difficulty must be > 0");
+    if args.site_secret == "dev-secret" {
+        log::warn!("Using default site secret; set --site-secret in production");
+    }
+
+    // Resolve + validate node auth before starting HTTP / upstream tasks so a
+    // bad fingerprint or oversized token exits instead of retrying forever.
+    let node_upstream = match args.node_addr {
+        Some(addr) => {
+            let auth_token = resolve_auth_token(args.auth_token, args.auth_token_file)?;
+            let tls_cert_sha256 =
+                resolve_tls_cert_sha256(args.tls_cert_sha256, args.tls_cert_sha256_file)?;
+            quic_transport::validate_auth_config(&auth_token, &tls_cert_sha256)?;
+            Some((addr, auth_token, tls_cert_sha256))
+        }
+        None => {
+            if args.auth_token.is_some()
+                || args.auth_token_file.is_some()
+                || args.tls_cert_sha256.is_some()
+                || args.tls_cert_sha256_file.is_some()
+            {
+                log::warn!("Ignoring auth/TLS pin flags in standalone mode (no --node-addr)");
+            }
+            None
+        }
+    };
+
+    let (solution_tx, solution_rx) = tokio::sync::mpsc::channel(16);
+    let state = state::PoolState::new(
+        U512::from(args.share_difficulty),
+        args.site_secret.clone(),
+        solution_tx,
+        state::Limits {
+            max_sessions: args.max_sessions,
+            max_tokens: args.max_tokens,
+        },
+    );
+
+    let limiter = Arc::new(rate_limit::SessionIssuerLimiter::new(SessionRateLimit {
+        max_per_ip: args.sessions_per_ip_per_min,
+        window: Duration::from_secs(60),
+    }));
+
+    // Periodic cleanup of expired sessions/tokens and stale rate-limit windows.
+    {
+        let state = state.clone();
+        let limiter = limiter.clone();
+        tokio::spawn(async move {
+            loop {
+                tokio::time::sleep(Duration::from_secs(30)).await;
+                state.gc();
+                limiter.gc();
+            }
+        });
+    }
+
+    // Job source. Node mode reconnects across transient disconnects (node
+    // restart); only a PermanentConnectError ends the task — then we tear
+    // down HTTP so the process does not look healthy on a bad token/pin.
+    match node_upstream {
+        Some((addr, auth_token, tls_cert_sha256)) => {
+            log::info!("Upstream: quantus-node at {}", addr);
+            let state_for_upstream = state.clone();
+            let upstream = tokio::spawn(async move {
+                upstream::run_node_client(
+                    state_for_upstream,
+                    addr,
+                    auth_token,
+                    tls_cert_sha256,
+                    solution_rx,
+                )
+                .await
+            });
+
+            tokio::select! {
+                biased;
+                result = upstream => {
+                    match result {
+                        Ok(Ok(())) => anyhow::bail!(
+                            "upstream task exited unexpectedly (should reconnect forever)"
+                        ),
+                        Ok(Err(e)) => {
+                            log::error!("Upstream permanent failure; shutting down: {e}");
+                            Err(e.into())
+                        }
+                        Err(e) => Err(anyhow::anyhow!("upstream task panicked: {e}")),
+                    }
+                }
+                _ = http::serve(state, limiter, args.http_addr, args.serve_dir) => Ok(()),
+            }
+        }
+        None => {
+            log::warn!("No --node-addr given: running STANDALONE with synthetic jobs");
+            let state_for_upstream = state.clone();
+            tokio::spawn(async move {
+                upstream::run_standalone(
+                    state_for_upstream,
+                    Duration::from_secs(args.standalone_job_secs),
+                    solution_rx,
+                )
+                .await
+            });
+            http::serve(state, limiter, args.http_addr, args.serve_dir).await;
+            Ok(())
+        }
+    }
+}
+
+fn resolve_auth_token(
+    auth_token: Option<String>,
+    auth_token_file: Option<PathBuf>,
+) -> anyhow::Result<String> {
+    resolve_required_secret(
+        auth_token,
+        auth_token_file,
+        "--auth-token",
+        "--auth-token-file",
+        "when --node-addr is set, pass --auth-token-file <PATH> to the node's \
+         miner-auth-token file (or --auth-token <TOKEN>)",
+    )
+}
+
+fn resolve_tls_cert_sha256(value: Option<String>, file: Option<PathBuf>) -> anyhow::Result<String> {
+    resolve_required_secret(
+        value,
+        file,
+        "--tls-cert-sha256",
+        "--tls-cert-sha256-file",
+        "when --node-addr is set, pass --tls-cert-sha256-file <PATH> to the node's \
+         miner-tls-cert-sha256 file (or --tls-cert-sha256 <HEX>; also printed in node logs)",
+    )
+}
+
+fn resolve_required_secret(
+    value: Option<String>,
+    file: Option<PathBuf>,
+    value_flag: &str,
+    file_flag: &str,
+    missing_msg: &str,
+) -> anyhow::Result<String> {
+    if let Some(value) = value {
+        let value = value.trim().to_string();
+        anyhow::ensure!(!value.is_empty(), "{value_flag} is empty");
+        return Ok(value);
+    }
+    if let Some(path) = file {
+        let contents = std::fs::read_to_string(&path)
+            .map_err(|e| anyhow::anyhow!("failed to read {file_flag} {}: {}", path.display(), e))?;
+        let value = contents.trim().to_string();
+        anyhow::ensure!(!value.is_empty(), "{file_flag} {} is empty", path.display());
+        return Ok(value);
+    }
+    anyhow::bail!("{missing_msg}");
+}
@@ -0,0 +1,104 @@
+//! Per-client rate limiting for unauthenticated endpoints.
+//!
+//! `/api/session` is free, so a hard map cap alone still lets an attacker
+//! churn entries at the TTL boundary. A per-IP issuance ceiling bounds that
+//! even when the global cap has headroom.
+
+use std::collections::HashMap;
+use std::net::IpAddr;
+use std::sync::Mutex;
+use std::time::{Duration, Instant};
+
+#[derive(Debug, Clone, Copy)]
+pub struct SessionRateLimit {
+    /// Max session issuances per IP per window.
+    pub max_per_ip: u32,
+    pub window: Duration,
+}
+
+impl Default for SessionRateLimit {
+    fn default() -> Self {
+        SessionRateLimit {
+            // 60 sessions/min/IP is generous for real users (a page load is
+            // one) but stops a single host from dominating issuance.
+            max_per_ip: 60,
+            window: Duration::from_secs(60),
+        }
+    }
+}
+
+#[derive(Default)]
+struct IpWindow {
+    count: u32,
+    window_start: Option<Instant>,
+}
+
+pub struct SessionIssuerLimiter {
+    cfg: SessionRateLimit,
+    by_ip: Mutex<HashMap<IpAddr, IpWindow>>,
+}
+
+impl SessionIssuerLimiter {
+    pub fn new(cfg: SessionRateLimit) -> Self {
+        SessionIssuerLimiter {
+            cfg,
+            by_ip: Mutex::new(HashMap::new()),
+        }
+    }
+
+    /// Returns Ok(()) if this IP may issue another session, Err otherwise.
+    pub fn check(&self, ip: IpAddr) -> Result<(), &'static str> {
+        let now = Instant::now();
+        let mut map = self.by_ip.lock().unwrap();
+        let entry = map.entry(ip).or_default();
+        let window_start = entry.window_start.get_or_insert_with(Instant::now);
+        if now.duration_since(*window_start) >= self.cfg.window {
+            *window_start = now;
+            entry.count = 0;
+        }
+        if entry.count >= self.cfg.max_per_ip {
+            return Err("rate_limited");
+        }
+        entry.count += 1;
+        Ok(())
+    }
+
+    /// Drop stale IP windows so the side table doesn't grow forever.
+    pub fn gc(&self) {
+        let cutoff = self.cfg.window * 2;
+        let now = Instant::now();
+        self.by_ip.lock().unwrap().retain(|_, w| {
+            w.window_start
+                .map(|t| now.duration_since(t) < cutoff)
+                .unwrap_or(true)
+        });
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::net::{Ipv4Addr, Ipv6Addr};
+
+    #[test]
+    fn allows_up_to_cap_then_rejects() {
+        let lim = SessionIssuerLimiter::new(SessionRateLimit {
+            max_per_ip: 2,
+            window: Duration::from_secs(60),
+        });
+        let ip = IpAddr::V4(Ipv4Addr::new(10, 0, 0, 1));
+        assert!(lim.check(ip).is_ok());
+        assert!(lim.check(ip).is_ok());
+        assert_eq!(lim.check(ip), Err("rate_limited"));
+    }
+
+    #[test]
+    fn ips_are_independent() {
+        let lim = SessionIssuerLimiter::new(SessionRateLimit {
+            max_per_ip: 1,
+            window: Duration::from_secs(60),
+        });
+        assert!(lim.check(IpAddr::V4(Ipv4Addr::new(1, 2, 3, 4))).is_ok());
+        assert!(lim.check(IpAddr::V6(Ipv6Addr::LOCALHOST)).is_ok());
+    }
+}
@@ -0,0 +1,527 @@
+//! Core pool state: the current mining job, captcha sessions with disjoint
+//! nonce ranges, and single-use share tokens.
+
+use std::collections::HashMap;
+use std::sync::atomic::{AtomicU64, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
+
+use primitive_types::U512;
+use rand::RngCore;
+
+/// How long a captcha session stays solvable after issuance.
+pub const SESSION_TTL: Duration = Duration::from_secs(120);
+/// How long a share token stays redeemable via /siteverify.
+pub const TOKEN_TTL: Duration = Duration::from_secs(300);
+/// Each session gets a disjoint slice of the nonce space this many nonces wide.
+/// 2^64 nonces is unreachable within a session TTL at browser hash rates.
+pub const SESSION_RANGE_BITS: u32 = 64;
+
+/// Caps on live entries in the in-memory maps. `/api/session` is free and
+/// unauthenticated, so without a ceiling a request flood becomes a memory-DoS
+/// on the anti-abuse service itself. At the ~300 bytes/entry ballpark the
+/// defaults bound each map to tens of MB.
+#[derive(Debug, Clone, Copy)]
+pub struct Limits {
+    pub max_sessions: usize,
+    pub max_tokens: usize,
+}
+
+impl Default for Limits {
+    fn default() -> Self {
+        Limits {
+            max_sessions: 100_000,
+            max_tokens: 100_000,
+        }
+    }
+}
+
+/// The job the pool is currently handing out to captcha sessions.
+#[derive(Debug, Clone)]
+pub struct Job {
+    /// Job id assigned by the upstream node (or synthesized in standalone mode).
+    pub job_id: String,
+    /// Block header hash to mine over.
+    pub header: [u8; 32],
+    /// Full network difficulty for this job (block found if a share meets it).
+    pub network_difficulty: U512,
+}
+
+/// A captcha session: one issued challenge, bound to a disjoint nonce range.
+#[derive(Debug, Clone)]
+pub struct Session {
+    pub id: String,
+    /// Snapshot of the job at issuance time. Shares are verified against this
+    /// header even if the pool has since moved to a newer job (grace period =
+    /// session TTL); only network-difficulty solutions for the *current* job
+    /// are submitted upstream.
+    pub job: Job,
+    pub nonce_start: U512,
+    pub nonce_end: U512, // exclusive
+    pub share_difficulty: U512,
+    pub issued_at: Instant,
+    pub solved: bool,
+}
+
+/// A verified share, redeemable exactly once by the protected site's backend.
+#[derive(Debug, Clone)]
+pub struct ShareToken {
+    pub created_at: Instant,
+    pub created_unix: u64,
+    pub consumed: bool,
+}
+
+#[derive(Debug, Default, Clone, serde::Serialize)]
+pub struct PoolStats {
+    pub sessions_issued: u64,
+    pub shares_accepted: u64,
+    pub shares_rejected: u64,
+    pub tokens_verified: u64,
+    pub blocks_found: u64,
+}
+
+pub struct PoolState {
+    /// Current job, None until the first job arrives from the job source.
+    current_job: Mutex<Option<Job>>,
+    sessions: Mutex<HashMap<String, Session>>,
+    tokens: Mutex<HashMap<String, ShareToken>>,
+    stats: Mutex<PoolStats>,
+    /// Monotonic counter used to carve disjoint nonce ranges out of U512 space.
+    range_counter: AtomicU64,
+    /// Difficulty for captcha shares (expected hashes per solve).
+    pub share_difficulty: U512,
+    /// Site secret checked by /siteverify.
+    ///
+    /// NOTE: single-tenant by design for now. All tokens are interchangeable
+    /// across any backend holding this secret; per-sitekey secrets land with
+    /// the host-registration follow-up.
+    pub site_secret: String,
+    limits: Limits,
+    /// Sender used to push network-difficulty solutions upstream.
+    pub solution_tx: tokio::sync::mpsc::Sender<FoundBlock>,
+}
+
+/// A share that met full network difficulty: a real block solution.
+#[derive(Debug, Clone)]
+pub struct FoundBlock {
+    pub job_id: String,
+    pub nonce: U512,
+}
+
+pub enum ShareOutcome {
+    /// Share accepted; token the client can hand to the protected site.
+    Accepted {
+        token: String,
+        block_found: bool,
+    },
+    Rejected(&'static str),
+}
+
+fn random_id() -> String {
+    let mut bytes = [0u8; 24];
+    rand::rng().fill_bytes(&mut bytes);
+    hex::encode(bytes)
+}
+
+fn unix_now() -> u64 {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map(|d| d.as_secs())
+        .unwrap_or(0)
+}
+
+impl PoolState {
+    pub fn new(
+        share_difficulty: U512,
+        site_secret: String,
+        solution_tx: tokio::sync::mpsc::Sender<FoundBlock>,
+        limits: Limits,
+    ) -> Arc<Self> {
+        assert!(
+            !share_difficulty.is_zero(),
+            "share difficulty must be non-zero"
+        );
+        Arc::new(PoolState {
+            current_job: Mutex::new(None),
+            sessions: Mutex::new(HashMap::new()),
+            tokens: Mutex::new(HashMap::new()),
+            stats: Mutex::new(PoolStats::default()),
+            range_counter: AtomicU64::new(0),
+            share_difficulty,
+            site_secret,
+            limits,
+            solution_tx,
+        })
+    }
+
+    /// Install a new job (from the upstream node or the standalone generator).
+    /// Existing sessions keep their job snapshot and remain solvable until TTL.
+    pub fn set_job(&self, job: Job) {
+        log::info!(
+            "New job {}: header=0x{} network_difficulty={}",
+            job.job_id,
+            hex::encode(job.header),
+            pow_core::format_u512(job.network_difficulty)
+        );
+        *self.current_job.lock().unwrap() = Some(job);
+    }
+
+    pub fn current_job(&self) -> Option<Job> {
+        self.current_job.lock().unwrap().clone()
+    }
+
+    /// Drop the current job (e.g. upstream disconnected). New sessions will
+    /// see `no_job_available` until the next `set_job`.
+    pub fn clear_job(&self) {
+        if self.current_job.lock().unwrap().take().is_some() {
+            log::info!("Cleared current job (upstream unavailable)");
+        }
+    }
+
+    /// Issue a new captcha session with a disjoint nonce range.
+    pub fn issue_session(&self) -> Result<Session, &'static str> {
+        let job = self.current_job().ok_or("no_job_available")?;
+
+        if self.sessions.lock().unwrap().len() >= self.limits.max_sessions {
+            return Err("at_capacity");
+        }
+
+        // Carve the next disjoint 2^64-nonce slice. The counter is 64 bits and
+        // slices are 2^64 wide, so slices live in the bottom 2^128 of the U512
+        // nonce space and never collide with each other.
+        let index = self.range_counter.fetch_add(1, Ordering::Relaxed);
+        let nonce_start = U512::from(index) << SESSION_RANGE_BITS;
+        let nonce_end = nonce_start + (U512::one() << SESSION_RANGE_BITS);
+
+        let session = Session {
+            id: random_id(),
+            job,
+            nonce_start,
+            nonce_end,
+            share_difficulty: self.share_difficulty,
+            issued_at: Instant::now(),
+            solved: false,
+        };
+
+        self.sessions
+            .lock()
+            .unwrap()
+            .insert(session.id.clone(), session.clone());
+        self.stats.lock().unwrap().sessions_issued += 1;
+        Ok(session)
+    }
+
+    /// Verify a submitted share and, if valid, mint a single-use token.
+    pub fn submit_share(&self, session_id: &str, nonce: U512) -> ShareOutcome {
+        let session = {
+            let mut sessions = self.sessions.lock().unwrap();
+            match sessions.get_mut(session_id) {
+                None => return self.reject("unknown_session"),
+                Some(s) => {
+                    if s.solved {
+                        return self.reject("session_already_solved");
+                    }
+                    if s.issued_at.elapsed() > SESSION_TTL {
+                        sessions.remove(session_id);
+                        return self.reject("session_expired");
+                    }
+                    // Mark solved optimistically; unmark on verification failure
+                    // would allow retries, but a failed share means a buggy or
+                    // malicious client, so we burn the session either way.
+                    s.solved = true;
+                    s.clone()
+                }
+            }
+        };
+
+        if nonce < session.nonce_start || nonce >= session.nonce_end {
+            return self.reject("nonce_out_of_range");
+        }
+
+        let (share_valid, hash) = pow_core::is_valid_nonce(
+            session.job.header,
+            nonce.to_big_endian(),
+            session.share_difficulty,
+        );
+        if !share_valid {
+            return self.reject("share_below_target");
+        }
+
+        // Jackpot check: does this share meet full network difficulty for the
+        // job it was issued against? Only submit if that job is still current.
+        // Target derivation delegates to pow-core's canonical JobContext.
+        let network_target =
+            pow_core::JobContext::new(session.job.header, session.job.network_difficulty).target;
+        let mut block_found = false;
+        if hash < network_target {
+            let still_current = self
+                .current_job()
+                .map(|j| j.job_id == session.job.job_id)
+                .unwrap_or(false);
+            if still_current {
+                log::info!(
+                    "BLOCK FOUND by captcha share! job={} nonce={:x}",
+                    session.job.job_id,
+                    nonce
+                );
+                // This is the revenue event: count it and be loud if the
+                // upstream queue can't take it.
+                match self.solution_tx.try_send(FoundBlock {
+                    job_id: session.job.job_id.clone(),
+                    nonce,
+                }) {
+                    Ok(()) => {
+                        block_found = true;
+                        self.stats.lock().unwrap().blocks_found += 1;
+                    }
+                    Err(e) => {
+                        log::error!(
+                            "DROPPED block solution for job {}: upstream queue unavailable ({})",
+                            session.job.job_id,
+                            e
+                        );
+                    }
+                }
+            }
+        }
+
+        let token = random_id();
+        {
+            let mut tokens = self.tokens.lock().unwrap();
+            // A full token map means someone is farming shares faster than
+            // TTL expiry; refuse rather than grow without bound. The client
+            // paid hashes for this, so make the refusal explicit.
+            if tokens.len() >= self.limits.max_tokens {
+                return self.reject("at_capacity");
+            }
+            tokens.insert(
+                token.clone(),
+                ShareToken {
+                    created_at: Instant::now(),
+                    created_unix: unix_now(),
+                    consumed: false,
+                },
+            );
+        }
+        self.stats.lock().unwrap().shares_accepted += 1;
+        ShareOutcome::Accepted { token, block_found }
+    }
+
+    /// Redeem a share token (the reCAPTCHA/Turnstile "siteverify" semantics):
+    /// valid exactly once, only with the correct site secret.
+    pub fn verify_token(&self, secret: &str, token: &str) -> Result<u64, &'static str> {
+        if !constant_time_eq::constant_time_eq(secret.as_bytes(), self.site_secret.as_bytes()) {
+            return Err("invalid_secret");
+        }
+        let mut tokens = self.tokens.lock().unwrap();
+        match tokens.get_mut(token) {
+            None => Err("invalid_token"),
+            Some(t) if t.consumed => Err("token_already_consumed"),
+            Some(t) if t.created_at.elapsed() > TOKEN_TTL => {
+                tokens.remove(token);
+                Err("token_expired")
+            }
+            Some(t) => {
+                t.consumed = true;
+                let ts = t.created_unix;
+                self.stats.lock().unwrap().tokens_verified += 1;
+                Ok(ts)
+            }
+        }
+    }
+
+    pub fn stats(&self) -> PoolStats {
+        self.stats.lock().unwrap().clone()
+    }
+
+    /// Drop expired sessions and tokens. Called periodically.
+    pub fn gc(&self) {
+        self.sessions
+            .lock()
+            .unwrap()
+            .retain(|_, s| s.issued_at.elapsed() <= SESSION_TTL);
+        self.tokens
+            .lock()
+            .unwrap()
+            .retain(|_, t| t.created_at.elapsed() <= TOKEN_TTL);
+    }
+
+    fn reject(&self, reason: &'static str) -> ShareOutcome {
+        self.stats.lock().unwrap().shares_rejected += 1;
+        ShareOutcome::Rejected(reason)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn test_state(
+        share_difficulty: u64,
+    ) -> (Arc<PoolState>, tokio::sync::mpsc::Receiver<FoundBlock>) {
+        let (tx, rx) = tokio::sync::mpsc::channel(4);
+        let state = PoolState::new(
+            U512::from(share_difficulty),
+            "secret".into(),
+            tx,
+            Limits::default(),
+        );
+        state.set_job(Job {
+            job_id: "1".into(),
+            header: [7u8; 32],
+            network_difficulty: U512::MAX, // effectively unreachable
+        });
+        (state, rx)
+    }
+
+    /// Brute-force a valid share within the session's range.
+    fn solve(session: &Session) -> U512 {
+        let mut nonce = session.nonce_start;
+        loop {
+            let (ok, _) = pow_core::is_valid_nonce(
+                session.job.header,
+                nonce.to_big_endian(),
+                session.share_difficulty,
+            );
+            if ok {
+                return nonce;
+            }
+            nonce += U512::one();
+        }
+    }
+
+    #[test]
+    fn accepts_valid_share_and_verifies_token_once() {
+        let (state, _rx) = test_state(2); // ~2 hashes per solve
+        let session = state.issue_session().unwrap();
+        let nonce = solve(&session);
+
+        let token = match state.submit_share(&session.id, nonce) {
+            ShareOutcome::Accepted { token, block_found } => {
+                assert!(!block_found);
+                token
+            }
+            ShareOutcome::Rejected(r) => panic!("rejected: {r}"),
+        };
+
+        assert!(state.verify_token("secret", &token).is_ok());
+        assert_eq!(
+            state.verify_token("secret", &token),
+            Err("token_already_consumed")
+        );
+        assert_eq!(state.verify_token("wrong", &token), Err("invalid_secret"));
+    }
+
+    #[test]
+    fn rejects_out_of_range_nonce() {
+        let (state, _rx) = test_state(1); // difficulty 1: every nonce is a valid hash
+        let session = state.issue_session().unwrap();
+        // A nonce outside the assigned slice must be rejected even though its
+        // hash trivially meets the share target.
+        let outside = session.nonce_end;
+        match state.submit_share(&session.id, outside) {
+            ShareOutcome::Rejected("nonce_out_of_range") => {}
+            other => panic!(
+                "expected out-of-range rejection, got {:?}",
+                matches_str(&other)
+            ),
+        }
+    }
+
+    #[test]
+    fn burns_session_after_one_submission() {
+        let (state, _rx) = test_state(1);
+        let session = state.issue_session().unwrap();
+        let nonce = session.nonce_start;
+        assert!(matches!(
+            state.submit_share(&session.id, nonce),
+            ShareOutcome::Accepted { .. }
+        ));
+        match state.submit_share(&session.id, nonce) {
+            ShareOutcome::Rejected("session_already_solved") => {}
+            other => panic!("expected burned session, got {:?}", matches_str(&other)),
+        }
+    }
+
+    #[test]
+    fn disjoint_ranges() {
+        let (state, _rx) = test_state(1);
+        let a = state.issue_session().unwrap();
+        let b = state.issue_session().unwrap();
+        assert!(a.nonce_end <= b.nonce_start || b.nonce_end <= a.nonce_start);
+    }
+
+    #[test]
+    fn block_found_signalled_when_share_meets_network_difficulty() {
+        let (tx, mut rx) = tokio::sync::mpsc::channel(4);
+        // Network difficulty == share difficulty == 2: any valid share is a block.
+        let state = PoolState::new(U512::from(2u64), "secret".into(), tx, Limits::default());
+        state.set_job(Job {
+            job_id: "1".into(),
+            header: [9u8; 32],
+            network_difficulty: U512::from(2u64),
+        });
+        let session = state.issue_session().unwrap();
+        let nonce = solve(&session);
+        match state.submit_share(&session.id, nonce) {
+            ShareOutcome::Accepted { block_found, .. } => assert!(block_found),
+            ShareOutcome::Rejected(r) => panic!("rejected: {r}"),
+        }
+        let block = rx
+            .try_recv()
+            .expect("block solution should be queued upstream");
+        assert_eq!(block.job_id, "1");
+        assert_eq!(block.nonce, nonce);
+    }
+
+    #[test]
+    fn session_cap_rejects_flood() {
+        let (tx, _rx) = tokio::sync::mpsc::channel(4);
+        let state = PoolState::new(
+            U512::from(1u64),
+            "secret".into(),
+            tx,
+            Limits {
+                max_sessions: 2,
+                max_tokens: 2,
+            },
+        );
+        state.set_job(Job {
+            job_id: "1".into(),
+            header: [7u8; 32],
+            network_difficulty: U512::MAX,
+        });
+        assert!(state.issue_session().is_ok());
+        assert!(state.issue_session().is_ok());
+        assert!(matches!(state.issue_session(), Err("at_capacity")));
+    }
+
+    #[test]
+    fn dropped_block_does_not_count_or_flag() {
+        // Zero-capacity channel: try_send always fails, simulating a stalled
+        // upstream. The share must still be accepted (the user did the work),
+        // but block_found must be false and blocks_found must stay 0.
+        let (tx, rx) = tokio::sync::mpsc::channel(1);
+        drop(rx); // closed channel -> try_send errors
+        let state = PoolState::new(U512::from(2u64), "secret".into(), tx, Limits::default());
+        state.set_job(Job {
+            job_id: "1".into(),
+            header: [9u8; 32],
+            network_difficulty: U512::from(2u64),
+        });
+        let session = state.issue_session().unwrap();
+        let nonce = solve(&session);
+        match state.submit_share(&session.id, nonce) {
+            ShareOutcome::Accepted { block_found, .. } => assert!(!block_found),
+            ShareOutcome::Rejected(r) => panic!("rejected: {r}"),
+        }
+        assert_eq!(state.stats().blocks_found, 0);
+    }
+
+    fn matches_str(o: &ShareOutcome) -> &'static str {
+        match o {
+            ShareOutcome::Accepted { .. } => "accepted",
+            ShareOutcome::Rejected(r) => r,
+        }
+    }
+}
@@ -0,0 +1,192 @@
+//! Job sources: either a QUIC connection to a real quantus-node (speaking the
+//! external miner protocol), or a standalone generator for demos/tests.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use primitive_types::U512;
+use quantus_miner_api::{
+    read_message, write_message, ApiResponseStatus, MinerMessage, MiningResult,
+};
+use tokio::sync::mpsc::Receiver;
+
+use crate::state::{FoundBlock, Job, PoolState};
+
+/// Connect to a node as an external miner and keep the pool's job current.
+/// Solutions arriving on `solutions` are pushed upstream as JobResults.
+///
+/// Transient disconnects (node restart, network blip) clear the current job
+/// and reconnect with backoff — the pool process stays up. Only
+/// [`quic_transport::PermanentConnectError`] (bad pin / rejected auth) returns
+/// `Err` so `main` can shut down the HTTP side.
+pub async fn run_node_client(
+    state: Arc<PoolState>,
+    node_addr: SocketAddr,
+    auth_token: String,
+    tls_cert_sha256: String,
+    mut solutions: Receiver<FoundBlock>,
+) -> Result<(), quic_transport::PermanentConnectError> {
+    let mut reconnect_delay = Duration::from_secs(1);
+    const MAX_RECONNECT_DELAY: Duration = Duration::from_secs(30);
+
+    // A block taken off the channel but not yet acknowledged by a successful
+    // upstream write. Kept across reconnects so a failed write doesn't drop
+    // the block (the node may still accept it after reconnection if the job
+    // hasn't moved on).
+    let mut pending: Option<FoundBlock> = None;
+
+    loop {
+        log::info!("Connecting to node at {}...", node_addr);
+        match quic_transport::connect(node_addr, &auth_token, &tls_cert_sha256).await {
+            Ok((connection, mut send, mut recv)) => {
+                log::info!("Connected to node at {}", node_addr);
+                let mut authenticated = false;
+
+                loop {
+                    // Retry any unsubmitted block before doing anything else.
+                    if let Some(block) = pending.take() {
+                        match submit_block(&mut send, &block).await {
+                            Ok(()) => {
+                                log::info!(
+                                    "Submitted block solution to node (job {})",
+                                    block.job_id
+                                );
+                            }
+                            Err(e) => {
+                                log::error!("Failed to submit block solution, will retry: {}", e);
+                                pending = Some(block);
+                                break; // reconnect and retry
+                            }
+                        }
+                    }
+
+                    tokio::select! {
+                        biased;
+
+                        reason = connection.closed() => {
+                            let msg = reason.to_string();
+                            if !authenticated && msg.to_ascii_lowercase().contains("auth") {
+                                state.clear_job();
+                                return Err(quic_transport::PermanentConnectError(format!(
+                                    "node rejected miner auth ({msg}); check --auth-token / miner-auth-token"
+                                )));
+                            }
+                            log::warn!("Node connection closed: {}", msg);
+                            break;
+                        }
+
+                        // A captcha share met network difficulty: submit it.
+                        block = solutions.recv() => {
+                            if let Some(block) = block {
+                                // Park it in `pending`; the top of the loop
+                                // submits it and keeps it on write failure.
+                                pending = Some(block);
+                            }
+                        }
+
+                        msg = read_message(&mut recv) => {
+                            match msg {
+                                Ok(MinerMessage::NewJob(request)) => {
+                                    authenticated = true;
+                                    match parse_job(&request.job_id, &request.mining_hash, &request.difficulty) {
+                                        Ok(job) => state.set_job(job),
+                                        Err(e) => log::warn!("Ignoring malformed job from node: {}", e),
+                                    }
+                                }
+                                Ok(other) => {
+                                    log::warn!("Unexpected message from node: {:?}", other);
+                                }
+                                Err(e) => {
+                                    log::warn!("Read error from node: {}", e);
+                                    break;
+                                }
+                            }
+                        }
+                    }
+                }
+
+                // Node gone / stream dropped: stop handing out captchas on a
+                // stale header until we reconnect and get a fresh NewJob.
+                state.clear_job();
+                if authenticated {
+                    reconnect_delay = Duration::from_secs(1);
+                }
+            }
+            Err(e) => {
+                if let Some(perm) = e.downcast_ref::<quic_transport::PermanentConnectError>() {
+                    state.clear_job();
+                    return Err(quic_transport::PermanentConnectError(perm.0.clone()));
+                }
+                log::warn!("Failed to connect to node: {}", e);
+                state.clear_job();
+            }
+        }
+
+        log::info!("Reconnecting in {:?}...", reconnect_delay);
+        tokio::time::sleep(reconnect_delay).await;
+        reconnect_delay = (reconnect_delay * 2).min(MAX_RECONNECT_DELAY);
+    }
+}
+
+async fn submit_block(send: &mut quinn::SendStream, block: &FoundBlock) -> std::io::Result<()> {
+    let result = MiningResult {
+        status: ApiResponseStatus::Completed,
+        job_id: block.job_id.clone(),
+        nonce: Some(format!("{:x}", block.nonce)),
+        work: Some(hex::encode(block.nonce.to_big_endian())),
+        hash_count: 0,
+        elapsed_time: 0.0,
+        miner_id: None,
+    };
+    write_message(send, &MinerMessage::JobResult(result)).await
+}
+
+fn parse_job(job_id: &str, mining_hash: &str, difficulty: &str) -> anyhow::Result<Job> {
+    let bytes = hex::decode(mining_hash)?;
+    let header: [u8; 32] = bytes
+        .try_into()
+        .map_err(|_| anyhow::anyhow!("mining_hash must be 32 bytes"))?;
+    let network_difficulty =
+        U512::from_dec_str(difficulty).map_err(|e| anyhow::anyhow!("bad difficulty: {:?}", e))?;
+    if network_difficulty.is_zero() {
+        anyhow::bail!("zero difficulty");
+    }
+    Ok(Job {
+        job_id: job_id.to_string(),
+        header,
+        network_difficulty,
+    })
+}
+
+/// Standalone mode: synthesize a fresh random job every `interval` so the
+/// captcha works without a running node. Solutions are logged and dropped.
+pub async fn run_standalone(
+    state: Arc<PoolState>,
+    interval: Duration,
+    mut solutions: Receiver<FoundBlock>,
+) {
+    use rand::RngCore;
+    let mut counter: u64 = 0;
+
+    loop {
+        let mut header = [0u8; 32];
+        rand::rng().fill_bytes(&mut header);
+        counter += 1;
+        state.set_job(Job {
+            job_id: format!("standalone-{}", counter),
+            header,
+            // Unreachable "network" difficulty: standalone mode never finds blocks.
+            network_difficulty: U512::MAX,
+        });
+
+        tokio::select! {
+            _ = tokio::time::sleep(interval) => {}
+            block = solutions.recv() => {
+                if let Some(block) = block {
+                    log::info!("(standalone) would submit block for job {}", block.job_id);
+                }
+            }
+        }
+    }
+}
@@ -18,7 +18,6 @@ default = ["baseline", "std"]
 # std feature wires through to dependencies that have separate std/no_std builds.
 std = [
 	"primitive-types/std",
-	"qp-poseidon-core/std",
 ]

 # Baseline/reference implementation using BigUint-backed modular exponentiation.
@@ -34,14 +33,9 @@ simd-poseidon2 = []
 gpu-interop = []

 [dependencies]
-log = { workspace = true }
 primitive-types = { workspace = true }
-num-bigint = { workspace = true }
-num-traits = { workspace = true }
 qp-poseidon-core = { workspace = true }
 qpow-math = { workspace = true }
-thiserror = { workspace = true }
-anyhow = { workspace = true }

 [dev-dependencies]
 hex = { workspace = true }
@@ -1,8 +1,56 @@
 use primitive_types::U512;
+use qp_poseidon_core::{Goldilocks, Poseidon2};

+pub use qp_poseidon_core::SPONGE_WIDTH;
 pub use qpow_math::{get_nonce_hash, is_valid_nonce, mine_range};

-/// Job context for Bitcoin-style PoW mining with double Poseidon2 hashing
+/// Sponge state (canonical u64 felts) after absorbing the 32-byte header and the
+/// high 32 bytes of the big-endian nonce — the first two of the five Poseidon2
+/// permutations of `get_nonce_hash`. This state is identical for every nonce in a
+/// batch as long as incrementing the nonce never carries into its high 256 bits,
+/// so GPU kernels can resume the sponge from here and skip 2 of 5 permutations.
+pub fn mining_midstate(header: [u8; 32], nonce_high_be: [u8; 32]) -> [u64; SPONGE_WIDTH] {
+    let poseidon2 = Poseidon2::new();
+    let mut state = [Goldilocks::ZERO; SPONGE_WIDTH];
+    for (i, chunk) in header.chunks_exact(4).enumerate() {
+        state[i] += Goldilocks::from_u64(u32::from_le_bytes(chunk.try_into().unwrap()) as u64);
+    }
+    poseidon2.permute_mut(&mut state);
+    for (i, chunk) in nonce_high_be.chunks_exact(4).enumerate() {
+        state[i] += Goldilocks::from_u64(u32::from_le_bytes(chunk.try_into().unwrap()) as u64);
+    }
+    poseidon2.permute_mut(&mut state);
+    state.map(|g| g.as_canonical_u64())
+}
+
+/// Format a U512 in a human-readable way (scientific notation for large numbers).
+pub fn format_u512(n: U512) -> String {
+    if n.is_zero() {
+        return "0".to_string();
+    }
+    let s = format!("{}", n);
+    let len = s.len();
+    if len <= 12 {
+        s
+    } else {
+        format!("{}e{}", &s[..4], len - 1)
+    }
+}
+
+/// Format a hash rate with appropriate units (H/s, KH/s, MH/s, GH/s).
+pub fn format_hashrate(hashes_per_sec: f64) -> String {
+    if hashes_per_sec >= 1_000_000_000.0 {
+        format!("{:.2} GH/s", hashes_per_sec / 1_000_000_000.0)
+    } else if hashes_per_sec >= 1_000_000.0 {
+        format!("{:.2} MH/s", hashes_per_sec / 1_000_000.0)
+    } else if hashes_per_sec >= 1_000.0 {
+        format!("{:.2} KH/s", hashes_per_sec / 1_000.0)
+    } else {
+        format!("{:.2} H/s", hashes_per_sec)
+    }
+}
+
+/// Job context for PoW mining with Poseidon2 hash_squeeze_twice
 #[derive(Debug, Clone)]
 pub struct JobContext {
    pub header: [u8; 32],
@@ -11,7 +59,12 @@ pub struct JobContext {
 }

 impl JobContext {
-    /// Build a new context from header and difficulty
+    /// Build a new context from header and difficulty.
+    ///
+    /// # Panics
+    ///
+    /// Panics if `difficulty` is zero (division by zero in target calculation).
+    /// Callers must validate that difficulty is non-zero before calling this function.
    pub fn new(header: [u8; 32], difficulty: U512) -> Self {
        // In Bitcoin-style PoW, target = max_target / difficulty
        let max_target = U512::MAX;
@@ -35,7 +88,7 @@ pub fn step_nonce(nonce: U512) -> U512 {
    nonce.saturating_add(U512::from(1u64))
 }

-/// Compute hash for the current nonce using Bitcoin-style double Poseidon2
+/// Compute hash for the current nonce using Poseidon2 hash_squeeze_twice
 pub fn hash_from_nonce(ctx: &JobContext, nonce: U512) -> U512 {
    let nonce_bytes = nonce.to_big_endian();
    qpow_math::get_nonce_hash(ctx.header, nonce_bytes)
@@ -189,4 +242,47 @@ mod tests {
        // With impossible difficulty, should not find solution
        assert!(result.is_none());
    }
+
+    #[test]
+    fn test_midstate_resumes_to_full_hash() {
+        let header = [7u8; 32];
+        let nonce = (U512::from(0xdeadbeefcafeu64) << 300) | U512::from(0x1234567890abcdefu64);
+        let nonce_be = nonce.to_big_endian();
+
+        let mut state =
+            mining_midstate(header, nonce_be[..32].try_into().unwrap()).map(Goldilocks::from_u64);
+        let poseidon2 = Poseidon2::new();
+        for (i, chunk) in nonce_be[32..].chunks_exact(4).enumerate() {
+            state[i] += Goldilocks::from_u64(u32::from_le_bytes(chunk.try_into().unwrap()) as u64);
+        }
+        poseidon2.permute_mut(&mut state);
+        state[0] += Goldilocks::ONE;
+        state[1] += Goldilocks::ONE;
+        poseidon2.permute_mut(&mut state);
+
+        let mut hash = [0u8; 64];
+        for i in 0..4 {
+            hash[i * 8..(i + 1) * 8].copy_from_slice(&state[i].as_canonical_u64().to_le_bytes());
+        }
+        poseidon2.permute_mut(&mut state);
+        for i in 0..4 {
+            hash[32 + i * 8..32 + (i + 1) * 8]
+                .copy_from_slice(&state[i].as_canonical_u64().to_le_bytes());
+        }
+
+        assert_eq!(
+            U512::from_big_endian(&hash),
+            qpow_math::get_nonce_hash(header, nonce_be)
+        );
+    }
+
+    #[test]
+    #[should_panic(expected = "division by zero")]
+    fn test_zero_difficulty_panics() {
+        // Zero difficulty causes division by zero in target calculation.
+        // Callers MUST validate difficulty > 0 before calling JobContext::new.
+        let header = [7u8; 32];
+        let zero_difficulty = U512::zero();
+        let _ctx = JobContext::new(header, zero_difficulty);
+    }
 }
@@ -0,0 +1,15 @@
+[package]
+name = "quic-transport"
+version.workspace = true
+edition.workspace = true
+publish = false
+description = "Shared QUIC client transport for connecting to a quantus-node's external-miner endpoint"
+
+[dependencies]
+anyhow = { workspace = true }
+quantus-miner-api = { workspace = true }
+quinn = "0.10"
+rustls = { version = "0.21", default-features = false, features = ["dangerous_configuration", "quic"] }
+serde_json = { workspace = true }
+sha2 = "0.10"
+tokio = { workspace = true, features = ["macros", "time"] }
@@ -0,0 +1,218 @@
+//! Shared QUIC client used by miner-service and pool-service to connect to a
+//! quantus-node's external-miner endpoint.
+//!
+//! The node uses a persisted self-signed certificate. Callers must supply the
+//! expected SHA-256 fingerprint of that certificate (lowercase hex); the
+//! transport rejects any other server cert.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+use std::time::Duration;
+
+use quantus_miner_api::{write_message, MinerMessage, MAX_MESSAGE_SIZE};
+use quinn::{ClientConfig, Endpoint};
+use rustls::client::ServerCertVerified;
+use sha2::{Digest, Sha256};
+
+/// Permanent misconfiguration / auth failure: callers must not reconnect-loop.
+#[derive(Debug)]
+pub struct PermanentConnectError(pub String);
+
+impl std::fmt::Display for PermanentConnectError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "{}", self.0)
+    }
+}
+
+impl std::error::Error for PermanentConnectError {}
+
+/// Parse and validate miner auth inputs that never become valid by retrying.
+///
+/// Call this at process startup (before workers / HTTP) so a bad fingerprint
+/// or oversized token exits immediately instead of looping forever.
+pub fn validate_auth_config(auth_token: &str, tls_cert_sha256_hex: &str) -> anyhow::Result<()> {
+    parse_fingerprint(tls_cert_sha256_hex)?;
+    validate_ready_frame(auth_token)?;
+    Ok(())
+}
+
+/// Parse a 64-hex (optional whitespace/`:` separators) SHA-256 fingerprint.
+pub fn parse_fingerprint(tls_cert_sha256_hex: &str) -> anyhow::Result<[u8; 32]> {
+    normalize_fingerprint(tls_cert_sha256_hex).ok_or_else(|| {
+        PermanentConnectError("invalid TLS cert SHA-256 fingerprint (expected 64 hex chars)".into())
+            .into()
+    })
+}
+
+fn validate_ready_frame(auth_token: &str) -> anyhow::Result<()> {
+    let frame = serde_json::to_vec(&MinerMessage::Ready {
+        token: auth_token.to_string(),
+    })
+    .map_err(|e| PermanentConnectError(format!("failed to serialize Ready frame: {e}")))?;
+    if frame.len() > MAX_MESSAGE_SIZE as usize {
+        return Err(PermanentConnectError(format!(
+            "auth token serializes to a {}-byte Ready frame; max is {} \
+             (shorten the token or avoid characters that require JSON escaping)",
+            frame.len(),
+            MAX_MESSAGE_SIZE
+        ))
+        .into());
+    }
+    Ok(())
+}
+
+/// Establish a QUIC connection to the node's external-miner endpoint, open
+/// the bidirectional stream, and send the initial `Ready { token }` message.
+///
+/// `tls_cert_sha256_hex` is the lowercase hex SHA-256 of the node's miner TLS
+/// certificate DER (contents of the node's `miner-tls-cert-sha256` file).
+///
+/// Wrong tokens are rejected by the node closing with "auth failed"; that is
+/// returned as [`PermanentConnectError`] so callers do not reconnect forever.
+pub async fn connect(
+    addr: SocketAddr,
+    auth_token: &str,
+    tls_cert_sha256_hex: &str,
+) -> anyhow::Result<(quinn::Connection, quinn::SendStream, quinn::RecvStream)> {
+    // Re-check here so library callers that skip startup validation still fail
+    // closed instead of looping on a deterministic parse error.
+    let expected = parse_fingerprint(tls_cert_sha256_hex)?;
+    validate_ready_frame(auth_token)?;
+
+    let mut crypto = rustls::ClientConfig::builder()
+        .with_safe_defaults()
+        .with_custom_certificate_verifier(Arc::new(PinnedCertVerifier { expected }))
+        .with_no_client_auth();
+    // Versioned with the wire protocol (node side: MINER_ALPN). `/2` = the
+    // authenticated `Ready { token }` protocol; a mismatched node/miner pair
+    // fails at the TLS handshake with "no application protocol" instead of an
+    // opaque auth/deserialize error.
+    crypto.alpn_protocols = vec![b"quantus-miner/2".to_vec()];
+
+    let mut client_config = ClientConfig::new(Arc::new(crypto));
+    let mut transport_config = quinn::TransportConfig::default();
+    transport_config.keep_alive_interval(Some(Duration::from_secs(5)));
+    transport_config.max_idle_timeout(Some(Duration::from_secs(15).try_into().unwrap()));
+    client_config.transport_config(Arc::new(transport_config));
+
+    let mut endpoint = Endpoint::client("0.0.0.0:0".parse().unwrap())?;
+    endpoint.set_default_client_config(client_config);
+
+    let connection = match endpoint.connect(addr, "localhost")?.await {
+        Ok(c) => c,
+        Err(e) => {
+            let msg = e.to_string();
+            // Wrong pin fails the TLS handshake deterministically.
+            if msg.to_ascii_lowercase().contains("fingerprint") {
+                return Err(PermanentConnectError(format!(
+                    "TLS certificate pin rejected ({msg}); check --tls-cert-sha256 / \
+                     miner-tls-cert-sha256"
+                ))
+                .into());
+            }
+            return Err(e.into());
+        }
+    };
+    let (mut send, recv) = connection.open_bi().await?;
+    write_message(
+        &mut send,
+        &MinerMessage::Ready {
+            token: auth_token.to_string(),
+        },
+    )
+    .await?;
+
+    // Ready is fire-and-forget on the wire; the node closes immediately on a
+    // bad token. Wait briefly so that surfaces as a permanent error here
+    // instead of "connected" + 1s reconnect churn.
+    const AUTH_REJECT_GRACE: Duration = Duration::from_secs(2);
+    tokio::select! {
+        reason = connection.closed() => {
+            let msg = reason.to_string();
+            if msg.to_ascii_lowercase().contains("auth") {
+                return Err(PermanentConnectError(format!(
+                    "node rejected miner auth ({msg}); check --auth-token / miner-auth-token"
+                ))
+                .into());
+            }
+            return Err(anyhow::anyhow!("connection closed during auth handshake: {msg}"));
+        }
+        _ = tokio::time::sleep(AUTH_REJECT_GRACE) => {}
+    }
+
+    Ok((connection, send, recv))
+}
+
+fn normalize_fingerprint(input: &str) -> Option<[u8; 32]> {
+    let hex: String = input
+        .chars()
+        .filter(|c| !c.is_whitespace() && *c != ':')
+        .map(|c| c.to_ascii_lowercase())
+        .collect();
+    if hex.len() != 64 || !hex.chars().all(|c| c.is_ascii_hexdigit()) {
+        return None;
+    }
+    let mut out = [0u8; 32];
+    for i in 0..32 {
+        out[i] = u8::from_str_radix(&hex[i * 2..i * 2 + 2], 16).ok()?;
+    }
+    Some(out)
+}
+
+/// Accepts only a server certificate whose SHA-256 fingerprint matches.
+struct PinnedCertVerifier {
+    expected: [u8; 32],
+}
+
+impl rustls::client::ServerCertVerifier for PinnedCertVerifier {
+    fn verify_server_cert(
+        &self,
+        end_entity: &rustls::Certificate,
+        _intermediates: &[rustls::Certificate],
+        _server_name: &rustls::ServerName,
+        _scts: &mut dyn Iterator<Item = &[u8]>,
+        _ocsp_response: &[u8],
+        _now: std::time::SystemTime,
+    ) -> Result<ServerCertVerified, rustls::Error> {
+        let actual = Sha256::digest(&end_entity.0);
+        if actual.as_slice() != self.expected.as_slice() {
+            return Err(rustls::Error::General(
+                "server certificate SHA-256 fingerprint mismatch".into(),
+            ));
+        }
+        Ok(ServerCertVerified::assertion())
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn rejects_short_fingerprint() {
+        let err = validate_auth_config("token", "nope").unwrap_err();
+        assert!(
+            err.downcast_ref::<PermanentConnectError>().is_some(),
+            "{err}"
+        );
+    }
+
+    #[test]
+    fn accepts_valid_fingerprint_and_token() {
+        let fp = "a".repeat(64);
+        validate_auth_config("deadbeef", &fp).unwrap();
+    }
+
+    #[test]
+    fn rejects_token_whose_ready_frame_exceeds_limit() {
+        let fp = "a".repeat(64);
+        // Escape-heavy token: raw length under a naive cap can still blow the frame.
+        let token = "\"".repeat(600);
+        let err = validate_auth_config(&token, &fp).unwrap_err();
+        assert!(
+            err.downcast_ref::<PermanentConnectError>().is_some(),
+            "{err}"
+        );
+        assert!(err.to_string().contains("Ready frame"), "{err}");
+    }
+}
@@ -0,0 +1,13 @@
+[package]
+name = "solver-wasm"
+version.workspace = true
+edition.workspace = true
+publish = false
+description = "Browser WASM solver for Quantus captcha shares (raw C ABI, no wasm-bindgen)"
+
+[lib]
+crate-type = ["cdylib", "rlib"]
+
+[dependencies]
+primitive-types = { workspace = true }
+qpow-math = { workspace = true }
@@ -0,0 +1,101 @@
+//! Browser captcha solver: grinds Poseidon2 nonce hashes over a block header.
+//!
+//! Exposed as a raw C-ABI WASM module (no wasm-bindgen) so it can be built
+//! with a bare `cargo build --target wasm32-unknown-unknown` and loaded with
+//! minimal JS glue. All I/O goes through one static buffer:
+//!
+//! ```text
+//! offset   0..32   header hash                     (in)
+//! offset  32..96   nonce, 64-byte big-endian       (in: start; out: found/next)
+//! offset  96..160  share target, 64-byte BE        (in)
+//! offset 160..224  winning hash, 64-byte BE        (out, only when found)
+//! ```
+//!
+//! JS calls `solve(iters)` in chunks (e.g. 5–20k iterations) to keep the
+//! worker responsive; a return of 0 means "keep going" (the nonce field has
+//! been advanced), 1 means the nonce field now holds a valid share.
+
+use primitive_types::U512;
+
+const HEADER_OFF: usize = 0;
+const NONCE_OFF: usize = 32;
+const TARGET_OFF: usize = 96;
+const HASH_OFF: usize = 160;
+const IO_SIZE: usize = 224;
+
+static mut IO: [u8; IO_SIZE] = [0u8; IO_SIZE];
+
+/// Pointer to the shared I/O buffer in WASM linear memory.
+#[no_mangle]
+pub extern "C" fn io_ptr() -> *mut u8 {
+    core::ptr::addr_of_mut!(IO) as *mut u8
+}
+
+/// Size of the shared I/O buffer, for JS-side sanity checks.
+#[no_mangle]
+pub extern "C" fn io_len() -> u32 {
+    IO_SIZE as u32
+}
+
+/// Grind up to `iters` nonces. Returns 1 if a share was found.
+#[no_mangle]
+pub extern "C" fn solve(iters: u32) -> u32 {
+    let io = unsafe { &mut *core::ptr::addr_of_mut!(IO) };
+
+    let mut header = [0u8; 32];
+    header.copy_from_slice(&io[HEADER_OFF..HEADER_OFF + 32]);
+    let mut nonce = U512::from_big_endian(&io[NONCE_OFF..NONCE_OFF + 64]);
+    let target = U512::from_big_endian(&io[TARGET_OFF..TARGET_OFF + 64]);
+
+    for _ in 0..iters {
+        let nonce_bytes = nonce.to_big_endian();
+        let hash = qpow_math::get_nonce_hash(header, nonce_bytes);
+        if hash < target {
+            io[NONCE_OFF..NONCE_OFF + 64].copy_from_slice(&nonce_bytes);
+            io[HASH_OFF..HASH_OFF + 64].copy_from_slice(&hash.to_big_endian());
+            return 1;
+        }
+        nonce = nonce.saturating_add(U512::one());
+    }
+
+    // Not found: persist progress so the next chunk resumes where we stopped.
+    io[NONCE_OFF..NONCE_OFF + 64].copy_from_slice(&nonce.to_big_endian());
+    0
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn solve_matches_pool_verification() {
+        // Difficulty 4 -> target = MAX/4; expect a solution within a few iters.
+        let header = [7u8; 32];
+        let difficulty = U512::from(4u64);
+        let target = U512::MAX / difficulty;
+
+        let io = unsafe { &mut *core::ptr::addr_of_mut!(IO) };
+        io[HEADER_OFF..HEADER_OFF + 32].copy_from_slice(&header);
+        let start = U512::from(1u64) << 64; // arbitrary "session range" start
+        io[NONCE_OFF..NONCE_OFF + 64].copy_from_slice(&start.to_big_endian());
+        io[TARGET_OFF..TARGET_OFF + 64].copy_from_slice(&target.to_big_endian());
+
+        let mut found = 0;
+        for _ in 0..64 {
+            found = solve(256);
+            if found == 1 {
+                break;
+            }
+        }
+        assert_eq!(found, 1, "should find a share at difficulty 4");
+
+        let nonce = U512::from_big_endian(&io[NONCE_OFF..NONCE_OFF + 64]);
+        let (valid, hash) = qpow_math::is_valid_nonce(header, nonce.to_big_endian(), difficulty);
+        assert!(
+            valid,
+            "pool-side verification must accept the solver's nonce"
+        );
+        let reported = U512::from_big_endian(&io[HASH_OFF..HASH_OFF + 64]);
+        assert_eq!(hash, reported);
+    }
+}
@@ -14,8 +14,8 @@
 #
 # Notes:
 # - CPUAffinity must be a subset of the service’s allowed cpuset (e.g., cgroups).
-# - The miner logs the detected cpuset mask at startup and, when metrics are enabled,
-#   exports a miner_effective_cpus gauge for dashboards.
+# - The miner counts CPUs via the process affinity mask and exports a
+#   miner_effective_cpus gauge on the always-on metrics endpoint.
 # - Adjust the CPU list and worker count to match your hardware and cpuset.
 #
 # Example below assumes an 8‑way logical CPU system (IDs 0‑7).
@@ -40,9 +40,8 @@ IOSchedulingPriority=4
 # Default time-sharing policy is fine for CPU-bound workloads when pinned.
 CPUSchedulingPolicy=other

-# Optional: pin worker count and enable metrics (uncomment and adjust as needed).
-# Environment=MINER_ENGINE=cpu-fast
-# Environment=MINER_WORKERS=4
+# Optional: pin worker count and change the metrics port (uncomment and adjust as needed).
+# Environment=MINER_CPU_WORKERS=4
 # Environment=MINER_METRICS_PORT=9900
 # Environment=RUST_LOG=info,miner=debug

@@ -15,11 +15,11 @@
 # - CPUAffinity MUST be a subset of the service’s allowed cpuset (cgroup v2:
 #     /sys/fs/cgroup/cpuset.cpus.effective). If your deployment already constrains
 #   the miner with a cpuset, you can omit CPUAffinity here to inherit that mask.
-# - The miner logs the detected cpuset mask at startup (debug) and, when metrics
-#   are enabled, exports miner_effective_cpus for dashboards.
-# - Set MINER_WORKERS to match the number of CPUs in your affinity mask if you want
-#   full utilization; otherwise the miner default (50% of effective CPUs) is used
-#   when --workers is omitted.
+# - The miner counts CPUs via the process affinity mask and exports
+#   miner_effective_cpus on the always-on metrics endpoint.
+# - Set MINER_CPU_WORKERS to match the number of CPUs in your affinity mask if you
+#   want full utilization; otherwise the miner auto-detects when --cpu-workers is
+#   omitted.

 [Service]
 # Bind miner to ALL logical CPUs intended for mining. EDIT this to match your host.
@@ -46,10 +46,9 @@ IOSchedulingPriority=0
 TasksMax=infinity
 MemoryMax=infinity

-# Optional: pin worker count and enable metrics (uncomment and adjust).
-# Ensure MINER_WORKERS matches the size of CPUAffinity (or the cpuset mask) to use all CPUs.
-# Environment=MINER_ENGINE=cpu-fast
-# Environment=MINER_WORKERS=16
+# Optional: pin worker count and change the metrics port (uncomment and adjust).
+# Ensure MINER_CPU_WORKERS matches the size of CPUAffinity (or the cpuset mask) to use all CPUs.
+# Environment=MINER_CPU_WORKERS=16
 # Environment=MINER_METRICS_PORT=9900
 # Environment=RUST_LOG=info,miner=debug

@@ -64,5 +63,5 @@ MemoryMax=infinity
 #   and compare with:
 #     cat /sys/fs/cgroup/cpuset.cpus.effective
 #
-# - The service logs the cpuset mask at startup and the metrics exporter (if enabled)
-#   exposes miner_effective_cpus for observability.
+# - The always-on metrics exporter exposes miner_effective_cpus for
+#   observability of the detected CPU count (MINER_METRICS_PORT changes the port).
@@ -15,14 +15,24 @@
 #
 # Notes:
 #   - The miner CLI supports environment variables for all flags (see README):
-#       MINER_PORT, MINER_METRICS_PORT, MINER_WORKERS, MINER_ENGINE,
-#       MINER_PROGRESS_CHUNK_MS,
-#       MINER_MANIP_SOLVED_BLOCKS, MINER_MANIP_BASE_DELAY_NS,
-#       MINER_MANIP_STEP_BATCH, MINER_MANIP_THROTTLE_CAP.
+#       MINER_NODE_ADDR, MINER_AUTH_TOKEN_FILE (or MINER_AUTH_TOKEN),
+#       MINER_TLS_CERT_SHA256_FILE (or MINER_TLS_CERT_SHA256),
+#       MINER_CPU_WORKERS, MINER_GPU_DEVICES, MINER_GPU_BATCH_SIZE,
+#       MINER_CPU_BATCH_SIZE, MINER_GPU_THROTTLE_MS, MINER_METRICS_PORT,
+#       MINER_ALLOW_INTEGRATED, MINER_VERBOSE.
 #     This unit relies on those env vars (set in /etc/default|/etc/sysconfig) so
 #     ExecStart can remain stable and simple.
-#   - The service logs the detected cpuset mask (if any) and emits a
-#     miner_effective_cpus metric when metrics are enabled.
+#   - The node requires miner auth: MINER_AUTH_TOKEN_FILE and
+#     MINER_TLS_CERT_SHA256_FILE (or their inline variants) are mandatory —
+#     the miner exits at startup without them. Copy the node's
+#     miner-auth-token and miner-tls-cert-sha256 files somewhere this service
+#     can read (e.g. /etc/quantus-miner/): ProtectHome=true below blocks
+#     paths under /home and /root, where the node's base path usually lives.
+#   - The miner counts CPUs via the process affinity mask (so CPUAffinity=
+#     is respected) and emits a miner_effective_cpus metric.
+#   - The Prometheus exporter is ALWAYS on and binds plaintext HTTP on
+#     0.0.0.0:9900 by default; MINER_METRICS_PORT only changes the port.
+#     Firewall the port if the host is reachable from untrusted networks.

 [Unit]
 Description=Quantus External Miner Service
@@ -53,7 +63,9 @@ Environment=RUST_LOG=info

 # Launch the miner. All configuration should be passed via environment vars
 # (preferred) or via EXTRA_MINER_FLAGS in the env file for additional arguments.
-ExecStart=/usr/local/bin/quantus-miner $EXTRA_MINER_FLAGS
+# MINER_AUTH_TOKEN_FILE and MINER_TLS_CERT_SHA256_FILE must be set in the env
+# file (see readme.md) or the miner exits immediately.
+ExecStart=/usr/local/bin/quantus-miner serve $EXTRA_MINER_FLAGS

 # Graceful shutdown and restart behavior
 Restart=always
@@ -18,6 +18,18 @@ Prerequisites
 - The quantus-miner binary installed at /usr/local/bin/quantus-miner (or adjust ExecStart).
 - A service account (recommended):
  - sudo useradd --system --no-create-home --shell /usr/sbin/nologin quantus
+- The node's miner auth token and TLS cert fingerprint (required since the node
+  authenticates miners). The node creates both on first start with
+  --miner-listen-port and logs their paths (default:
+  `<base-path>/chains/<chain>/miner-auth-token` and
+  `<base-path>/chains/<chain>/miner-tls-cert-sha256`). Copy them where the
+  service can read them — the unit sets ProtectHome=true, so paths under /home
+  or /root are not readable:
+  - sudo install -d -m 0755 /etc/quantus-miner
+  - sudo install -m 0640 -o root -g quantus \
+      "<base-path>/chains/<chain>/miner-auth-token" /etc/quantus-miner/miner-auth-token
+  - sudo install -m 0644 \
+      "<base-path>/chains/<chain>/miner-tls-cert-sha256" /etc/quantus-miner/miner-tls-cert-sha256

 Install (unit)
 1) Copy the service file
@@ -26,21 +38,22 @@ Install (unit)
 2) Create a writable working directory (managed by systemd via StateDirectory)
   sudo install -d -o quantus -g quantus /var/lib/quantus-miner

-3) (Optional) Provide environment variables
+3) Provide environment variables (auth vars are REQUIRED)
   - Debian/Ubuntu:   sudoedit /etc/default/quantus-miner
   - RHEL/CentOS/Fed: sudoedit /etc/sysconfig/quantus-miner

-   Common variables (examples):
-   MINER_ENGINE=cpu
-   MINER_PORT=9833
-   MINER_METRICS_PORT=9900         # enable Prometheus exporter
-   MINER_WORKERS=4                 # leave unset to use default (50% of effective CPUs)
-   MINER_PROGRESS_CHUNK_MS=2000
-   # Throttling engine (cpu-chain-manipulator) knobs:
-   # MINER_MANIP_SOLVED_BLOCKS=0
-   # MINER_MANIP_BASE_DELAY_NS=500000
-   # MINER_MANIP_STEP_BATCH=10000
-   # MINER_MANIP_THROTTLE_CAP=0
+   Common variables (examples). Note: systemd EnvironmentFile= does not strip
+   inline comments — keep each assignment on its own line with nothing after
+   the value.
+   MINER_NODE_ADDR=127.0.0.1:9833
+   # Required: node auth token + TLS cert pin (see Prerequisites)
+   MINER_AUTH_TOKEN_FILE=/etc/quantus-miner/miner-auth-token
+   MINER_TLS_CERT_SHA256_FILE=/etc/quantus-miner/miner-tls-cert-sha256
+   # CPU worker threads; leave unset to auto-detect (~50% of available CPUs)
+   MINER_CPU_WORKERS=4
+   MINER_GPU_DEVICES=0
+   # Prometheus exporter port (the exporter is always on; this only changes the port)
+   MINER_METRICS_PORT=9900
   # Extra CLI flags (kept stable ExecStart):
   # EXTRA_MINER_FLAGS="--some-future-flag value"

@@ -65,31 +78,44 @@ Dedicated hardware (miner only)
  sudo systemctl restart quantus-miner

 Configuration reference (environment variables)
- MINER_ENGINE
-  - cpu (default), cpu-chain-manipulator
-  - gpu for high-performance GPU mining
- MINER_PORT
-  - HTTP API port (default 9833)
+- MINER_NODE_ADDR
+  - Address of the node's miner QUIC endpoint (default 127.0.0.1:9833).
+- MINER_AUTH_TOKEN_FILE (required, or MINER_AUTH_TOKEN inline)
+  - Path to a copy of the node's miner-auth-token file. The file variant is
+    preferred so the secret stays off the command line and out of `systemctl show`.
+- MINER_TLS_CERT_SHA256_FILE (required, or MINER_TLS_CERT_SHA256 inline)
+  - Path to a copy of the node's miner-tls-cert-sha256 file (the fingerprint is
+    also printed in the node's startup logs).
+- MINER_CPU_WORKERS
+  - CPU worker threads. If unset, auto-detected.
+- MINER_GPU_DEVICES
+  - Number of GPU devices to use. If unset, auto-detected.
+- MINER_GPU_BATCH_SIZE / MINER_CPU_BATCH_SIZE
+  - Nonces/hashes per cancellation check (defaults 1000000 / 10000).
+- MINER_GPU_THROTTLE_MS
+  - Delay between GPU batches in milliseconds (default 0 = no throttle).
 - MINER_METRICS_PORT
-  - Enable Prometheus exporter when set (e.g., 9900). If unset, metrics exporter is disabled.
- MINER_WORKERS
-  - Worker threads (logical CPUs). If unset, defaults to ~50% of effective CPUs (clamped to [1, effective-1]).
- MINER_PROGRESS_CHUNK_MS
-  - Target milliseconds for per-thread progress updates (default 2000ms).
- Throttling engine (cpu-chain-manipulator) knobs
-  - MINER_MANIP_SOLVED_BLOCKS, MINER_MANIP_BASE_DELAY_NS, MINER_MANIP_STEP_BATCH, MINER_MANIP_THROTTLE_CAP
+  - Prometheus exporter port (default 9900). The exporter is ALWAYS on and
+    binds plaintext HTTP on all interfaces (0.0.0.0); this variable only
+    changes the port — there is no disable or loopback-only option. Firewall
+    the port or restrict it to your monitoring network.
+- MINER_ALLOW_INTEGRATED
+  - Allow integrated GPUs even when discrete GPUs are present.
 - EXTRA_MINER_FLAGS
  - Optional extra CLI flags appended to ExecStart.

 CPU affinity, cpusets, and workers
- The miner detects the effective CPU capacity (logical CPUs) visible to the process by preferring cgroup v2 cpuset (cpuset.cpus.effective), falling back to v1, else using all logical CPUs.
- At startup (debug level), the miner logs the detected cpuset mask (if any).
- A Prometheus gauge miner_effective_cpus is emitted (when metrics are enabled) with the effective count for dashboards/alerts.
- If --workers (or MINER_WORKERS) exceeds effective CPUs, it is clamped and a warning is logged.
- If omitted, the miner defaults to ~50% of effective CPUs (but always at least 1 and less than or equal to effective-1).
+- The miner counts available CPUs via the process CPU affinity mask (num_cpus),
+  so a systemd CPUAffinity= setting (or a cgroup cpuset) is reflected in the count.
+- If MINER_CPU_WORKERS is unset, the miner uses ~50% of the detected CPUs
+  (at least 1) and logs the choice at startup ("Auto-detected N CPU workers").
+- An explicit MINER_CPU_WORKERS value is used as-is — it is NOT clamped to the
+  affinity mask, so an oversized value oversubscribes the pinned CPUs.
+- A Prometheus gauge miner_effective_cpus is emitted with the detected count for dashboards/alerts.
 - When pinning CPUAffinity at the systemd level:
  - Ensure CPUAffinity is a subset of the cgroup cpuset mask.
-  - Consider setting MINER_WORKERS to match the number of CPUs in the affinity mask if you want full utilization, or rely on the default 50% policy.
+  - Set MINER_CPU_WORKERS to the number of CPUs in the affinity mask for full
+    utilization, or leave it unset for the ~50% default.

 Security hardening (in the unit)
 - NoNewPrivileges=true
@@ -101,6 +127,9 @@ Security hardening (in the unit)
 - RestrictSUIDSGID=true
 - SystemCallFilter=@system-service
 Adjust or relax as needed for your environment.
+Note: the Prometheus exporter always listens on 0.0.0.0:<MINER_METRICS_PORT>
+(default 9900) — the unit cannot disable it or bind it to loopback. Firewall
+the port if the host is reachable from untrusted networks.

 Validation and troubleshooting
 - Check service status and logs:
@@ -110,20 +139,28 @@ Validation and troubleshooting
  taskset -cp "$pid"
 - Verify cpuset mask (cgroup v2):
  cat /sys/fs/cgroup/cpuset.cpus.effective
- Metrics:
-  - If MINER_METRICS_PORT is set, curl http://127.0.0.1:<port>/metrics
+- Metrics (always on, default port 9900):
+  - curl http://127.0.0.1:${MINER_METRICS_PORT:-9900}/metrics
  - Look for miner_effective_cpus and per-job/thread metrics.
 - Common pitfalls:
  - ExecStart path wrong (ensure /usr/local/bin/quantus-miner exists and is executable).
  - Service user/group missing (create quantus or adjust User/Group).
+  - Miner exits immediately with "miner auth token required" / "TLS cert fingerprint required":
+    set MINER_AUTH_TOKEN_FILE and MINER_TLS_CERT_SHA256_FILE in the env file (see Prerequisites).
+  - Auth/TLS files unreadable: ProtectHome=true blocks /home and /root; copy the
+    files to /etc/quantus-miner/ and make them readable by the quantus user.
+  - Node rejects the miner after the node's base path changed or its credential
+    files (miner-auth-token, miner-tls-cert.der/-key.der) were deleted and
+    regenerated: re-copy both files and restart. (A plain purge-chain does NOT
+    rotate them — it removes only the database, so the same token and cert are
+    reloaded on the next start.)
  - CPUAffinity not a subset of the cgroup cpuset (adjust cpuset or affinity).
-  - MINER_ENGINE set to gpu-* (currently unimplemented, exits with clear error).
  - Insufficient permissions to write WorkingDirectory (systemd StateDirectory creates /var/lib/quantus-miner with correct ownership).

 Operational tips
- For shared machines: prefer 10-shared-hardware.conf and leave MINER_WORKERS unset (defaults to ~50%).
- For dedicated machines: use 20-dedicated-hardware.conf and set MINER_WORKERS to the number of CPUs in CPUAffinity (or omit CPUAffinity to inherit cpuset).
- Use RUST_LOG=info,miner=debug temporarily to verify startup detection (cpuset mask, effective CPUs) and to observe mining loop behavior; then turn back down to reduce log volume.
+- For shared machines: prefer 10-shared-hardware.conf and leave MINER_CPU_WORKERS unset (auto-detect).
+- For dedicated machines: use 20-dedicated-hardware.conf and set MINER_CPU_WORKERS to the number of CPUs in CPUAffinity (or omit CPUAffinity to inherit cpuset).
+- Use RUST_LOG=info,miner=debug temporarily to verify startup detection (worker auto-detection, GPU discovery) and to observe mining loop behavior; then turn back down to reduce log volume.

 Support
 - Repository: https://github.com/Quantus-Network/quantus-miner
@@ -1,4 +1,4 @@
 [toolchain]
-channel = "stable"
+channel = "1.93.0"
 components = ["clippy", "rustfmt"]
 profile = "minimal"
@@ -1,207 +0,0 @@
-use primitive_types::U512;
-use quantus_miner::*;
-use resonance_miner_api::*;
-use warp::test::request;
-use warp::Filter; // Import shared API types
-
-#[tokio::test]
-async fn test_mine_endpoint() {
-    let state = MiningState::new();
-    let state_clone = state.clone();
-    let state_filter = warp::any().map(move || state_clone.clone());
-
-    let mine_route = warp::post()
-        .and(warp::path("mine"))
-        .and(warp::body::json())
-        .and(state_filter.clone())
-        .and_then(handle_mine_request);
-
-    // Test valid request
-    let valid_request = MiningRequest {
-        job_id: "test".to_string(),
-        mining_hash: "a".repeat(64),
-        distance_threshold: "1000".to_string(),
-        nonce_start: "0".repeat(128),
-        nonce_end: "1".repeat(128),
-    };
-
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&valid_request)
-        .reply(&mine_route)
-        .await;
-
-    assert_eq!(resp.status(), 200);
-    let body: MiningResponse = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::Accepted);
-    assert_eq!(body.job_id, "test");
-
-    // Test duplicate job ID
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&valid_request)
-        .reply(&mine_route)
-        .await;
-
-    assert_eq!(resp.status(), 409);
-    let body: MiningResponse = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::Error);
-    assert!(body.message.is_some());
-    assert!(body.message.unwrap().contains("Job already exists"));
-
-    // Test invalid request
-    let invalid_request = MiningRequest {
-        job_id: "".to_string(), // Empty job ID
-        mining_hash: "a".repeat(64),
-        distance_threshold: "1000".to_string(),
-        nonce_start: "0".repeat(128),
-        nonce_end: "1".repeat(128),
-    };
-
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&invalid_request)
-        .reply(&mine_route)
-        .await;
-
-    assert_eq!(resp.status(), 400);
-}
-
-#[tokio::test]
-async fn test_result_endpoint() {
-    let state = MiningState::new();
-    let state_clone = state.clone();
-    let state_filter = warp::any().map(move || state_clone.clone());
-
-    // First create a job
-    let job = MiningJob::new([0; 32], U512::from(1000), U512::from(0), U512::from(1000));
-    state.add_job("test".to_string(), job).await.unwrap();
-
-    let result_route = warp::get()
-        .and(warp::path("result"))
-        .and(warp::path::param())
-        .and(state_filter.clone())
-        .and_then(handle_result_request);
-
-    // Test existing job
-    let resp = request()
-        .method("GET")
-        .path("/result/test")
-        .reply(&result_route)
-        .await;
-
-    assert_eq!(resp.status(), 200);
-    let body: MiningResult = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::Running);
-    assert_eq!(body.job_id, "test");
-
-    // Test non-existent job
-    let resp = request()
-        .method("GET")
-        .path("/result/nonexistent")
-        .reply(&result_route)
-        .await;
-
-    assert_eq!(resp.status(), 404);
-    let body: MiningResult = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::NotFound);
-}
-
-#[tokio::test]
-async fn test_cancel_endpoint() {
-    let state = MiningState::new();
-    let state_clone = state.clone();
-    let state_filter = warp::any().map(move || state_clone.clone());
-
-    // First create a job
-    let job = MiningJob::new([0; 32], U512::from(1000), U512::from(0), U512::from(1000));
-    state.add_job("test".to_string(), job).await.unwrap();
-
-    let cancel_route = warp::post()
-        .and(warp::path("cancel"))
-        .and(warp::path::param())
-        .and(state_filter.clone())
-        .and_then(handle_cancel_request);
-
-    // Test cancel existing job
-    let resp = request()
-        .method("POST")
-        .path("/cancel/test")
-        .reply(&cancel_route)
-        .await;
-
-    assert_eq!(resp.status(), 200);
-    let body: MiningResponse = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::Cancelled);
-    assert_eq!(body.job_id, "test");
-
-    // Test cancel non-existent job
-    let resp = request()
-        .method("POST")
-        .path("/cancel/nonexistent")
-        .reply(&cancel_route)
-        .await;
-
-    assert_eq!(resp.status(), 404);
-    let body: MiningResponse = serde_json::from_slice(resp.body()).unwrap();
-    assert_eq!(body.status, ApiResponseStatus::NotFound);
-}
-
-#[tokio::test]
-async fn test_concurrent_access() {
-    let state = MiningState::new();
-    let state_clone = state.clone();
-    let state_filter = warp::any().map(move || state_clone.clone());
-
-    // Create multiple jobs concurrently
-    let mut handles = vec![];
-    for i in 0..10 {
-        let state = state.clone();
-        let handle = tokio::spawn(async move {
-            let job = MiningJob::new([0; 32], U512::from(1000), U512::from(0), U512::from(1000));
-            state.add_job(format!("test{}", i), job).await
-        });
-        handles.push(handle);
-    }
-
-    // Wait for all jobs to be created
-    for handle in handles {
-        assert!(handle.await.unwrap().is_ok());
-    }
-
-    // Verify all jobs exist
-    for i in 0..10 {
-        assert!(state.get_job(&format!("test{}", i)).await.is_some());
-    }
-
-    // Test concurrent result checks
-    let result_route = warp::get()
-        .and(warp::path("result"))
-        .and(warp::path::param())
-        .and(state_filter.clone())
-        .and_then(handle_result_request);
-
-    let mut result_handles = vec![];
-    for i in 0..10 {
-        let route = result_route.clone();
-        let handle = tokio::spawn(async move {
-            request()
-                .method("GET")
-                .path(&format!("/result/test{}", i))
-                .reply(&route)
-                .await
-        });
-        result_handles.push(handle);
-    }
-
-    // Verify all result checks succeed
-    for handle in result_handles {
-        let resp = handle.await.unwrap();
-        assert_eq!(resp.status(), 200);
-        let body: MiningResult = serde_json::from_slice(resp.body()).unwrap();
-        assert_eq!(body.status, ApiResponseStatus::Running);
-    }
-}
@@ -1,310 +0,0 @@
-// Verifying that our API spec is in sync with the implementation
-// Manually generated from the OpenAPI spec but it should still flag when we get out of sync and need to update.
-
-use primitive_types::U512;
-use quantus_miner::*;
-use resonance_miner_api::*; // Use the shared API types
-use serde_json::json;
-use warp::test::request;
-use warp::Filter; // For sending custom/invalid JSON
-
-// Helper function to setup the routes with a fresh state for each test
-fn setup_routes() -> (
-    MiningState,
-    impl Filter<Extract = impl warp::Reply, Error = warp::Rejection> + Clone,
-) {
-    let state = MiningState::new();
-    let state_clone = state.clone();
-    let state_filter = warp::any().map(move || state_clone.clone());
-
-    let mine_route = warp::post()
-        .and(warp::path("mine"))
-        .and(warp::body::json::<MiningRequest>()) // Expect valid MiningRequest
-        .and(state_filter.clone())
-        .and_then(handle_mine_request);
-
-    let result_route = warp::get()
-        .and(warp::path("result"))
-        .and(warp::path::param::<String>())
-        .and(state_filter.clone())
-        .and_then(handle_result_request);
-
-    let cancel_route = warp::post()
-        .and(warp::path("cancel"))
-        .and(warp::path::param::<String>())
-        .and(state_filter.clone())
-        .and_then(handle_cancel_request);
-
-    let routes = mine_route.or(result_route).or(cancel_route);
-    (state, routes)
-}
-
-// --- /mine Endpoint Tests ---
-
-#[tokio::test]
-async fn spec_mine_valid_request() {
-    let (_state, routes) = setup_routes();
-    let valid_req = MiningRequest {
-        job_id: "job-valid-1".to_string(),
-        mining_hash: "a".repeat(64),
-        distance_threshold: "12345".to_string(),
-        nonce_start: "0".repeat(128),
-        nonce_end: "f".repeat(128),
-    };
-
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&valid_req)
-        .reply(&routes)
-        .await;
-
-    // Spec: 200 OK, response matches MiningResponseAccepted
-    assert_eq!(
-        resp.status(),
-        200,
-        "Spec requires 200 OK for valid /mine request"
-    );
-    let body: MiningResponse = serde_json::from_slice(resp.body())
-        .expect("Valid response should deserialize to MiningResponse");
-    assert_eq!(
-        body.status,
-        ApiResponseStatus::Accepted,
-        "Spec requires status 'accepted'"
-    );
-    assert_eq!(body.job_id, valid_req.job_id);
-    assert!(
-        body.message.is_none(),
-        "Spec shows no message field for success"
-    );
-}
-
-#[tokio::test]
-async fn spec_mine_duplicate_job_id() {
-    let (state, routes) = setup_routes();
-    let valid_req = MiningRequest {
-        job_id: "job-duplicate-1".to_string(),
-        mining_hash: "b".repeat(64),
-        distance_threshold: "54321".to_string(),
-        nonce_start: "1".repeat(128),
-        nonce_end: "e".repeat(128),
-    };
-
-    // Add the job first
-    let job = MiningJob::new(
-        hex::decode(valid_req.mining_hash.clone())
-            .unwrap()
-            .try_into()
-            .unwrap(),
-        U512::from_dec_str(&valid_req.distance_threshold).unwrap(),
-        U512::from_str_radix(&valid_req.nonce_start, 16).unwrap(),
-        U512::from_str_radix(&valid_req.nonce_end, 16).unwrap(),
-    );
-    state.add_job(valid_req.job_id.clone(), job).await.unwrap();
-
-    // Try to submit again
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&valid_req)
-        .reply(&routes)
-        .await;
-
-    // Spec: 409 Conflict, response matches MiningResponseError
-    assert_eq!(
-        resp.status(),
-        409,
-        "Spec requires 409 Conflict for duplicate job ID"
-    );
-    let body: MiningResponse = serde_json::from_slice(resp.body())
-        .expect("Valid error response should deserialize to MiningResponse");
-    assert_eq!(
-        body.status,
-        ApiResponseStatus::Error,
-        "Spec requires status 'error'"
-    );
-    assert!(
-        body.message.is_some(),
-        "Spec requires message field for error"
-    );
-    assert!(body.message.unwrap().contains("Job already exists"));
-    assert_eq!(body.job_id, valid_req.job_id);
-}
-
-// Test cases for invalid formats (hash length, nonce length, difficulty non-numeric)
-// These should ideally result in 400 Bad Request
-#[tokio::test]
-async fn spec_mine_invalid_hash_format() {
-    let (_state, routes) = setup_routes();
-    let invalid_req = json!({
-        "job_id": "job-invalid-hash",
-        "mining_hash": "a".repeat(63), // Too short
-        "distance_threshold": "1000",
-        "nonce_start": "0".repeat(128),
-        "nonce_end": "f".repeat(128)
-    });
-
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&invalid_req)
-        .reply(&routes)
-        .await;
-
-    // Spec: 400 Bad Request, response matches MiningResponseError
-    assert_eq!(
-        resp.status(),
-        400,
-        "Spec requires 400 Bad Request for invalid hash format"
-    );
-    let body: MiningResponse =
-        serde_json::from_slice(resp.body()).expect("Error response should deserialize");
-    assert_eq!(body.status, ApiResponseStatus::Error);
-    assert!(body.message.is_some());
-    assert!(body
-        .message
-        .unwrap()
-        .contains("mining_hash must be 64 hex characters"));
-}
-
-// Add similar tests for invalid nonce_start, nonce_end, difficulty format...
-
-#[tokio::test]
-async fn spec_mine_missing_required_field() {
-    let (_state, routes) = setup_routes();
-    // Missing 'difficulty'
-    let invalid_req = json!({
-        "job_id": "job-missing-field",
-        "mining_hash": "c".repeat(64),
-        "nonce_start": "0".repeat(128),
-        "nonce_end": "f".repeat(128)
-    });
-
-    let resp = request()
-        .method("POST")
-        .path("/mine")
-        .json(&invalid_req)
-        .reply(&routes)
-        .await;
-
-    // Warp's default JSON body deserialization handles missing fields
-    // It should return 400 Bad Request based on warp::body::json()
-    assert_eq!(
-        resp.status(),
-        400,
-        "Spec requires 400 Bad Request for missing required field"
-    );
-    // The body might be a plain text error from Warp, not necessarily our JSON structure
-    // We can check if the body indicates a deserialization error
-    let body_bytes = resp.body();
-    let body_string = String::from_utf8_lossy(body_bytes);
-    assert!(
-        body_string.contains("Failed to deserialize") || body_string.contains("missing field"),
-        "Response body should indicate deserialization failure"
-    );
-}
-
-// --- /result/{job_id} Endpoint Tests ---
-
-#[tokio::test]
-async fn spec_result_job_running() {
-    let (state, routes) = setup_routes();
-    let job_id = "job-running-1".to_string();
-    // Add a job
-    let job = MiningJob::new([1u8; 32], U512::from(1000), U512::from(0), U512::from(1000));
-    state.add_job(job_id.clone(), job).await.unwrap();
-
-    let resp = request()
-        .method("GET")
-        .path(&format!("/result/{}", job_id))
-        .reply(&routes)
-        .await;
-
-    // Spec: 200 OK, response matches MiningResult with status 'running'
-    assert_eq!(resp.status(), 200);
-    let body: MiningResult =
-        serde_json::from_slice(resp.body()).expect("Should deserialize to MiningResult");
-    assert_eq!(body.status, ApiResponseStatus::Running);
-    assert_eq!(body.job_id, job_id);
-    assert!(body.nonce.is_none()); // Should be none initially
-    assert!(body.work.is_none()); // No work yet
-                                  // hash_count and elapsed_time should be present
-                                  // assert_eq!(body.hash_count, 0); // Initial state
-                                  // assert!(body.elapsed_time >= 0.0);
-}
-
-// Add tests for spec_result_job_completed, spec_result_job_failed (need to manipulate state)
-// Need to carefully simulate a completed/failed state or run miner loop briefly
-
-#[tokio::test]
-async fn spec_result_job_not_found() {
-    let (_state, routes) = setup_routes();
-    let job_id = "job-not-found-1";
-
-    let resp = request()
-        .method("GET")
-        .path(&format!("/result/{}", job_id))
-        .reply(&routes)
-        .await;
-
-    // Spec: 404 Not Found, response matches MiningResultNotFound
-    assert_eq!(resp.status(), 404);
-    let body: MiningResult = serde_json::from_slice(resp.body())
-        .expect("Should deserialize to MiningResult (NotFound variant)");
-    assert_eq!(body.status, ApiResponseStatus::NotFound);
-    assert_eq!(body.job_id, job_id);
-    assert!(body.nonce.is_none());
-    assert!(body.work.is_none());
-    assert_eq!(body.hash_count, 0);
-    assert_eq!(body.elapsed_time, 0.0);
-}
-
-// --- /cancel/{job_id} Endpoint Tests ---
-
-#[tokio::test]
-async fn spec_cancel_existing_job() {
-    let (state, routes) = setup_routes();
-    let job_id = "job-cancel-1".to_string();
-    let job = MiningJob::new([2u8; 32], U512::from(1000), U512::from(0), U512::from(1000));
-    state.add_job(job_id.clone(), job).await.unwrap();
-
-    let resp = request()
-        .method("POST")
-        .path(&format!("/cancel/{}", job_id))
-        .reply(&routes)
-        .await;
-
-    // Spec: 200 OK, response matches MiningResponseCancelled
-    assert_eq!(resp.status(), 200);
-    let body: MiningResponse =
-        serde_json::from_slice(resp.body()).expect("Should deserialize to MiningResponse");
-    assert_eq!(body.status, ApiResponseStatus::Cancelled);
-    assert_eq!(body.job_id, job_id);
-    assert!(body.message.is_none());
-
-    // Verify job is actually cancelled
-    let job = state.get_job(&job_id).await.unwrap();
-    assert_eq!(job.status, JobStatus::Cancelled);
-}
-
-#[tokio::test]
-async fn spec_cancel_non_existent_job() {
-    let (_state, routes) = setup_routes();
-    let job_id = "job-cancel-not-found-1";
-
-    let resp = request()
-        .method("POST")
-        .path(&format!("/cancel/{}", job_id))
-        .reply(&routes)
-        .await;
-
-    // Spec: 404 Not Found, response matches MiningResponseNotFound
-    assert_eq!(resp.status(), 404);
-    let body: MiningResponse = serde_json::from_slice(resp.body())
-        .expect("Should deserialize to MiningResponse (NotFound variant)");
-    assert_eq!(body.status, ApiResponseStatus::NotFound);
-    assert_eq!(body.job_id, job_id);
-    assert!(body.message.is_some());
-}
-
-// Removed placeholder test