refactor(mcp): integrate evaluator with production service

.github/workflows/ci.yml

@@ -57,6 +57,21 @@ jobs:
       - uses: Swatinem/rust-cache@v2
       - run: cargo test --workspace
 
+  # Keep the frozen evaluator's contracts and executable scenario inventory
+  # checked independently of a candidate acceptance run. Provider scenarios
+  # remain in the separate candidate acceptance lane because this gate does
+  # not invoke a candidate or use host state/network services.
+  mcp-eval-contracts:
+    name: MCP evaluator contracts
+    needs: clippy
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+      - run: cargo test --locked -p icm-mcp-eval --no-fail-fast
+      - run: cargo run --locked -p icm-mcp-eval -- verify-design --suite-root crates/icm-mcp-eval
+
   # Validates the fully-static musl binary shipped for old-glibc distros
   # (issue #330: the glibc binaries need symbols newer than Debian Bookworm's
   # glibc 2.36). Embeddings are dropped because `ort`/onnxruntime ships only

CLAUDE.md

@@ -22,6 +22,12 @@ Créer une mémoire long-terme intelligente qui:
 
 ## Structure du workspace
 
+> Note de maintenance : le workspace actuel comprend `crates/icm-core`,
+> `crates/icm-store`, `crates/icm-mcp`, `crates/icm-cli`, et le harness de
+> développement `crates/icm-mcp-eval`. Ce dernier contient les contrats MCP
+> gelés et le runner hermétique ; voir son [README](crates/icm-mcp-eval/README.md).
+> Les phases et chemins historiques ci-dessous décrivent l'ancien prototype.
+
 ```
 icm/
 ├── Cargo.toml (workspace)

CONTRIBUTING.md

@@ -173,6 +173,30 @@ cargo fmt --all --check && cargo clippy --workspace --all-targets -- -D warnings
 
 ---
 
+## MCP evaluator changes
+
+Changes to MCP behavior should be exercised through the workspace evaluator
+when applicable. Keep synthetic fixtures, frozen contracts/goldens,
+normalization, metrics, and isolation assertions in `crates/icm-mcp-eval`;
+do not reimplement production protocol, framing, dispatch, or schemas there.
+Ordinary MCP scenarios use the production `icm-mcp` service in-process, while
+CLI, provider, proxy, HTTP, transport-edge, and process-isolation scenarios
+must remain candidate-process tests.
+
+Before opening a PR that changes the evaluator or MCP service, run:
+
+```bash
+cargo fmt --all -- --check
+cargo clippy --workspace --all-targets -- -D warnings
+cargo test --locked --offline -p icm-mcp-eval
+cargo run -p icm-mcp-eval --locked --offline -- verify-design \
+  --suite-root crates/icm-mcp-eval
+```
+
+Use a dedicated non-Git workspace for the two-root self-test. Never use real
+user state, credentials, provider configuration, network services, or a fixed
+port. See the evaluator README for candidate build and staging instructions.
+
 ## Questions?
 
 - **Bug reports & features**: [Issues](../../issues)