Add HTTPCore cancellation-poisoned pool eval case

[vishaltandale00]

Part of #278

Summary

• Adds the frozen HTTPCore 1.0.2 cancellation-poisoned connection-pool case at upstream commit 79fa6bf0dfcf3820d1ae7e52a2d268f33022c5a4 and tree 834aaf7041c78aa49597e691e6ce9fc41d6c0bc6.
• Adds the visible debugging task, immutable snapshot, exact Python and uv environment, materializer, sealed public-seam verifier, catalog/service integration, and deterministic tests.
• Uses evaluator-owned cancellation hooks and loopback-only servers to verify connection-slot release, subsequent-request success, repeated cancellation, cleanup, and the frozen upstream regression suite without prescribing a source patch.
• Adds the complete admission portfolio: untouched red baseline, two materially different green fixes, and seven adversarial mutants with independent predicates. No suite manifest or other case is included.

Verification

• RELAYER_HTTPCORE_CACHE=/private/tmp/relayer-httpcore-feasibility.NcnwtE/cache npm run eval:httpcore-admission — passed all 10 variants; sealed verifier digest sha256:db8e646df351986fea8700934db8b2787835376afefed4e08dd1feb10bb262d5.
• npx vitest run packages/eval-runner/test/httpcore-cancellation-case.test.ts test/eval-service-simulated-user.test.mjs test/eval-app-integration.test.mjs — 3 files, 16 tests passed.
• npm run build — passed on the final source snapshot.
• git diff --cached --check — passed before commit.
• npm run check — did not complete green on this storage-constrained host. Independent reruns passed each encountered unrelated timing boundary, but the final monolithic run stopped in existing macOS evidence tests with ENOSPC while copying the sealed Homebrew Node runtime and an intermittent undersized provider-video frame. The host had about 282 MiB free after evaluator-owned cleanup.

Adversarial review

Reviewer: adversarial_review. Commit: 63a55ad. Pre-commit staged digest: 51bcf5258f23d601a48c4fcc4346f7013a9b79e903dcd8af011119e1bdb60db6. Scope: complete HTTPCore case, normalized mutant patches, seals, timeout/grace behavior, flushed receipt handling, catalog/service integration, and focused tests. Verdict: PASS with no P0/P1/P2 findings.

Authority boundary

This is behavioral-verifier admission, not security-sandbox certification for arbitrarily hostile candidate code inside the Python worker. Host execution isolation and attacks on process internals remain explicitly out of scope and are recorded in the sealed receipt.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.