fix(review): preserve full-hour sessions in .github/workflows/opencode-review.yml

[seonghobae]

Outcome

Preserve legitimate long-running OpenCode review sessions and retain valid model work without weakening any review, check, or ruleset gate.

Current-head evidence

• DeepSeek V3 and DeepSeek R1-0528 were each terminated by the protected-base 600-second cap on the previous head.
• GPT-4.1 returned a 31-line review in 86 seconds, but the protected-base normalizer rejected only the redundant exact path:line citation in adversarial probe evidence.
• Every cadence class now allows 5,400 seconds per model because changed-file count is not a repository-complexity proxy.
• The model pool remains bounded at 11,700 seconds inside a 12,000-second watchdog and a 205-minute Actions step.
• The normalizer may restore only redundant path:line text when one receipt already matches the sealed current-head source line. Mismatched, duplicate, circular, malformed, or unobserved evidence remains rejected.

Verification

• Python: 622 passed
• Focused normalizer branch coverage: 100.00 percent, 105 passed
• Strix/OpenCode integration suite: PASS after approximately 40 minutes
• Ruff, actionlint syntax, bash syntax, and git diff check: PASS

No branch protection, approval count, required workflow, review identity, or security gate is relaxed.

[cwl-noema-review]

Noema PydanticAI review

All check and dependency gates pass, with evidence confirming that the PR's behavioral intent (to preserve full-hour model-pool review sessions in OpenCode workflows, robust R package-aware evidence, and reliable recovery from provider timeouts or HTTP 5xx failures) is implemented thoroughly. The diff delivers:

• Increased model pool and dynamic cadence session/budget timeouts (from 10–60min to 90–195min bounds) everywhere in .github/workflows/opencode-review.yml and run_opencode_review_model_pool.sh, matching approval/test expectations.
• Improved handling of paid session timeouts: after a hung or killed attempt, the runner attempts to recover persisted partial results via OpenCode session export by title, before falling through, tested and contract-guarded in test_opencode_model_pool_runner.py and test_opencode_agent_contract.py.
• Improvements for R package evidence and test handling (conditionally awaiting completed package checks, improved result reporting, new output bindings) in opencode-review.yml and the test gate (test_strix_quick_gate.sh confirms workflow behavior).
• Extra resilience and coverage in error classification, session export, logic to advance/skipping on provider failure signatures.
• All tests, SAST, dependency, and workflow checks are green; no regressions or behavioral divergences are reported by CI, contract, or regression tests.

All new/changed behavior is protected by tests. No blocking issue is evidenced. Approving.

Findings

• No blocking findings.

• Result: APPROVE

• Verdict: approve

• Confidence: high

• Head SHA: df408934021a206be6029838546c72094ca528ef

• Reviewer credential: noema-github-app

[seonghobae]

현재 main 기준으로 이 PR의 범위를 재확인했습니다.\n\n- 와 가 모든 cadence의 모델 실행 한도 5,400초, 총 budget 11,700초, Actions step 205분 계약을 이미 포함합니다.\n- 정상화기의 검증된 복구 변경은 별도 PR #563이 독립적으로 보유합니다.\n- 이 PR은 기준점 에서 갈라져 현재 main과 충돌하며, 그대로 재기반하면 이미 반영된 장기 세션 변경과 #563 범위를 다시 합치는 중복 PR이 됩니다.\n\n기능 손실 없이 현재 main + #563으로 대체되었으므로 이 PR은 superseded 상태로 닫습니다. #563은 #578 병합 후 현재-head 증거를 다시 실행해 별도로 처리하겠습니다.

[seonghobae]

정정된 current-head 근거입니다.

• 현재 main은 8684f90286ff60234b1f78642e9c38a378245545입니다.
• 3e7e33c와 e56fab8가 모든 cadence의 모델 실행 한도 5,400초, 총 budget 11,700초, Actions step 205분 계약을 이미 포함합니다.
• 정상화기의 검증된 path:line 복구 변경은 별도 PR #563이 독립적으로 보유합니다.
• 이 PR은 기준점 d1406fc에서 갈라져 현재 main과 충돌하며, 그대로 재기반하면 이미 반영된 장기 세션 변경과 fix(review): bind verified evidence in scripts/ci/opencode_review_normalize_output.py #563 범위를 다시 합치는 중복 PR이 됩니다.

따라서 현재 main + #563으로 대체된 superseded PR로 닫은 상태를 유지합니다.