fix(judge): add bounded correction retry for agent judge

[DancingCircles]

Summary

• add exactly one validation-guided correction retry for strict agent_judge decode or semantic validation failures
• resume the Judge session when supported, with one serialized first-turn fallback message otherwise
• preserve every raw Judge response and isolate retry engine artifacts under outputs/judge/run/retry/
• aggregate both attempts into the final Judge session without changing report schemas
• add Judge, Evaluator, deterministic E2E, and credential-gated QoderCLI coverage

Relationship to #197

This builds on #197, which introduced the strict program-owned response contract and deterministic criterion mapping. This PR completes the maintainer-approved PR2 scope: one bounded correction retry and raw-output artifact preservation.

Closes #192

Behavior

• retry only after strict JSON decoding or semantic contract validation fails
• do not retry a valid FAIL judgment, explicit cancellation, context materialization failure, or an agent error without recoverable output
• share the existing Judge/case timeout budget across both attempts; no timeout reset or backoff
• use agent.SessionResumer.RunTurn when a session ID is available; otherwise call Run with one serialized user message containing the original prompt, first raw response, and correction instructions
• keep the last attempt authoritative while aggregating duration, token/turn usage, and stable-deduplicated artifacts
• keep native structured output, configurable retry policies, and report schema/UI changes out of scope

Artifact layout

outputs/judge/run/
├── stdout.json (or other first-attempt engine artifacts)
├── raw-response-attempt-1.txt
├── raw-response-attempt-2.txt       # only when correction runs
└── retry/
    └── stdout.json (or other retry engine artifacts)

Test plan

• go test -race ./internal/judge ./internal/evaluator ./internal/cli
• make fmt
• make verify
• go test -tags e2e -timeout 1200s -count=1 ./e2e
• go run ./cmd/skill-up debug judge examples/judge-debug-agent.json --output /tmp/grading.json
• verified the deterministic mock performs exactly initial → correction, returns PASS, and preserves both raw responses plus retry artifacts
• verified the QoderCLI ultimate test skips without full-E2E prerequisites; the trusted Model E2E runner will execute it when QoderCLI and credentials are available

Local Docker note

go test -race ./... passes for every package except the unchanged upstream test internal/runtime.TestNoneRuntime_ExecStaleLockDoesNotHangNextCommand, which reproducibly reports that a background child survives timeout in this Docker environment. internal/runtime has no diff from upstream/main; all other packages pass together under -race.

[zpzjzj]

Review summary: 0 Blockers, 2 Major findings, and 1 Info item. I recommend addressing the Major findings before merging.

Coverage: reviewed all 10/10 changed files, including the correction contract, agent adapters and callers, evaluator/report metrics, artifact handling, tests, fixtures, documentation, E2E paths, and CI-only coverage. All current CI checks are green.

Info: please use an English commit subject. Replace fix(judge): 增加有界纠错重试 with fix(judge): add bounded correction retry.

[DancingCircles]

Thanks for the thorough review. Both Major findings are addressed in 422bdaf, with focused adapter/aggregation regressions plus deterministic and full quick E2E coverage. The follow-up commit and PR title use English subjects. I kept the already-reviewed first commit intact rather than force-pushing rewritten history; the final PR title is fix(judge): add bounded correction retry for agent judge.

[zpzjzj]

Review summary: 0 Blockers, 2 Major findings, and 2 Info items. I recommend addressing the Major findings before merging.

Coverage: reviewed all 11/11 changed files at 422bdaf, including the correction contract, all built-in and custom agent paths, metric aggregation, artifact collection and archival, evaluator/report consumers, unit tests, fixtures, documentation, deterministic E2E, credential-gated paths, and current CI. make test, make verify, and the targeted correction E2E pass; all GitHub checks are green. The local full E2E suite reached and passed the new correction test, but the pre-existing QoderCLI live test failed because the successful external run reported zero token usage.

Info:

• Commit 7c9e1b1 still has a Chinese subject and body. Adding a later English commit and changing the PR title does not make the historical commit information English; please reword that commit.
• The Summary still says "deterministic three-message fallback", while the implementation and Behavior section now correctly describe one serialized user message. Please update the stale Summary text.

The two Major findings from the previous review have been fixed in 422bdaf; the comments below are distinct issues found during the complete re-review.

[DancingCircles]

Both new Major findings are addressed in aed0cdd, with a real evaluator-level CustomAgent regression covering fresh-turn accounting and separate framework input/output snapshots. The stale Summary wording has also been updated. Local verification: focused race tests pass, make verify reports 0 issues, all packages except the documented unchanged Docker-only internal/runtime case pass under race, and the full quick E2E suite passes.

For the remaining historical commit Info item: rewording 7c9e1b1 requires a non-fast-forward force push over the already-reviewed branch. I have not rewritten that remote history. If English-only individual commit history is mandatory, would you prefer a replacement clean PR branch; otherwise an English squash-merge commit using the current PR title preserves an English final history?

[zpzjzj]

LGTM