Resume incomplete strict teacher coverage

[bingran-you]

What changed

• Reuses completed jobs/teacher/attempt-* directories when strict teacher collection resumes.
• Recomputes selection across preserved attempts and runs only task IDs that still lack an eligible rollout.
• Restarts an interrupted attempt when its terminal health marker is missing.
• Treats an insufficient teacher manifest as incomplete work to continue, not as a completed state to reject.
• Allows one safe persisted-plan change on resume: increasing teacher.max_attempts. Decreasing it or changing dataset/model/reward/training semantics still fails closed.

Why

The merged-main red-wine full run completed a healthy deterministic 8/14 baseline and three strict 397B teacher attempts, then stopped at 11/16 eligible trajectories as configured. All existing baseline and teacher artifacts are valid and expensive, but the current resume path cannot continue them:

• collect_verified_teacher_rollouts always starts again at attempt 1.
• _collect_and_convert_teacher_data treats an insufficient manifest as a reusable completed stage, then rejects it.
• The saved-plan guard rejects increasing only the retry budget.

That makes strict all-task coverage non-resumable even though the next correct action is simply attempt 4 on the five missing task IDs.

Validation

• 222 package contract tests pass.
• 36 focused teacher/pipeline tests pass.
• Ruff check/format, Python compilation, and git diff --check pass.
• New coverage proves completed-attempt reuse, missing-task-only continuation, interrupted-attempt restart, insufficient-manifest recollection, safe retry-budget increase, and rejection of all other teacher recipe drift.
• Live preserved state contains three complete attempts, 11 selected training-ready reward-1 trajectories, and exactly five remaining task IDs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 03b4cacde6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Gate teacher-attempt reuse on resume

In a non-resume run that reuses an existing run name or jobs directory, this new scan treats any prior attempt-* directory with a health file as completed and skips straight to later attempts, because collect_verified_teacher_rollouts has no resume flag and is called from the normal collection path as well as resume. That can silently build the new SFT data from stale teacher rollouts from an earlier recipe/task snapshot instead of recollecting attempt 1; the reuse should only happen when the pipeline is actually resuming.

Useful? React with 👍 / 👎.