feat(cli): build local evals from workload traces

[aamir]

Summary

Developers can now turn a captured Understudy workload into a private, human-reviewable local eval draft with understudy evals build. The CLI freezes an exact cohort, downloads only bounded and hash-verified traces, reconstructs lineage and tool behavior locally, and generates the verifier environment without uploading customer payloads or calling a model provider.

This is the local-authoring half of the CLI-first eval workflow. It depends on understudy-platform#643 for exact create-response recovery; publishing verifier packages and hosted model/prompt sweeps remain follow-up work.

Design decisions

• Keep shared authority on the backend: tenant scope, redacted catalog, immutable cohort, bounded export.
• Keep eval semantics local: lineage, tool interpretation, task contracts, oracle, sentinels, and human review.
• Persist only a redacted checkpoint between retries; delete payload-bearing partial attempts.
• Fail closed on changed selection options, concurrent output ownership, stale locks, lineage mismatches, untrusted URLs, redirects, oversized payloads, and generated Python identifier collisions.
• Publish the final project directory only after frozen and compiled capture counts agree.

Privacy

The command writes private local files that may contain prompts, completions, and tool payloads. It performs no upload and calls no model provider.

Validation

• npm run build
• npm run typecheck
• Focused CLI workflow tests: 6 passed
• Final build/resume/concurrency regression tests: 2 passed after hardening
• Trace-foundry and materialization tests: 61 passed
• Public skill validation: 43 skills
• npm package dry-run smoke: passed

The repository-wide test command was not run because unrelated installer tests can mutate the operator's real ~/.understudy; the affected workflow suites were run directly.

---

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-30T01:17:44.574151Z	9acf42f	PR opened

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[devin-ai-integration]

Devin Review found 2 potential issues.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)