[JacobStephens2]

Summary

Prepare both Arms before Trials, caching thirdshift by commit and recording each Harness's real path and version.

prepare-arms
  resolve each Harness once → real path + version
  baseline / candidate commit
    cached → reuse builds/<commit>/thirdshift
    missing → detached worktree → cargo build --release --locked → cache
  atomically publish both Arm records

Adds thirdshift evals prepare-arms and documents the record and cache layout. Git replacement objects cannot change the requested build. Trial execution and Experiment scheduling remain in #1001 and #1003; this change supplies their preparation inputs.

Closes #1002

Evidence

• Before: seeded_arm_commits_are_recorded_without_building failed because the preparation command was absent. A stderr-only Harness version failed as empty; a Git replacement built the candidate twice.
  After: cargo test --test eval_arms passes all 7 tests, including those regressions, failed-build cleanup, concurrent cache reuse, distinct Harness versions and a link swapped between Arm launches.
• cargo clippy --all-targets -- -D warnings and cargo fmt --check pass.
• /usr/bin/time -p scripts/affected-tests origin/issue-977 passes: 43 site tests and 2,392 Rust tests passed, with 56 Rust tests skipped by the repository configuration. Total wall time was 1,595.25 s including compilation and site checks; Rust compilation took 1m 30s, and Nextest execution took 1,447.242 s (11 slow tests).
• CI is green for 0623ff0.

Test seam: the whole binary with real Git repositories, seeded test binaries, and fake Cargo/Harness executables. Assertions inspect JSON output, the saved Arm record and the executable at its recorded path. The link-swap test invokes that path directly; production Trial integration belongs to #1001/#1003. Tests compile no Arm builds.

Review coverage

Used thirdshift-code-review with issue-977 pinned to its fetched head cea6615b2e039e1521d6b0e62e6abfb4668c3c1a. Standards: 0 findings. Spec: 0 findings.

Both reviewers read all 8 changed files. Unread changed files: none. Generated exclusions: none; generated manual changes were inspected. Final reports and each axis's files-read list are saved in .thirdshift-review-jwdqiI/standards.md and .thirdshift-review-jwdqiI/spec.md in the worktree.

Unaddressed findings

Standards

None.

Spec

None.

Merge Danger

Door: two-way

Blast Radius: Evals

Adds build-cache entries and a new Arm record.

Built with codex · gpt-6.1-sol · xhigh