[tonyketcham]

Stacked on #296.

What

Phase 3 of the code-context benchmark. Two new tests:

1. Questions no single path names. behaviour.json (16) asks about behaviour spread across folders, with spread and control groups. multibranch.json (12) asks about many branches at once.

2. Code anchors a refactor left stale. build-repo.mjs --refactor adds 10 (1x) or 11 (10x) refactor commits on main without updating Proof:

   • two folder moves;
   • file renames, and symbol renames (evict_pages → reclaim_pages, warm_district → prewarm_district, upgrade_steps → migration_steps, hash_blocks → hash_state_blocks);
   • one module split and one merge;
   • one behaviour change: the hash interval goes from 30 to 60 frames, while the accepted Decision still says 30.

   refactor.json (19) asks about the moved code. Its groups are follow, no-follow, drift, location and control. The refactor map is written beside the keymap, outside the repository.

Details

• New test-bed checks: renamed, staleAnchor, codeAt, anchorDrift, goneText and refAll.
• check.mjs refuses the new sets on a bed without a 10x refactor map.
• run.mjs leaves out the 17 code and branch questions the refactor makes stale.
• Without --refactor, both existing beds rebuild byte for byte.

Checks

• check.mjs: ok, 270 questions; ok, 91 questions on the 10x refactor bed; ok, 44 questions on p2-bed-x10.
• Prettier is clean.

Offline baseline with today's reads

• Every --path query on moved code returns no governing records: 10 of 10.
• Four cards cite moved code, and none of them says the code moved.

The paid baseline comes next.

---

Generated by Claude Code