feat(bench): add a git test bed with code-context and branch questions

[tonyketcham]

Summary of changes

This is layer 13 of a stack. It sits on the --path/--refs PR below it.

It adds the benchmark for Phase 2: whether agents can find the records behind code, and see what other branches propose. It changes benchmark files only.

1. scripts/build-repo.mjs --out <dir> builds a git repository from the committed 1x graph and spec/repo.mjs. The build is deterministic: fixed dates, seeded ids and a fixed commit author. The repository holds:

   • a small fake codebase for the fictional studio;
   • commits with Proof: trailers;
   • code Citations and implementation notes written through the real Proof writer;
   • four engineers' branches with unmerged records, and draft PRs in .flatbread/proof/prs.json;
   • planned conflicts between branches: two accepted Decisions on one question, one record replaced twice, a record with a different state on two branches, a retraction, and a second compaction.

   The key map is written outside the repository.

2. Questions. questions/code.json and questions/branch.json have 22 questions each, in opt-in sets. They were written without access to the new reads, and are worded the way an engineer asks, for example "Before I edit …, which live Decision governs …?" or "Does any open branch propose replacing …?".

3. check.mjs --repo <dir> proves every answer with two new check types:

   • atRef runs any check against the graph as a branch has it.
   • codeAnchor and trailer prove which records govern a piece of code.

   check.mjs still prints ok: 270 questions, 498 records by default, and each new set prints ok: 22 questions.

4. run.mjs --repo <dir> gives each session its own clone of the test bed, with no key map, spec or questions. In that mode every arm's task text explains the clone and allows read-only git, so the no-brief arm can use git log, git show and git diff too. --prepare-only builds the folders and starts no session.

5. A new findcode strategy adds two lines to the find text: use --path before opening code, adding --refs all when a file may exist only on a branch, and use --refs all before building on a Decision. Without --repo, every existing text is byte-identical.

Offline check, joining this test bed with the reads below:

Set	Evidence in top 5	Answer on the cards
Code context	21 of 22	17 of 19
Branch	17 of 22	14 of 15 (counted on evidence cards only, since every --refs all reply names all branches)

A paid trial of findcode against no-brief comes next, in a results PR.

Please don't delete this checklist! Before submitting the PR, please make sure you do the following:

• I added doc comments to any new public exports, and inline comments to any hard-to-understand areas
• My changes generate no new console errors locally
• If applicable, try to include a test that fails without this PR but passes with it. These are benchmark scripts. check.mjs proves all 44 new answers, two builds give the same commit hashes, and compare.mjs on the full Haiku runs still prints 221.42 against 226.25.

Does this introduce any non-backwards compatible changes?

• Yes
• No

Does this include any user config changes?

• Yes
  • If so, I have updated the relevant areas of documentation
• No

---

Generated by Claude Code