fix: pre_build must not stage untracked human work

[Jammy2211]

Closes #232.

run_workspace ran black "$d/" then git add "$d/" over scripts/, notebooks/ and slam_pipeline/ across 13 repos. Both operations reach untracked files, so any uncommitted work in a workspace's scripts/ was reformatted on disk and pushed inside the "pre build" commit — to a public repo, silently, exit 0.

Same leak class as #126, which was fixed for dataset/ and config/ by dropping their staging (#156) while leaving the scripts/ path open. Near-miss during the 2026-08-07 release, caught only because the operator moved the file out by hand.

Reproduced first, against the unmodified script on throwaway fixture repos with real bare remotes: the private file was committed as "pre build" and pushed to the remote, exit 0. The same fixture with this change aborts at the preflight, leaves the file byte-identical (md5 unchanged — black never reached it), and creates no commit.

What changed

A fail-fast preflight sweeps every repo for untracked files under the directories the run mutates, before the first repo is touched. It has to precede everything: run_workspace commits and pushes each repo before moving to the next, so a per-repo check aborting midway would leave earlier repos already published. It reports every offending repo and path in one pass, so one run surfaces the whole problem. It also aborts on a missing checkout, which previously surfaced as a bare cd error partway through a run.

This answers the open atomicity question in docs/pre_build_failure_audit.md §6 — "worth a fail-fast pre-pass (all repos validated before any push)?" — which that audit costed as a follow-up. Marked resolved there.

Staging narrows to git add -u (tracked edits and deletions) plus files created by the run, added by explicit path, so the directory-wide form cannot return. New notebooks from generate.py are still staged — which is why git add -u alone is insufficient.

The repo list moves into a WORKSPACE_SPECS array. Two passes now read it, and a second hand-maintained list would drift out of step — the preflight would then silently skip a repo it exists to protect.

No --allow-dirty override, deliberately: an override is precisely the operator vigilance this replaces.

Files

• pre_build.sh — the two legs above.
• tests/test_pre_build_staging.py (new) — runs the real script against fixture git repos with black/python/gh stubbed and real bare remotes. Text assertions cannot prove a gate fires; these do. Six cases: WIP abort, multi-repo reporting, gitignored files not blocking a release, generated notebooks still staged, tracked deletions, missing checkout.
• tests/test_pre_build_skill.py — repo-list regex retargeted at the array; two assertions added (no directory-form git add; preflight precedes every mutation).
• skills/pre_build/pre_build.md — the preflight documented as a numbered step.
• docs/pre_build_failure_audit.md — §6 atomicity question marked resolved.

Test state

302 passed / 7 failed / 3 skipped locally. The 7 failures are pre-existing and environmental — ipynb-py-convert will not build in this sandbox — and the identical 7 fail on pristine main with these changes stashed. All 10 pre_build tests pass.

---

Generated by Claude Code