test: tolerate copier's temp-clone cleanup race on a dirty tree

[ichoosetoaccept]

Summary

generate_project in tests/conftest.py now tolerates one specific copier exit: a successful render whose temp-clone cleanup crashed.

Why

poe test failed whenever the working tree was dirty and passed when it was clean. One unrelated newline reproduced it. CI never saw it — CI always checks out clean — so the suite was green exactly when nobody needed it and red exactly when somebody was mid-change.

On a dirty tree copier takes its dirty-file overlay path, which does extra git work inside the temp clone it makes of this repo. Something still holds a handle under that clone's .git when _cleanup calls rmtree(ignore_errors=False), so copier exits non-zero after every file has already been written. The render succeeds; only the cleanup fails.

It reads as flakiness and is not. The failure lands on whichever test happens to render first, so the count and identity of "failing" tests vary per run, and re-running one in isolation usually passes — which sends triage toward "just a flake" and then toward "it started when I changed X". Both are wrong.

Narrow on purpose, in three ways: stderr must name a path copier itself calls copier._vcs.clone.*, it must carry an rmtree/Errno 66 signature, and the destination must actually contain a rendered pyproject.toml. A render that genuinely failed writes no pyproject.toml, and every other non-zero exit still fails the test with copier's own stderr. A broader "ignore non-zero exits" would be a real liability in a suite whose entire job is asserting on rendered output.

Test plan

• Fast suite on a dirty tree: 202 passed. The same tree previously gave 2 failed, 13 passed in TestTemplateUpdateCheck alone, with the failures on tests unrelated to the edit.
• Fast suite on a clean tree: unchanged, the tolerance path is never entered.

Closes DOT-606

Greptile Summary

The fixture now retries a narrowly classified temporary-clone cleanup failure instead of accepting any nonzero Copier result. Focused tests cover the accepted cleanup traceback and its rejection gates.

• Adds a maximum of three render attempts for the intermittent dirty-tree cleanup race.
• Requires a cleanup-frame traceback, Copier temporary-clone path, no chained exception, and rendered pyproject.toml before retrying.
• Adds regression coverage based on representative cleanup and render-error tracebacks.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the current implementation only returns after a successful Copier exit, and the previously reported coverage and partial-render suppression issues are fixed.

Important Files Changed

Filename	Overview
tests/conftest.py	Replaces failure suppression with bounded retry behavior while preserving failure for every unsuccessful final attempt.
tests/test_template.py	Adds focused coverage for the retry classifier’s accepted traceback and rejection boundaries.

_{Reviews (6): Last reviewed commit: "fix(test): retry the cleanup race instea..." | Re-trigger Greptile}

[ichoosetoaccept]

This change is part of the following stack:

• fix(template): repoint sync-with-uv at the upstream project #349
  • fix(template): stop chmodding scripts the template does not own #347
    • test: tolerate copier's temp-clone cleanup race on a dirty tree #348 ◀

_{Change managed by git-spice.}

[linear-code]

DOT-606