test: isolate persisted violation traces from committed examples

[fatihcvs]

Summary

Fixes #72. Persist the M4 violation artifact under pytest's tmp_path instead of overwriting the tracked example on every run. Keep the existing tool-call assertion and verify the complete JSON round trip plus the violated refund rule.

The committed example and its annotated documentation remain unchanged. I reproduced the random seed-ID change while running the previous unit suite; after this change the full unit suite leaves examples/traces/ clean.

Type of change

• Tests

Checklist

• Read CONTRIBUTING.md
• Linked related issue
• Core stays free of framework SDK imports
• Tests updated and full unit suite run
• Documentation claims remain accurate
• One concern per PR
• Authorized offline testing only

Test plan

• pytest tests/unit -q --tb=short: 376 passed on Python 3.12.14, with MUTINY_SAMPLE_OFFLINE=1 and PYTHONUTF8=1.
• git diff --exit-code -- examples/traces: clean after the suite.
• git diff --check: passed.

Based independently on main; does not include #83.

[vercel]

@fatihcvs is attempting to deploy a commit to the priyanshu's projects Team on Vercel.

A member of the Team first needs to authorize it.

[CodewithJha]

Thanks @fatihcvs — this hits #72 / MUT-024 cleanly: persistence is asserted under pytest tmp_path, so unit runs no longer rewrite examples/traces/m4_refund_limit_violation.json. The committed example and annotated docs stay the hand-maintained fixture. GitHub Actions green on 3.11/3.12; ignoring Vercel auth.

[CodewithJha]

Merged — thank you @fatihcvs. Exactly the kind of tests-hygiene lock Mutiny needs: the M4 persistence assertion still holds, and pytest no longer rewrites a tracked example trace.

If you'd like to continue working on Mutiny, these would be good related areas to look at:

• #75 — CLI / loader / reliability tests still miss several product paths
• #57 — mutiny run error-path coverage (adapter/config failures should not traceback)
• #69 — in-process second project load can reuse the first agent module; a tests-shaped isolation bug

No pressure — only if you want another one.