fix: preserve matched conditions in require_args evidence

[Sonike]

Summary

• A compliant require_args call with a matching condition incorrectly reported matched_when=False. Track whether any relevant call satisfies the condition so successful evidence reports True, including when later calls do not match.
• Keep scanning for later violations and preserve the existing violation and proximity results. Extend the policy matrix to cover matching/nonmatching conditions, multiple-call ordering, a later violation, absent tools, and unconditional requirements.
• Fixes [BUG] MUT-037: Successful require_args records matched_when=False #79.

Type of change

• Docs / community / DX only
• Bug fix
• Tests
• CLI / DX
• Adapter (new or Adapter docs: add policy.yaml operator cheatsheet #1)
• Core engine (policy, campaign, minimize, regression)
• Hosted API / UI (secondary surface)

Checklist

• I read CONTRIBUTING.md (where to ask, how reviews work, how to run tests)
• Linked related issue (Fixes #N) or Discussion when applicable
• Core stays free of framework SDK imports (adapters own glue)
• Tests added or updated when behavior changes (uv run pytest tests/unit -q)
• Docs / README claims match reality (no PyPI / multi-adapter claims unless shipped)
• One concern per PR when possible
• Authorized testing only — no open-internet attack-proxy behavior

Test plan

Validated with Python 3.12.13 and the offline sample mode (MUTINY_SAMPLE_OFFLINE=1), with isolated temporary database paths:

• uv run --offline pytest tests/unit/test_policy_evaluator.py -q --tb=short: 5 failures before the fix; 123 passed after.
• uv run --offline pytest tests/unit -q --tb=short: 380 passed, including package build/content checks.
• uv run --offline pytest tests/integration tests/reliability -q --tb=short: 140 passed (one existing Starlette deprecation warning).
• git diff --check: passed.

[vercel]

@Sonike 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 @Sonike — this hits #79 / MUT-037 cleanly: the require_args success path now tracks whether any relevant call matched when, so a compliant issue_refund(amount=250, approved=True) reports matched_when=True instead of the hardcoded False. Multi-call ordering, later violations, unmatched when, and absent tools stay correct. GitHub Actions green on 3.11/3.12; ignoring Vercel auth.

[CodewithJha]

Merged — thank you @Sonike. Exactly the kind of focused Core evaluator fix Mutiny needs: compliant require_args evidence now reports matched_when=True when the when clause actually matched, without changing violation or proximity semantics.

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

• #1 — policy operator cheatsheet; you already know the evaluator evidence shape
• #65 — Core campaign knobs that currently no-op or stall (use_boundary_seeds, elite_count)
• #7 — richer per-rule comment headers in the sample policy (XS, comments-only)

No pressure — only if you want another one.