workflow-verify-enhanced

Verification accuracy (~87%) far exceeds generation accuracy (~63%), so spending model budget on STRONGER tests pays off. Two gate extensions in mcp-verify:

• property_test(path) — the local model generates 3–5 Hypothesis @given property tests (falsifiable invariants: round-trips, bounds, ordering, agreement with a slow reference), runs them, and returns minimal counterexamples. Hallucinated properties (those that fail to import/collect) are filtered out. PGS lifted fix rate 23.1%→53.8% over a TDD baseline (arXiv:2506.18315); Anthropic's agentic PBT found real bugs in NumPy/HuggingFace (arXiv:2510.09907).
• mutation_test(path, test_path) — mutmut mutates the module and runs your tests against each mutant. A surviving mutant is a test that fails to catch a real bug — a gap. Meta ACH: mutation beats coverage (arXiv:2501.12862).

How to use

1. Get unit tests green first (verify).
2. property_test the core logic functions — if it returns a counterexample, your function (or your understanding of its contract) is wrong; fix it.
3. mutation_test the changed files — for each surviving mutant, add a test that kills it, then re-run. Only declare done when no mutant survives.

Both are time-bounded and opt-in for the primary gate (ENABLE_PROPERTY_TEST=true) because property generation adds wall time and the model can hallucinate properties — treat a generated property's failure as a hypothesis to investigate, not a verdict.