Mutation-testing foundation

exhibit_a.verdict.mutation_testing measures how strongly a generated test constrains suspect code. It is deliberately a scoring layer, not an evidence gate: flip_check.py remains the only module that can admit a Case.

Process

1. discover_mutations tokenizes caller-selected Python source files and emits a deterministic ordered set of allowlisted operator replacements. Optional suspect line sets narrow the surface; input paths are repository-relative and cannot traverse outside the checkout.
2. score_mutations reruns the candidate on the unmodified pass-state baseline. If that baseline is not deterministically green, no kill rate is reported.
3. Each mutant is independently applied to a fresh disposable executor copy. Docker execution retains the no-network, dropped-capability, read-only-container policy.
4. A mutant is killed only when every run fails with the same extractable signature, survived only when every run passes, and invalid when unsupported, flaky, timed out, or broken at the harness/environment layer.

The score reports killed / (killed + survived). Invalid mutants remain visible but are excluded from the denominator. No score can upgrade, downgrade, or fabricate the raw verdict. The current bug-repro judge emits VERIFIED, PARTIAL, or UNCERTAIN; FAILED is reserved for a future claim judge that can deterministically disprove its stated goal.

Initial operator set

• equality and ordering boundaries: ==, !=, <, <=, >, >=
• arithmetic operators: +, -, *, //, %
• Boolean literals: True, False

These are syntax-preserving token edits rather than model-generated patches. The executor independently rejects any replacement outside the allowlist, even if a caller constructs a SourceMutation directly. The current foundation targets UTF-8 Python source and requires callers to identify suspect files; broader language and mutation-operator studies belong in later research commits.