Select the test files to run for each mutant — coverage-driven test selection.

Runs after Mutare.Runner.Baseline has confirmed the suite green and measured the timing; this module is purely about coverage and never green-checks or times anything. Keeping the two apart is the point: folding coverage into the baseline made baseline_ms a sum of per-file process boots (an inflated timeout cap), and meant the suite was never confirmed green together — only file-by-file.

It is one instrumented suite run. The probe runs mix test once at baseline with the coverage-capture flag set, so the metamutant self-records — in the test process, synchronously — which mutant ids each test file covers, plus a process-agnostic aggregate of every id that ran at all (Mutare.Coverage.Recorder implements the capture, Mutare.Coverage reads the dump). No :cover, no per-file subprocess fan-out, and no async-formatter race that loses fast async: false modules' coverage.

Three modes, set by :test_selection, a granularity ladder from safest/slowest to fastest — :full ⊃ :coverage ⊃ :tests:

• :tests (default) — per mutant, run only the individual test cases that covered its line, via mix test <file> --only test:<name> (one --only per covering test). Finer than :coverage: where :coverage runs a whole covering file, :tests runs just the tests in it that touched the line. It falls back to :coverage's whole-file selection for any id whose coverage can't be pinned to runnable tests — an id in the wholefile set (covered through a setup_all/on_exit, which map to a module-scoped context, not a single test), or one covered via an unlabeled process (whole suite). So the narrowing is applied only where every covering attribution is a concrete, runnable test.
• :coverage — per mutant, run only the test files that covered its line (a setup_all attributes to its own module's file via the __ex_unit__/2 stacktrace frame; a Task via its caller chain; a test-registered on_exit via its closure frame); a mutant whose code ran in an unlabeled process (a bare spawn, a setup-registered on_exit closure, or a setup_all whose work happened in a spawned Task) runs the whole suite — even if some file also attributes it, since that partial attribution would otherwise mask the unlabeled coverage and produce a false survivor; a mutant that never ran at all is :no_coverage (skipped, and kept out of the score's denominator).
• :full — no per-file selection: every covered mutant runs the whole suite, the rest are :no_coverage. Safest for suites with cross-file dependencies — including a setup_all with cross-module global side effects, which :coverage attributes to its own file only (see Mutare.Coverage.Recorder).

:tests and the file-granular safety margin it trades away

:coverage is file-granular on purpose: if any test in a file covers the line, the whole file runs — so a sibling test that kills the mutant indirectly (fails because a covering test in the same async: false module ran the mutated line and left corrupt shared state, without the sibling touching the line itself) is still run. :tests narrows to the covering tests only, dropping that sibling — so on a stateful, cross-test-dependent suite :tests can turn such a kill into a false survivor. It is the default on the assumption that suites are predominantly async: true with per-test-isolated state (where a sibling that never runs the line cannot observe the mutation, so nothing is lost); :coverage is the opt-out (--per-file) and :full the fully conservative escape hatch.

The narrowing never manufactures a false kill: every name emitted is a test that actually ran in one of the included files during the probe, so at least one test always matches and mix test --only never hits its "no test executed" error (which would otherwise surface as a harness error, not a kill).

Coverage is advisory, never authoritative. Anything uncertain — a non-zero probe exit, an unreadable dump, or missing capture tables — degrades to :run_all: we never skip a mutant on doubt. A valid empty dump means none of the emitted mutants ran; focused selection can legitimately leave the entire aggregate empty. But because :run_all makes every covered mutant run the whole suite — prohibitive on a large project — a failed probe run is retried once before degrading: the baseline was green moments earlier, so a probe failure is usually a flaky test, and one extra suite run is cheap next to a whole run's selection quality. The probe run is wall-clock capped for the same reason (by default a generous multiple of the per-mutant cap, since instrumentation adds overhead a plain baseline doesn't have; :probe_timeout sets an explicit cap instead): a pathological interaction between the coverage capture and the target's hot loops must degrade to :run_all, not hang the whole run at the probe stage forever.