Scoped mutation run splits by project #309

Requirement

scripts/mutation-changed.sh runs Stryker once per changed project instead of once over the whole solution.

Stryker 5.0's solution mode captures per-test coverage once per mutated project, and every capture runs all tests in all assemblies (StrykerRunner.AnalyzeCoverage → MicrosoftTestPlatformRunnerPool.CaptureCoverageTestByTest, which ignores the project). All five captures record the same coverage. A scoped run on one Worker file took 9:22 in solution mode and 2:15 with --project Anichron.Worker.csproj, with the same 31 mutants and the same 100% score.

Acceptance criteria:

• Projects are mutation targets by an MSBuild convention: IsMutationTarget is true for every non-test project (Directory.Build.props), and Anichron.Testing opts out in its csproj.
• scripts/mutation-targets.sh lists the target projects from Anichron.slnx through MSBuild. mutation-changed.sh uses it to group the changed files by project.
• One Stryker run per changed target project (--project <Name>.csproj plus that project's -m files), one after another, each with its own TMPDIR.
• Every project's run goes ahead even when an earlier one fails. The script exits with the worst status.
• A changed file outside any target project is named in one line ("not mutated: …") and does not fail the run. A test file that maps to no source file still exits 1, as today.
• mutation-guard.sh checks that every project opted out of IsMutationTarget has a !**/<Name>/** entry in stryker-config.json's mutate, and that every whole-project exclusion belongs to an opted-out project. The local full sweep still runs in solution mode and needs that entry.
• docs/mutation-testing.md and CLAUDE.md state no fixed run durations.
• ADR 0009 records why mutation runs per project, with the measurements.

First caller

scripts/mutation-changed.sh, run before every production change (CLAUDE.md). The CI matrix issue reuses mutation-targets.sh.

Design

Not needed because the change is confined to the scoped-run script, an MSBuild property and the guard, and measures the same mutants against the same tests. The trade-off is recorded in ADR 0009, written with this issue.