perf(mutation): stop booting a Postgres container for every mutant

[KenTaniguchi-R]

Split out from #99 — that PR is pure CI config; this one changes mutation scores and deserves its own review.

Problem

Stryker was spending most of its runtime starting databases, not testing mutants. A 9-file diff kept the mutation job busy for 30+ minutes on the CI runner, with 18 Postgres containers alive simultaneously and the 16GB mini swapping (991k pageouts).

Cause

Three things compound:

1. Stryker runs vitest once per mutant batch across concurrency workers — default is cpuCores - 1, so 9 on the Mac mini.
2. vitest.config.ts declares globalSetup, which boots a testcontainers Postgres and replays the entire migration chain. That fires on every worker.
3. vitest.related (default true) already limits which tests run — but globalSetup runs regardless of which tests are selected. So Postgres booted even to run money.test.ts.

Net effect: 8.92s of database bootstrap to run 29ms of tests.

Fix

A Stryker-only vitest config via the supported vitest.configFile option, with no globalSetup, excluding tests/ and the five colocated src tests that need a live database. Plus ignoreStatic: true, which skips mutants that only execute at file load — precisely the ones that force an environment reload.

vitest.config.ts is untouched, so pnpm test and CI's test job are unaffected.

Measured

	Before	After
One test file (29ms of tests)	6.66s	1.37s
Full unit suite (53 files, 390 tests)	—	6.47s, 0 containers
Stryker on money.ts — dry run	24s, 192 tests	3s, 111 tests
Stryker on money.ts — mutation phase	2m42s	1m52s
Peak Postgres containers	7 (this machine) / 18 (mini)	0

Measured on an M4 Pro (14 cores, 48GB). The gap is far wider on the mini — 10 cores and 16GB with swapping is where 7 containers became 18 and minutes became 30.

The score change, stated plainly

Mutation score on money.ts moves 77.32% → 75.26%. The composition is better than the number:

before: 71 killed, 4 timeout, 22 survived
after:  73 killed, 0 timeout, 24 survived

It kills more mutants outright. The old run's 4 timeouts — which Stryker counts as kills — were mutants timing out on container contention, not tests catching them. Part of the old score was infrastructure noise scored as success.

Mutants covered only by DB-backed tests now report NoCoverage. That's honest signal about unit-test strength, and it makes ci.yml's claim that Stryker "currently runs just the unit suite" true — it wasn't before.

Verification

• pnpm typecheck — clean
• pnpm lint — 0 errors (1 warning, pre-existing on main, verified by stashing)
• pnpm test — 113 files / 751 tests pass in 32s
• Full unit suite under the new config — 53 files / 390 tests pass, zero containers
• End-to-end Stryker runs both ways on money.ts, numbers above

Note

coverageAnalysis: "perTest" is kept for documentation only — the vitest runner ignores it and always uses perTest, but ignoreStatic documents perTest as a prerequisite.

🤖 Generated with Claude Code