fix(tests): stop exhausting test-db connections and migrate once

[KenTaniguchi-R]

The problem

pnpm test at vitest's default concurrency reported 137 failures that were not real:

Test Files  56 failed | 56 passed (112)
     Tests  137 failed | 479 passed | 123 skipped (739)
  Duration  339.05s

Errors were 57P01 ProcessInterrupts — Postgres terminating connections — attributed to whichever tests were in flight, not to any actual defect.

Root cause

Each integration file opens its own pg.Pool against the single shared testcontainer. pg pools default to 10 connections. On a 14-core machine vitest runs ~14 workers, demanding up to 140 connections against a server whose max_connections is 100 (verified against postgres:17-alpine).

Postgres starts dropping connections, and the failures land wherever they land. That also explains the flakiness: three runs on identical code gave 5 → 1 → 0 failures.

The fix

Connections. max_connections=300 on the container; each pool capped at 4. The suite needs a couple of concurrent queries per file, not ten.

Migrations — the dominant cost. All 52 integration files replayed the full 9-migration chain: 468 migration runs per suite. Now migrated once into a template database in global setup, with each file cloning it via CREATE DATABASE ... TEMPLATE. Concurrent clones briefly conflict (55006), so that specific error retries with backoff. When no template is present the old path still replays migrations, so a bare DATABASE_URL keeps working.

Durability. fsync, synchronous_commit, and full_page_writes off — the server is destroyed at teardown.

Results

pnpm test, default concurrency, no worker cap:

	Before	After
Result	137 failed / 479 passed	727 passed / 727
Duration	339s	21s

16x faster, and green. Five consecutive integration runs all passed 301/301, where the suite previously varied run to run.

This should also help the CI backlog: there is one self-hosted runner (macmini-ledgr) and every branch's jobs queue behind it serially, so shortening each job compounds.

Also: ESLint

.stryker-tmp/** added to globalIgnores. Stryker leaves sandbox copies of the entire tree behind, and linting them produced 981 of 1024 errors, failing pnpm lint on duplicated copies of our own code. Lint is now 0 errors (1 pre-existing warning, _accountType in src/lib/money.ts).

Verification

Run in an isolated worktree off main:

• pnpm install --frozen-lockfile — clean
• pnpm typecheck — clean
• pnpm lint — 0 errors (verified with a sandbox present that previously errored)
• pnpm test — 727 passed / 112 files
• 5x consecutive integration runs — 301/301 each

🤖 Generated with Claude Code