fix(ci): surface test failures in logs and de-flake perf tests

[ffflorian]

Problem

Two separate issues on main:

1. Test failures are invisible in CI logs. scripts/test-summary.js ran each package with stdio: "pipe" and printed only scraped counts, discarding all failure output. The failing run on main showed just ❌ @codemcp/knowledge-mcp-server: 21/22 passed.

2. Worse — the script could report red as green. Pass/fail was derived from scraped numbers, and the regexes didn't match vitest output containing skipped tests (Tests 1 failed | 22 passed | 2 skipped (25)). Unmatched packages were silently dropped from the report and the totals. Verified against the old script with a deliberately failing test: it printed 🎉 All tests completed successfully! and exited 0.

3. Flaky perf assertions in packages/mcp-server/src/__tests__/performance.test.ts:

• expect(time).toBeLessThan(10) on a single cold sample, including module init and JIT warm-up — 2.29ms locally, under 4x headroom on a shared CI runner.
• expect(memoryDiff).toBeLessThan(10MB) on a single heapUsed delta — reported -0.36MB locally, i.e. GC noise dominated the signal entirely.

Changes

• Test runner streams child output to stdout/stderr while buffering it, and derives success from exit codes, never parsing. Parsing is presentation-only, so an unrecognised summary line can no longer flip a red run green. Failed suites are named explicitly at the end.
• Perf timing: warm up, then assert median of 9 samples against 50ms.
• Perf memory: assert aggregate growth across 50 creations with a 1MB-per-server budget.
• Root test now runs turbo run build first. packages/cli integration tests exec dist/index.js, but the root script bypassed turbo so its existing dependsOn: ["build"] never applied — a stale dist produced failures unrelated to the current source.
• Added dist/ to .prettierignore (root + per package). Prettier was checking build artifacts; CI only passes today because format:check happens to run before build.

Verification

• Failing-test injection now yields exit 1, the suite named, and the full assertion diff, file, line and code frame in the log.
• Skipped-test counts parse correctly (22/25 passed, 1 failed, 2 skipped).
• Perf tests reproducible across runs: median 0.13–0.17ms, 92.6–92.7KB per server (vs. the old ±0.36MB swing).
• Full CI sequence green locally: format:check:all, lint:all, build, test — 265 tests, 4 consecutive runs.

Not verified: behaviour on a genuinely contended CI runner. The headroom argument is inference from local numbers, not a measurement.

[mrsimpson]

Ouch... that was really poor earlier.