[jackhai9]

Summary

Three focused commits that make local runs trustworthy, restore the weekly coverage job, and shorten PR CI.

1. Enforce the pinned Node runtime (chore)

• The runtime was pinned only in .nvmrc; shells that skipped nvm use (including non-interactive agent shells) ran on Node 22. There, import.meta.main is undefined, so the exact-file runner and coverage CLIs skipped all work and exited 0 — a false green for test:affected / test:coverage.
• package.json devEngines makes npm install/ci/run/test/exec refuse any other runtime (EBADDEVENGINES).
• Repository CLIs assert the .nvmrc version at module load, before their entry guards (reusing the existing assertProjectNodeVersion, moved to scripts/node-runtime.mjs).
• A contract test keeps devEngines aligned with .nvmrc; AGENTS.md and docs/test-policy.md document per-command runtime selection.

2. Repair the weekly coverage job (fix(coverage))

• Since the multi-tab specs (feat: integrate chart storage optimization into orderbook v2.7.218 #306, fix: stop other-symbol tabs from restoring deleted drawings #307, fix: preserve data panel positions across viewport changes #317), the fixture covered only the default page: reloading a page opened with context.newPage() or closing the default page mid-test failed 11 scenarios in coverage mode only.
• Coverage now follows every page in the scenario context (started before it can navigate, written before it closes); the report accepts several CDP sessions per scenario and rejects duplicate sessions. A self-test spec covers open/reload/close.
• With complete collection the gate showed 85.52%: the byte-pinned copies of Binance's native webpack factories (mirror-module.js 7/438, drawing-save-module.js 5/206 branches) were counted as production source. They are now mapped like bundled dependencies; hand-written chart-storage modules stay in scope.
• chart-save-coalescer.js was 2 branches short of its 90% critical gate; two real failure paths gain behavior tests.

3. Speed up PR CI (perf(ci))

• Playwright ran 621 scenarios on one worker (~8.8 min of a 10-14 min job). Spec files use isolated contexts with routed fixtures, so they now run on half the logical CPUs (2 workers on 4-vCPU runners); tests inside a file stay ordered. The instrumented coverage layer stays serial.
• Changes limited to .gitignore files no longer select the full suite.

Validation

• npm test on Node 24.16.0 — 2626 passed; npm run lint:tests, npm run test:test-policy (64 passed), and git diff --check passed.
• Node 22.17.0: npm test now fails with EBADDEVENGINES; running scripts/test-selection/node-runner.mjs or scripts/test-coverage/run.mjs directly exits 1 with Use the project Node version 24.16.0 from .nvmrc (nvm use); received 22.17.0 (before: the runner exited 0 with no output).
• Coverage mode, the three previously failing spec files: 11 failed / 19 passed before, 30 passed after (59 captures, 16 scenarios with several page sessions, all session IDs unique).
• Full npm run test:coverage locally: 622 browser scenarios passed (serial, 9.3 min); aggregate branch coverage 90.17%; gate passed, including the critical modules.
• Full Playwright suite, plain mode, local 8-CPU machine: 1 worker 7.7 min, 2 workers 3.9 min, 4 workers 2.4 min; 622 passed in every run. In coverage mode with 4 workers one timing-sensitive scenario exceeded its 3 s expectation (it passes 3/3 alone), so the instrumented layer stays serial.
• Not exercised: live Binance pages and Tampermonkey (no userscript source or behavior change). PR CI timing is observed on this PR.

🤖 Generated with Claude Code