ci: drop duplicate unit-test run, cache Size base dist, typecheck with tsgo

[thymikee]

Summary

Three measured speed cuts to the dev loop and CI, none of which remove signal. Follow-up to the tsdown migration (#1087), where the investigation happened.

• pnpm typecheck now runs tsgo (the native-preview compiler this repo already trusts for declaration emit in the tsdown build): 21.7s → 5.3s locally, taking check:tooling from ~85s pre-migration to ~18s end to end. tsc remains available as typecheck:tsc, and the CI Typecheck job gets the same speedup since it calls pnpm typecheck.
• Removed the Unit Tests CI job. Coverage runs the exact same unit + provider-integration suites under coverage thresholds on every PR, so the job duplicated ~64s of tests per PR with no extra signal. Integration Tests still runs provider-integration as its own gate.
• Size workflow: now skips docs-only changes (same paths-ignore list as CI) and caches the base commit's dist keyed on base SHA — dist is fully determined by that commit, so re-running a PR against the same base skips the base rebuild. Startup medians are still measured fresh per run, keeping the base/PR startup comparison same-machine. The cache saves immediately after the base measurement so the PR-side build can never poison the base key.

Not included, checked and rejected during investigation: vitest --no-isolate (breaks ~30 tests relying on fresh module state, and slower) and --pool=threads (27 errors) — the unit suite legitimately needs forks+isolation; Android AVD snapshot caching (whole smoke run is only ~3m22s on KVM runners, marginal payoff for the maintenance cost).

⚠️ If branch protection lists Unit Tests as a required status check, that entry needs to be removed/replaced (e.g. with Coverage) or PRs will block on a check that no longer exists.

Validation

Ran pnpm check:tooling green in 18.5s with tsgo in the loop, and verified tsgo is actually checking: an injected type error fails with TS2322 plus a noUnusedLocals TS6133, and the clean tree passes. Parsed both edited workflows with a YAML loader and confirmed the resulting job lists. The Size cache flow was desk-checked for the two orders: on miss, save runs right after the base measurement (before the PR build overwrites dist); on hit, the base build is skipped and the gitignored dist survives the base checkout.

Touched files: 3 (package.json, .github/workflows/ci.yml, .github/workflows/size.yml).

🤖 Generated with Claude Code

https://claude.ai/code/session_01FqeW8sA2ZnvnftdvpCqFMS

---

Generated by Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-04 14:40 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.5 MB	1.5 MB	-208 B
JS gzip	489.2 kB	489.2 kB	-50 B
npm tarball	588.4 kB	588.3 kB	-58 B
npm unpacked	2.1 MB	2.1 MB	-162 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.5 ms	27.6 ms	+0.1 ms
CLI --help	51.6 ms	51.8 ms	+0.2 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/logcat.js	-208 B	-50 B

[thymikee]

Review status: no actionable code blockers found.

I checked the workflow/package changes directly: Unit Tests is removed while Coverage remains the unit + provider-integration gate with thresholds; Size now ignores docs-only changes and saves the base dist cache before the PR build can overwrite dist; pnpm typecheck points at tsgo while typecheck:tsc keeps the legacy compiler available. Local validation passed for both pnpm typecheck and pnpm typecheck:tsc, and GitHub checks are green (21/21), so the edited workflows are syntactically accepted and running.

Operational caveat: as the PR body notes, if branch protection still requires the old Unit Tests status, that repository setting must be updated outside this PR. Added ready-for-human.