perf: speed up unit tests and streamline checks

[thymikee]

Summary

Speed up the slow unit-test tail by replacing real waits, subprocess entrypoints, repeated fixture generation, and native/tool invocations with in-process entrypoints, controlled test seams, fake timing, and reusable fixtures. Behavioral assertions and coverage remain in place; runner adoption drops from roughly 28 seconds to tens of milliseconds in the full suite.

Make repository workflows clearer for agents and newcomers: document focused, affected, and full validation tiers; give common workflows stable script names; route test-app changes to an explicit cached typecheck lane; collapse platform perf aliases into pnpm perf --platform ios|android; and remove stale script/docs duplication.

Remove the misleading build:dev aggregate. Development guidance now points to the exact platform build, while npm prepack delegates to a tested package:npm release guard that builds TypeScript, iOS, macOS, tvOS, visionOS, the macOS helper, packaged Apple source, and both Android helper APKs. The macOS helper is clean-built for publishing so Swift caches cannot leak absolute paths between worktrees.

Validation

• pnpm check:tooling
• pnpm check — tooling, build, and Fallow passed; the saturated aggregate unit run passed 4,766/4,770 tests, and the four unrelated load-affected files then passed 137/137 in isolation
• pnpm test:maestro-compat — 32 files and 201 tests passed
• pnpm test-app:typecheck
• pnpm check:affected:test — 32 tests passed
• pnpm perf --platform invalid — confirmed the single script forwards and validates --platform
• Publish path — TypeScript, iOS, macOS, and tvOS builds passed; clean macOS-helper build, Apple source packaging, and both Android APK builds passed; visionOS awaits restoration of the temporarily removed local SDK component
• npm pack --ignore-scripts --dry-run --json — verified the 209-file tarball includes both Android artifact sets, Apple runner source, macOS-helper source, and the TypeScript distribution

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-29 16:54 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.88 MB	1.87 MB	-2.6 kB
JS gzip	604.0 kB	602.9 kB	-1.0 kB
npm tarball	722.5 kB	720.7 kB	-1.8 kB
npm unpacked	2.53 MB	2.52 MB	-6.9 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.2 ms	26.1 ms	-0.1 ms
CLI --help	56.6 ms	56.3 ms	-0.3 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/recording-export-quality.js	+648 B	+170 B
dist/src/apps.js	+64 B	+17 B
dist/src/session.js	+15 B	+8 B
dist/src/internal/daemon.js	-52 B	+5 B

[thymikee]

Reviewed exact head dea6e446: no code findings. The speedups preserve the production contracts: real waits become fake timers or explicit promise barriers, subprocess coverage retains an end-to-end CLI case while pure CLI logic runs in process, native/tool invocations use the existing scoped providers/executor override, and the screenshot worker is bypassed only in high-level pixel-behavior tests. Generic real ZIP and tar extraction coverage remains alongside the faster fixture seam. The new test-app selector maps TS/JS/JSON changes to the CI-backed isolated typecheck and fail-opens for this PR’s own tooling changes. Scoped exact-head validation passed 202/202 touched-suite tests plus 27 check-affected model tests; GitHub has no confirmed failures and remaining smoke/Coverage/FreeRange jobs are in progress. Ready for human review.

[thymikee]

Follow-up review on a0318436f is clean.

• Repaired the scheduled perf jobs to use the canonical pnpm perf --platform ios|android surface.
• Added a path-filtered macOS Package Smoke workflow that executes the real pnpm prepack path when package inputs change.
• Added regression assertions for the canonical commands and direct workflow inputs.

Validation: focused manifest tests, format, lint, typecheck, build/tooling, layering, metadata, and Fallow passed. The local package run compiled the iOS, macOS, tvOS, and visionOS runner paths; the complete Package Smoke job is queued on this head. Independent Terra re-review found no remaining actionable regression.