build: strip runner unit-test blocks from package

[thymikee]

Summary

Strip AGENT_DEVICE_RUNNER_UNIT_TESTS Swift blocks from the npm-packaged Apple runner source by generating a sanitized dist/apple-runner copy during prepack.

Teach runner source resolution to prefer checkout source locally and fall back to the packaged source in installed packages, so local runner unit tests keep their gated helpers.

Validation

• node scripts/package-apple-runner-source.mjs
• npm pack --dry-run --ignore-scripts --json --cache .tmp/npm-cache
• package manifest assertion for dist/apple-runner presence, top-level apple-runner absence, and no generated Swift AGENT_DEVICE_RUNNER_UNIT_TESTS guards
• node ./node_modules/oxfmt/bin/oxfmt --check ...
• ./node_modules/.bin/oxlint . --deny-warnings
• ./node_modules/.bin/tsgo -p tsconfig.json
• node --experimental-strip-types scripts/layering/check.ts
• node scripts/sync-mcp-metadata.mjs --check
• ./node_modules/.bin/tsdown
• ./node_modules/.bin/vitest run src/__tests__/package-exports.test.ts src/__tests__/apple-runner-package-source.test.ts src/platforms/apple/core/__tests__/runner-source.test.ts src/platforms/apple/core/__tests__/runner-client.test.ts

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-06 10:52 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.5 MB	1.5 MB	+180 B
JS gzip	491.8 kB	491.9 kB	+91 B
npm tarball	597.4 kB	588.3 kB	-9.0 kB
npm unpacked	2.1 MB	2.1 MB	-45.6 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	25.7 ms	26.5 ms	+0.8 ms
CLI --help	50.7 ms	49.6 ms	-1.1 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/runner-client.js	+135 B	+61 B
dist/src/record-trace-recording.js	+45 B	+30 B

[thymikee]

Ready for human review.

Review pass on #1128:

• GitHub CI is clean: 22/22 checks passing.
• The original CI blockers were addressed: Bundle Size now handles base commits where the helper script does not exist, and Fallow is clean after the package script and recording resolver cleanup.
• Local/worker validation covered the package-source tests, recording-script resolver test, package-source generation, Fallow, build, and size-report base/current comparison.
• Independent Codex review found no actionable blockers. I attempted the required claude -p review for this codex/ branch, but the sandbox reviewer blocked it because it would send PR-specific repo context to an external Claude service; I did not route around that.

Residual risk: installed-package Apple runner behavior depends on prepack generating dist/apple-runner, which is wired and tested at the package-script level, but broader installed-package device smoke coverage is still the main remaining confidence gap.