[Teyocesu]

Summary

• repairs the AudioTap resampler redeclaration without removing or weakening its assertions
• splits ClassScribe, AudioTap, and Windows development gates into parallel jobs
• replaces repeated build/package work with the minimum development checks
• keeps release/distributable validation in release.yml
• skips multiplatform builds for docs/** and Markdown-only changes

Real timings

Baseline is successful run #17. Final is green run #30 attempt 2; it used fresh hosted runners and no dependency/build cache.

Gate	Before	After	Savings
macOS critical path	9m37s	1m25s	8m12s (85.2%)
macOS AudioTap	serialized inside macOS	48s, parallel	removed from ClassScribe critical path
Windows	3m54s	1m19s	2m35s (66.2%)
Workflow wall-clock	9m43s	1m35s	8m08s (83.7%)

The latest pre-change Windows run was 2m25s, so the final 1m19s is also 45.5% faster than the current baseline. One no-cache run hit a NuGet/network outlier (3m14s Windows; 3m19s workflow); an immediate identical rerun produced the final table above.

What changed

PR/main CI retains:

• swift test for ClassScribe with complete strict concurrency
• swift test for all AudioTap tests with complete strict concurrency
• locked Windows restore
• Windows Release compilation through dotnet test plus all 182 tests
• explicit verification that global.json resolves a compatible installed SDK

Normal CI no longer repeats:

• explicit SwiftPM resolve, Debug build, and Release build before swift test
• Xcode Debug, Release, and Diarizer builds
• development bundle assembly and codesign verification
• Windows self-contained publish, native runtime validation, process smoke, portable staging, ZIP, and SHA256

release.yml still owns the exhaustive release build, distributable/DMG validation, Windows package and installer work, smoke checks, checksums, and final asset validation. It is unchanged.

Runner and cache evidence

• macOS runner: 3 CPUs, 7 GiB RAM. Removing -j 2 let SwiftPM choose its own parallelism and passed.
• Windows runner exposed SDK 10.0.303; global.json 10.0.302 with latestPatch resolved to 10.0.303 and passed. actions/setup-dotnet was removed.
• No Swift cache was added: the simplified ClassScribe job already met the target.
• A locked NuGet cache experiment was rejected: the archive was ~615 MB and a hit spent 33s restoring it, producing a 2m20s Windows job. The final workflow has no NuGet cache.
• push/main remains for direct-push coverage; PR checks remain the everyday gate. No commit-message parsing was introduced.

Hidden AudioTap failures exposed

Once the redeclaration compiled, three stale assertions from earlier [skip ci] production changes became visible. Tests now assert the documented current contracts: an empty callback is a transport-health heartbeat, and two input frames remain two master frames before/after flush. Production behavior was not changed.

No tests were removed, skipped, or made continue-on-error.