Audit follow-up: correctness fixes, JSON output, tests, CI, packaging

[Copilot]

Acting on the prior audit of this repo, this PR ships the highest-leverage subset: real correctness bugs in the audio pipeline, a missing machine-readable output format, and the trust-signal infrastructure (tests, CI, packaging, changelog) that the project was lacking.

Correctness fixes

• WAV save produced static. audio_capture.save_wav wrote raw float32 frames through stdlib wave, which always emits a WAVE_FORMAT_PCM header — most players then decoded the bytes as int32. Now clipped + converted to int16 PCM.
• 16 kHz resampling was nearest-neighbour decimation. Aliased badly at 48 kHz → 16 kHz and measurably hurts WER. Now uses scipy.signal.resample_poly (polyphase + anti-aliasing) when available, with a linear-interp fallback.
• Transcriber was hard-coded to CPU. Now probes torch.cuda.is_available() and picks a CTranslate2-compatible compute type (float16 only on CUDA, int8 on CPU).
• SIGINT handler leaked. start_recording now restores the previous handler on exit, so successive tray-mode recordings and the parent shell aren't affected.

New output format

• --format json (and included in --format all). Versioned schema:

  {"version": 1, "segments": [{"start": 2.34, "end": 5.12, "speaker": "Speaker 1", "text": "..."}]}

Packaging & dev experience

• pyproject.toml adds a meeting-recorder console script and optional dep groups (windows, pyannote, dev). Flat layout preserved via py-modules so existing imports keep working.
• faster_whisper import made lazy in transcriber.py — importing CLI helpers (and tests) no longer requires the full ML stack.
• Upper bounds on pyannote.audio<4 and torch<3 (both ship breaking releases routinely); same bounds mirrored across requirements*.txt.

Tests & CI

• tests/ (22 tests) — format_* helpers, txt/srt/json writers, _get_output_paths for every --format, format inference from extension, energy diarizer (empty input, label format, reset, max-speaker cap, distinct-tone case). conftest.py puts the repo root on sys.path so the flat-layout modules import without an editable install.
• .github/workflows/ci.yml — ruff lint + pytest matrix on Linux/macOS/Windows × Python 3.10/3.11/3.12. Heavy runtime deps not installed in CI; the lazy imports keep the test surface lightweight.
• Ruff config in pyproject.toml with conservative rule set (E/F/W/I/UP).

Documentation

• CHANGELOG.md (Keep-a-Changelog), CONTRIBUTING.md (setup, tests, lint).
• README: CI + platform badges, "Why this vs. alternatives?" comparison table (whisper.cpp, Buzz, WhisperX, cloud), JSON output example, "Known limitations" section that's honest about per-chunk diarization drift and the energy heuristic's failure modes.

Intentionally deferred

Each of these warrants its own PR with real-device testing on multiple OSes:

• Cross-platform audio capture (PulseAudio / CoreAudio / ScreenCaptureKit).
• src/ package restructure.
• Per-chunk → end-to-end diarization pipeline.
• Local-LLM summarisation, calendar integration, voiceprint speaker naming.