v2.3.0 — Batch Auto-Advance, Privacy Architecture, Output Format Expansion

Batch Auto-Advance

start_batch now self-advances through the queue without polling. An on('exit') handler fires when each whisper-cli process completes, immediately spawning the next job with zero API calls consumed. A mutex prevents double-spawn between concurrent exit handler and check_batch_progress calls. Unattended overnight batch processing now works as designed.

Privacy Architecture

Full privacy mode implementation:

WHISPER_PRIVACY_MODE=true — all tool responses return metadata only (filename, word count, save path). No transcript text transmitted to Claude's API under any circumstances.
WHISPER_CONSENT_ACKNOWLEDGED=true — suppresses the one-time session consent gate for non-sensitive content.
privacy_mode per-call parameter on transcribe_audio, transcribe_batch, start_batch, and check_progress. Overrides global env var in either direction, no restart required.
Per-operation confirmation gate for privacy mode. Session consent gate for standard mode.
Full compliance documentation in PRIVACY.md covering HIPAA, GDPR, attorney-client privilege, FERPA, SOX, PCI-DSS, and NDA/trade secret.

Output Format Expansion

vtt — WebVTT subtitle output, available in transcribe_audio, generate_subtitles, start_batch, and background mode
lrc — LRC lyrics/karaoke format
csv — CSV with timestamps
output_format default changed from text to timestamps across all tools

Bug Fixes

gpu_device parameter now correctly maps to --device flag instead of invalid -g
output_format was not forwarded to background jobs — fixed
Silent catch {} in background job output move swallowed failures — now surfaces a detailed error message

Documentation

All docs brought to v2.3.0 parity
--device flag correction applied across all 10 multilingual ROADMAP files
Temp directory auto-cleanup documented (7-day retention, runs at startup)