Add duration-priced voice cost governance

[oratis]

Summary

• introduce a separate duration-priced media ledger instead of encoding audio as fake tokens
• parse audio duration server-side for WebM/MP4/WAV/OGG and reject invalid or overlong clips (10-minute default, LISA_VOICE_MAX_SECONDS override)
• gate signed-in cloud transcription through paid-only account admission, a renewable turn lease, global spend guards, and idempotent settlement
• price ElevenLabs Scribe v2 at $0.22/hour and OpenAI Whisper at $0.006/minute, then apply the existing 1.4x face-price margin
• move the ElevenLabs default from scribe_v1 to current scribe_v2
• route dictation polish through normal LLM admission/settlement; estimate conservatively if a provider omits token usage
• include media spend/duration/operation counts in the billing usage response
• preserve local/BYO-key behavior and audit operator-funded shared-token cloud usage

Dependencies

Stacked on #318, which is stacked on #317. Merge order: #317 → #318 → this PR, retargeting each successor to main as its parent merges.

Pricing sources

• OpenAI Whisper: https://developers.openai.com/api/docs/models/whisper-1
• ElevenLabs Scribe: https://elevenlabs.io/pricing/api?price.section=speech_to_text

Validation

• npm run typecheck
• focused media/voice/admission tests — 27 passed
• npm test — 1456 passed, 1 skipped, 0 failed
• npm run build
• npm audit --omit=dev — 3 pre-existing moderate findings in the Google GenAI/MCP/Hono chain; no finding in music-metadata