whisper-transcriber - local ASR and agent tooling

A general-purpose long-form-audio transcription pipeline (works over meetings, talks, or any recorded audio): faster-whisper with GPU/CPU fallback and auto-retry, per-device model caching, and explicit VRAM lifecycle management, plus a Claude Code retrieval skill over the resulting transcripts.

Also includes a shelved experiment in local-LLM transcript cleanup (Ollama, 7B models quantised to fit an 8GB RTX 3070): the models kept summarising content when asked only to edit it, despite iteratively strengthened prompts and a comparative evaluation harness judged by a frontier LLM. Kept as an honest capability finding rather than hidden - models that size could not reliably follow even a simple edit-without-condensing constraint. I switched to Claude Code retrieval skills over the raw transcripts instead, which work well.