recording-transcriptionlisted

# recording-transcription ## Overview - The audio is **a file the user handed over** (phone, voice recorder, on-site capture) or text draft. - The result goes to `memorb-ingest` for the impact scan. ## The premise > A phone recording is the single-microphone problem in another guise: whoever sits near the phone comes through clearly, whoever sits far away may be muffled or talked over. **Do not assume the clear parts are the user speaking.** Work out the speaker from the content — question and answer pairs, forms of address, who owns the topic. > For a one-on-one with a manager or a colleague, **wording accuracy and timestamps matter more than they do for an ordinary meeting** — anything touching commitments, numbers, or dates should be marked uncertain rather than smoothed over by the model into something that merely reads well. ## Known environment limits (measured 2026-07-12 — read this before you start) - This sandbox's network allowlist covers package-ecosystem domains such as `pypi.org` and `github.com`, but **not `huggingface.co`**. `pip install faster-whisper` works fine; at run time it wants to fetch model weights from HF and that connection will always fail here. The `check` command detects this up front, so you do not discover it halfway through a background job. - **Any folder mounted from the real machine is append/overwrite-only — nothing can be deleted.** That covers the vault and the `outputs` scratch folder alike: `rm` on a file you have already writte