docs(devbook): record the meeting transcription decision (ADR 0049) - #1711

What changed

What the research found (Microsoft Learn, read 2026-11-10):

• AAC in ADTS is not accepted. The documented formats are mp3, mp4, mpeg, mpga, m4a, wav and webm, and "other formats return an error". So the desktop remuxes to M4A without re-encoding.
• The limit is 25 MB per request. The transparency note also gives a length limit of 15 to 30 minutes per synchronous request. That is why the ADR sends a recording in remuxed parts of at most 5 minutes.
• gpt-4o-transcribe supports only json output, with no verbose_json and so no segment timestamps. The required timestamps come from each part's offset instead.

Acceptance criteria

Build & Test

This is a devbook-only change with nothing runnable, so no build or test ran.

• node .devbook/_tools/devbook-meta/build.mjs --check exited 0.
• node tools/devbook/check-metadata.mjs reported 217 files, 0 blocking. The advisory counts are unchanged and none of them are on the touched files.

Review findings left open

None. One fresh-context review found 1 blocker, 4 majors and 6 minors, and all were fixed. The blocker: a phone edit made on the body from before the transcript would overwrite the whole note after the audio was already gone. The fix is that the desktop keeps the sections it wrote, by sha256, and appends them again on pull.

What could not be proved

• The format claims were not tested against the live deployment. No ADTS file was sent to Foundry to confirm it is refused. The 15–30-minute length limit comes from the transparency note, not from a measurement. The 5-minute part length is a judgement inside that limit.
• Timestamps are coarse, about every 5 minutes. The decision asked for timestamps, and the model gives none. Validate that part-offset timestamps are good enough.
• The repair on pull (§6) goes beyond the decision you gave me. It was added to keep the "never replaces" promise once the audio is gone. A second desktop that never wrote the sections cannot repair them.
• The settled decision says nothing about consent. The ADR says configuring Foundry on the desktop is the only consent, and names the three ADR 0023 rules this reverses.

Assumptions

• ADR number 0049, not 0048. 0048 was left for plan entry 1's recording-format ADR. Entry 1 landed it on plan/meeting-recording as "Meeting audio is recorded as AAC in an ADTS stream", and the merge of that base lists both. 0049 now cites 0048 for the phone's format (ADTS, audio/aac, 24 MB parts) and for the audio types a note accepts.
• The desktop reuses AzureFoundryTranscriptionClient, adding an API-key credential. It does not add a second client.
• The summary uses the existing chat deployment, so "configured" means AzureFoundrySettings.IsConfigured.

An unattended execute-plan run opened this draft for plan entry 78334a1c-f50b-429a-b21b-17781572cacf (+meeting-recording, spec-transcription). Personal Validation is pending: validate on this draft.

🤖 Generated with Claude Code