Porting to the work M5 (16 GB)

The point of the exercise. Every number in Measurements came from the home machine — M5 Max, 128 GB — and the target is a base M5 with 16 GB. Chunked at 30 s this should fit, but should is not measured, and that is one of the two things still open.

The checklist, in order

1. rustup toolchain, then git clone, ./fetch-models.sh, cargo build --release.
2. cargo run --release -- probe — confirms the Core Audio bindings link and reports the execution provider state. This needs no permissions.
3. cargo run --release -- transcribe <model> <recording> on a real recording, run under /usr/bin/time -l, and check maximum resident set size stays near the figure in Weight prepacking (~1.3 GB). If it does, the 16 GB machine is fine for batch transcription.
4. Then ./make-app.sh, and run the tap via open -a.

What to expect at step 4

If the tap returns silence, that is TCC and not a bug. See capture for the two independent failures and the truth table that separates them — the first thing to check is the launch method, because a terminal launch fails identically no matter how the app is signed.

On a managed Mac this is exactly what a PPPC profile from Jamf or Intune exists to grant. Ask for System Audio Recording. You do not need Screen Recording — that is the whole reason this uses a process tap rather than ScreenCaptureKit, and asking for the larger permission would give away the argument. A PPPC profile denying kTCCServiceAudioCapture produces symptoms identical to a stale cdhash, so rule the signing identity out before raising a ticket.

The number that decides it

The memory budget is the binding constraint, not throughput:

~1.3 GB during a transcription burst, chunked at 30 s with weight prepacking off; negligible while capturing.

That figure is from Weight prepacking; the ~2.2 GB in Speech recognition predates it.

Unchunked, memory scales linearly with audio length — an extrapolated ~35 GB for a 30-minute session, which fails on this machine outright. So step 3 is not a formality: it is the whole port. At ~20–30× realtime on CPU with prepacking off there is no throughput question to answer.

Diarization is a separate concern and a much smaller one — 249 MB peak on a 10.2-minute track, because the sliding window does not accumulate. It is a separate verb partly so its two graphs are never resident alongside Parakeet, which is what keeps the 16 GB target reachable. See ADR-0007.

Still open

1. The base M5 (16 GB). Every number is from the 128 GB machine.
2. Tap creation under the work machine’s TCC policy. Enumeration needs no permission; creating a tap needs System Audio Recording.