ETA from wall-clock throughput instead of per-file transcribe times

[TN019]

Reported from a live run: the bottom bar was predicting 3000+ minutes remaining.

Why it was wrong

The estimate was mean(per-file transcribe time) × files remaining, where the per-file clock ran from the transcribing status to done. Two things fell out of that:

• Every remaining file was priced as a full transcription. A file that already has subtitles is skipped in milliseconds, and nothing in the estimate knew that.
• Nothing outside transcription was counted. The wait in extract_q.get() — iCloud downloads, ffmpeg extraction, a starved transcriber — happens before the status flips to transcribing, so it contributed zero. The one-time whisper model load did the opposite: engine.load() sits inside the timed window, so it landed entirely on the first file and dragged the mean for the rest of the run.

The reported run, measured from history and the log:

finished 624 files in 128 min (20 transcribed, 603 skipped = 97%)
real throughput: 12.3 s/file

files left      old ETA      new ETA
       100       830 min        20 min
       200      1660 min        41 min
       400      3320 min        82 min

A 97%-skip batch was being quoted at roughly 40x its real cost.

The change

ETA = elapsed / files finished × files remaining

Wall-clock throughput accounts for everything the batch actually spends — downloads, extraction stalls, the model load amortised across the run, translation running alongside — and lets cheap files cost what they cost instead of needing a special case. It also deletes more than it adds: _durations, _active_since and _track_eta are gone.

Per-file measurements were considered first (drop the first sample, use a median, pre-count files that will be skipped). Wall-clock subsumes all three and needs no extra I/O over hundreds of iCloud paths at batch start.

Known limit: it assumes the rest of the batch skips at roughly the rate seen so far, so a run whose skipped files are all front-loaded will read low until it catches up — but it does catch up as the mix changes, which the old formula never did in the downward direction.

Also

downloading was missing from the viewmodel's active-status set, so the bottom bar showed no current file while one was being fetched from iCloud — it was added to the run page's list in #33 but not this one.

Tests

201 passed / 5 skipped. Three new: throughput extrapolation, a batch of instant skips staying cheap (the reported shape), and no estimate before the first file finishes while the row still reports as busy.