From a multi-agent bug-hunt + adversarial verification pass (0 critical/
high; 15 mediums). Fixed 13; 2 deferred as bigger refactors.
Correctness:
- srs: "Hard" no longer overshoots "Good" for reviewed cards (reps>=2). It
compounded ease AND x1.2; now grows x1.2 off the previous interval only.
+ regression test.
- enrich/dates: stop reading the modal verb "may" as the month May (phantom
calendar events). "may" needs an ordinal/year/date-preposition now.
- enrich/bib: disambiguate colliding BibTeX keys (smith2020, smith2020a, …)
— duplicates corrupted reference-manager imports.
- learn/glossary: junk-term guard used && (dead); now || so all-stopword
terms like "there" are actually skipped.
- transcription/engineImpl.web: don't collapse Whisper's null end-timestamp
to a zero-length [t,t] segment; estimate from the next chunk or window
duration (fixes citation/seam anchors).
- transcription/pipeline: re-check the abort signal AFTER each chunk so
Cancel works on single-chunk audio (it previously still saved).
- db/repo.native: use withExclusiveTransactionAsync for reassign /
deleteCourse / upsertVectors / createFlashcards (withTransactionAsync is
not isolated on a shared connection → interleaved/half-applied writes).
- db/repo.native: don't cache a rejected open/migrate promise — a transient
first-open failure no longer bricks storage for the whole session.
- stores/transcriptsStore: sequence-guard refresh() so overlapping
focus/typing/filter refreshes can't resolve out of order and show stale
results.
- audio/wav: decode WAVE_FORMAT_EXTENSIBLE (0xFFFE) via its SubFormat GUID
+ add 24-bit PCM — common ffmpeg/Windows WAVs no longer hard-fail native
import. + tests.
Performance / footprint:
- audio/decode.native: decode straight to 16kHz (decodeAudioData sampleRate
hint) so the JS side never holds a full-rate buffer or runs the resample
loop — big memory/OOM win on long lectures.
- models/catalog display: Settings model sizes are now backend-aware (the
no-GPU WASM path pulls ~2x fp32 weights; was advertising ~half).
Feature gap:
- download: native exports were silent no-ops. New download.native.ts writes
to cache + opens the share sheet (expo-sharing); transcript/ICS/Anki-CSV/
BibTeX/RIS exports now work on device.
Deferred (bigger): lexical-search recall ceiling (needs a full-corpus rank
path on both repos); triple in-memory copy of the encoded file during
transcribe (media is keyed by a not-yet-existing transcript id).
Validated: tsc clean, 282 tests pass, web export clean (native deps not
bundled), arm64 APK compiles with expo-sharing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The WASM backend couldn't create a session — "TransposeDQWeightsFor
MatMulNBits Missing required scale" — so transcription failed to start on
mobile (which we route to WASM) and on any no-WebGPU device. Root cause:
the default quantized Whisper decoder (q8/q4, decoder_model_merged) uses
MatMulNBits ops that the onnxruntime-web bundled with transformers.js
4.2.0 cannot load on WASM. Reproduced and bisected in a browser harness:
across both Xenova and onnx-community repos, q8/q4/string-fp32 all fail on
WASM with this error, while an EXPLICIT per-file fp32 decoder
({ encoder_model: 'fp32', decoder_model_merged: 'fp32' }) loads and runs
(2/2 chunks, no error) on a no-GPU machine.
So WASM now requests that explicit fp32 decoder; WebGPU keeps fp16. Larger
download on the WASM path, but it actually loads and runs — q8 never did
on WASM. (4.2.0 is the latest transformers.js, so bumping isn't an option.)
tsc clean, 279 tests pass, web export OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two issues behind "it just shows downloaded 100% / loads forever / crashes
after a couple seconds":
1. Progress was invisible during transcription. The 'transcribing' stage
only fired AFTER the first (slow) chunk, so the UI sat at "Loading
model… 100%" through the entire first inference and looked frozen.
- pipeline now emits a transcribing kickoff (progress 0) BEFORE the
first chunk and carries chunkIndex/chunkCount on every event.
- transcribeStore threads those through; the Library job card shows a
spinner, "Transcribing… N%", and "part i of N" so a long file's
progress is legible and obviously advancing.
2. Web crash on mobile. We picked WebGPU + fp16 whenever navigator.gpu
existed, but mobile WebGPU drivers crash on sustained fp16 Whisper
inference (transcribes a chunk, then the GPU process dies). Mobile web
now uses cross-origin-isolated multi-threaded WASM (slower but stable);
desktop keeps WebGPU.
Native's "loads forever" was the same missing-feedback problem — it was
grinding through chunks with no UI signal; the chunk counter now shows it.
tsc clean, 279 tests pass (pipeline progress test updated for the kickoff
event), web export OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Whisper ".en" models reject `task`/`language`; transformers.js throws
"Cannot specify `task` or `language` for an English-only model" if either
is passed — and the web engine passed `task: 'transcribe'` (plus a
possibly-undefined `language`) unconditionally. Since the default model is
tiny.en, web/PWA transcription failed for every default user.
Only add `task`/`language` for multilingual models now (translation also
requires a multilingual model), and never pass an explicit `undefined`
language (which trips the same check). Mirror the same gate on native
(whisper.rn) so a .en model is never asked to do a language/translate it
can't.
tsc clean, 279 tests pass, web export OK.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
NilsBriggen