[ctoth]

Replaces the MIDI.js sample renderer with a SoundFont synthesizer (spessasynth_lib) that plays one bank, MuseScore General (MIT), and adds the script that builds that bank.

Client

• src/midi/SoundfontOutput.ts is now a WorkletSynthesizer behind the same MidiOutputSink (send, sendAt, close).
  • ready() (:36) checks that the client audio output exists. It then loads the worklet with addModule from a Vite ?url import of spessasynth_lib/dist/spessasynth_processor.min.js (:2, :39), creates the synth on virtualMidiService.getAudioOutput().context and connects it to that output's destination (:41-42).
  • The bank downloads in the background from https://mongoose.world/sounds/soundfonts/MuseScore_General.sf3, with credentials: 'omit' (:49).
  • Until it has loaded, every message goes to the built-in synth (:87). Program, controller and pitch-bend messages sent in that time are kept, the latest per channel and controller, and replayed into the synth once the bank loads (:59).
  • After the switch, note-offs and CC 64/120/123 also go to the built-in synth, so notes it started still end (:93).
  • sendAt(data, s) calls sendMessage(data, 0, { time: context.currentTime + s }) (:95).
  • addSoundBank never resolves on a bank the worklet rejects; the worklet posts soundBankError instead (lib basic_synthesizer_core.ts:386-391). So the load also listens for that event (:55-56).
  • If the bank fails to load, the built-in synth stays in use and the status says so.
  • close() aborts the download, destroys the synth and closes the built-in synth (:110-115).
• src/midi/soundfonts.ts is down to SOUNDFONT_OUTPUT (device id unchanged) and SOUNDFONT_NAME.
• Removed: the three MIDI.js collections, src/midi/programs.ts, SoundfontPicker.tsx and .css, and the picker in preferences.tsx.
• The MIDI tab's sampled-instruments card now describes the one bank and links the MuseScore General license (MidiStatus.tsx:599-611).
• src/stores/preferencesStore.ts: removed the midiJsSoundfont field and its normalisation (:174). There is no migration and no STORAGE_VERSION change; a leftover midiJsSoundfont key in saved data is ignored. The enabled default, STORAGE_VERSION 3 and the migration from Make MIDI automatic: on by default, output connects on the first MOO message #202 are untouched (this branch merges master at f955956).
• MidiService.ts:294,300,314: new constructor, and the device name is now Soundfont Synthesizer · MuseScore General.
• vite.config.ts:37: the worklet file is a ?url asset, so optionalChunks (which only drops chunks, optionalChunks.ts:33) can't keep it out of the **/*.js precache. A globIgnores entry keeps it out of the service-worker install. It is 402,596 bytes.
• package.json pins spessasynth_core 4.3.22, spessasynth_lib 4.3.14 and stb-vorbis 0.0.6 exactly. Both libs declare their own dependency as "latest"; npm ls shows both nested dependencies deduped onto these pins.
• scripts/check-soundfonts.mjs (run in CI by test:audio-cache) now serves core's one-preset sample bank (BasicSoundBank.getSampleSoundBankFile(), :30) at the bank URL (:31, :38). The first request returns 503 and the check expects "MuseScore General could not load". After Reconnect it expects "MuseScore General ready." The collection loop is gone. I have not run this script locally; CI runs it.

Bank build: scripts/build-soundbank.mjs

node scripts/build-soundbank.mjs <outDir> [--source-dir <dir>] [--quality <q>]. Needs ffmpeg with libvorbis. It writes to <outDir>:

• MuseScore_General.flac.sf3, the FLAC master: core's setSampleFormat({format:"compressed"}) with ffmpeg FLAC, written to a temp file so STREAMINFO is filled in (:216-218). This is SF3-shaped with FLAC payloads, for rebuilding, not playback.
• MuseScore_General.sf3: ffmpeg libvorbis at each sample's own rate, quality 6 (:247-249).
• MuseScore_General_License.md, copied from the upstream directory (:344).

Gates (any failure exits 1):

• every sample really compressed (core's compressSample swallows encoder errors);
• the FLAC master decodes byte-identical to the source PCM;
• the structure of both outputs (presets, zones, generators, modulators, samples, loop points) matches the source;
• every SF3 sample's stb-vorbis decoded length equals the source length.

A longer decode is allowed only when both of these hold (loopOnlyModes :171, alignment :288):

• offset 0 has the lowest RMS error of all shifts in [-64, 64];
• every instrument zone using the sample has sampleModes & 3 == 1 and generators 1/12/3/50 all 0.

This covers an stb-vorbis 0.0.6 final-frame trim bug (wasm/stb_vorbis.c:3413-3443): ffmpeg's last granule and ffmpeg's own decode both equal the source length for the three samples below.

Run on MuseScore_General.sf2 0.2.0 (215,614,036 bytes), exit 0:

source: MuseScore_General.sf2 215614036 bytes, 309 presets, 204 instruments, 1246 samples
GATE PASS: every sample FLAC-encoded
GATE PASS: FLAC master decodes byte-identical to source PCM (1246/1246 samples identical)
GATE PASS: FLAC master structure matches source
GATE PASS: every sample Vorbis-encoded
GATE PASS: SF3 presets, generators, modulators and loop points match source
  allowed longer decode: sample 79 Cello-G#1: source 37067, decoded 37376, rms at offset 0 -45.5 dBFS, best other offset 1 -32.3 dBFS, zones: Cello_F/Cello_M/Cello_P mode 1 offsets 0/0/0/0
  allowed longer decode: sample 80 Cello-F1: source 33493, decoded 33920, rms at offset 0 -46.2 dBFS, best other offset 1 -34.1 dBFS, zones: Cello_F/Cello_M/Cello_P mode 1 offsets 0/0/0/0
  allowed longer decode: sample 599 OcarE1: source 60416, decoded 60544, rms at offset 0 -43.9 dBFS, best other offset -1 -34.1 dBFS, zones: Ocarina mode 1 offsets 0/0/0/0
GATE PASS: every SF3 sample decodes (stb-vorbis) to the source length, or longer only where it is aligned and never played past the loop end (1246/1246 samples, 3 allowed longer)
MuseScore_General.flac.sf3: 95679054 bytes (91.25 MiB)
MuseScore_General.sf3: 32986146 bytes (31.46 MiB)
all gates passed

The structure counts all match (modulators 1769, samples 1246, loop points 1246, instruments 204, zones 3859, generators 11854, presets 309). The worst whole-sample Vorbis error is Applause at -29.8 dBFS RMS; the worst within 64 samples of a loop point is Applause at -32.3 dBFS.

The bank is not uploaded by this PR. Until MuseScore_General.sf3 is at the URL above, Soundfont Synthesizer stays on the built-in synth and says the bank could not load.

Line endings

src/MidiService.ts had mixed CRLF and LF line endings on master. Editing it rewrote every line as CRLF, so the diff shows about 160 changed lines. git diff --ignore-cr-at-eol shows the real change is 9 lines.

Tests

src/midi/SoundfontOutput.test.ts is rewritten against a mocked spessasynth_lib: it failed 8 of 8 against the old renderer and passes 8 of 8 now. After the merge, SoundfontOutput.test.ts and preferencesStore.test.ts pass 26 of 26. I ran only those two files; CI runs the rest.