3.8 KiB
JamesDSP integration report — 2026-10-10
JamesDSP now uses the shared AudioGraph, remains opt-in, and is delivered as a lazy feature. The upstream port is pinned to a982b53c4ab347d1d9c50f4ce7ee3c2735eaa817. It adds four-channel convolution, generated and IndexedDB-stored IRs, spatial/HRTF processing, and 8–96 kHz sample-rate conversion. Settings expose IR selection/upload and bounded controls for convolution and spatial audio. Long IRs are constrained on low-resource devices. Standard remains an exact bypass route; failed DSP/WASM initialization returns to Standard. iOS playback remains gated by liveAudioIOS.
The legacy lossy-update timeout came from the new worker hashing/copying the entire legacy shell during activate, while the b77938a page reloads on a short timer after the user presses Refresh UI. The worker now stages and verifies legacy assets during install; activation only commits the staged cache. The lossy migration harness passes both engines, preserving the user-initiated legacy update flow.
The DSP WASM is now marked offlineOptIn: "jamesDsp". Normal completion caches the 103 required files and defers all 9 optional DSP files, including the ~1.2 MB WASM. Enabling “download audio effects for offline” persists the opt-in and resumes completion. Cold transfer is back near the pre-DSP baseline.
| Verification | Result |
|---|---|
| Frontend tests | 271 passed, 0 failed |
| Server tests | Previously run after integration: all pass except the two known yt_dlp environment failures; remaining server tests passed individually |
| Offline completion | Chromium + WebKit: 112/112 after explicit opt-in; 103 required + 9 optional deferred by default; zero missing/failures |
| Lazy / hidden timers | Both browsers pass; lazy asserts DSP WASM remains uncached by default; no hidden-timer page errors |
| UI geometry / icons | 384 geometry cases and 24 layout/theme icon cases passed |
| Settings persistence | Old-client Refresh UI preserved all existing keys in both engines; only intentional additions/migrations are the disabled jamesDsp default and 5→10 EQ migration. Save-before-play works before and after update. DSP preset/parameters and offline opt-in persist. Linux WebKit lacks native OPFS, so its save-play assertion is explicitly unsupported. |
| DSP harness | Both engines pass: four-channel IR changes output (interaural RMS 0.00735), spatial stage changes it (0.02154 vs dry 0), 32 kHz context, exact bypass, offline reload with stored IR, one DSP heap, corrupt-WASM fallback |
Migration (node perf/migration.mjs --browser all) passed lossy and unthrottled after the worker fix. Lossy wire totals: Chromium 652,568 bytes / 26.4 s; WebKit 746,826 bytes / 20.1 s. Unthrottled: Chromium 664,110 bytes / 11.2 s; WebKit 790,706 bytes / 7.7 s. Each run showed one banner/reload, offline launch, and playback guard.
Three-run cold LTE medians (node perf/baseline.mjs --runs 3 --browser all --profile lte --scenario cold):
| Engine | FCP before → now | LCP now | Wire before → now | Requests now | Positive app.js responses |
|---|---|---|---|---|---|
| Chromium | 1360 → 1348 ms | 1348 ms | 648 KB → 1.82 MB → 673,992 B | 117 | 1 |
| WebKit | 895 → 908 ms | 976 ms | 976 KB → 2.14 MB → 1,020,231 B | 161 | 1 |
The wire columns are pre-DSP baseline → initial DSP integration before the opt-in correction → current. The correction removes roughly 1.2 MB from first install while leaving the DSP assets available after explicit opt-in.
The harness JSON is perf/.tmp/dsp-cold-final.json; it records 274 ms Chromium long-task total and 0 ms in WebKit. No real iPhone, lock-screen session, thermal behavior, or listening/ear-tuning test was possible. Verify iOS opt-in, background audio, CPU use with long IRs, and the perceived level/space before enabling this for users. Phase 7 deployment was not run.