Add JamesDSP integration plan

plans/jamesdsp-integration-plan.md

@@ -0,0 +1,34 @@
+# JamesDSP in ytplayer — integration plan (2026-10-09)
+
+## What exists
+- **Standalone PWA/port**: `~/development/personal/jamesdsp-web` (GPL-3 port of JamesDSPManager's portable core to WASM). `build.sh` -> `web/dsp/jamesdsp.{js,wasm}` (1.2 MB wasm), ES modules `web/jamesdsp-node.js`, `params.js`, `worklet.js`. Docs: `INTEGRATION.md`, `PLAN.md`, `VERIFICATION.md` (Chromium-only verified: 44.1/48 kHz, 15-band EQ, compander, bass, convolver, stereo, reverb, crossfeed, tube, liveprog/DDC/arbitrary EQ; fixed 128 MiB WASM heap per node; no real-phone/Safari testing).
+- **Prototype in ytplayer**: branch `codex/jamesdsp` (3 commits on old base 93d3b05): vendor/jamesdsp/{controller,jamesdsp-node,params,worklet}.js + wasm, `JamesDSP` IIFE in app.js, Settings group (enable, 5 presets, 3 EQ sliders, bass, crossfeed, limiter), sw.js precache, beta compose stack (`docker-compose.beta.yml`, beta.worship.hesed.sbs, already a deploy target "codex/jamesdsp"), `tests/jamesdsp.spec.js`.
+
+## Problems the prototype does not solve (why it can't just be merged)
+1. **Second AudioContext + second MediaElementSource**: app.js already owns `EQ` (app.js ~5500: 10-band biquad EQ, Level/limiter, VocalReducer, widener) which calls `createMediaElementSource` on `els.video`/`els.audio`. An element can have only ONE source, ever. The two systems fight; whichever attaches first wins and the other throws/silences.
+2. **iOS background playback**: Safari suspends Web Audio when the screen locks; the existing EQ is disabled on iOS unless `liveAudioIOS` is on, and iOS saved songs get EQ rendered into a copy (`EqRender`, `eqAudioUrl`). The prototype ignores this -> likely silence on lock screen.
+3. **Dual elements / audio-continuity** (`audio-continuity.js`, master/secondary, `Player.soundEl`) and crossfade/preload swaps.
+4. **Assets**: now manifest-driven, lazy `assets.json` groups with `contract`, N-1 retention, offline completion job, per-file hashed URLs. The wasm/worklet must be a lazy group (`feature:jamesdsp`), in the manifest, cached for offline, never in first paint/install-critical set, and fetched only after opt-in (or idle-prefetch only on Wi-Fi). The worklet module URL must be a hashed immutable URL resolvable offline; wasm bytes are fetched on the main thread and passed via processorOptions (already designed so).
+5. **CORS/proxy**: streaming `<audio>` must be same-origin (`/api/...` proxy) or have `crossOrigin='anonymous'` set before src; OPFS blob URLs are fine. Verify every playback path (stream, saved/OPFS, P2P, rendered eqAudioUrl).
+6. **Memory/CPU**: 128 MiB heap per node; one node only, created lazily, destroyed on disable; low-end phone CPU for FIR EQ/reverb/convolver; settings changes suspend audio briefly.
+7. **Settings persistence**: new `settings.jamesDsp` must survive updates/profile sync (see settings-persist harness `perf/settings-persist.mjs`), default OFF.
+8. **Licence**: DSP port is GPL-3.0-or-later; ytplayer's licence/distribution must be checked and `frontend/vendor/jamesdsp/LICENSE.txt` + corresponding-source pointer + attribution in Settings/About. Owner decision flagged below.
+
+## Design decision (Claude, final)
+**One audio graph, one context.** Do not add a second pipeline. Refactor the existing `EQ` graph into a small shared "AudioGraph" owner (single AudioContext, one MediaElementSource per element, retained) with an ordered insert slot: `source -> [JamesDSP node OR legacy EQ/Level/Vocal/Widener chain] -> destination`. Modes (Settings -> Playback & sound -> "Sound engine"): **Standard** (today's EQ, default) | **JamesDSP** (replaces the 10-band EQ/vocal/widener while on; Level/limiter remain via JamesDSP's own limiter+gain, loudness gain preserved). Never both EQ chains active -> no double processing. Bypass = exact (setBypass) and, when engine = Standard, JamesDSP is not loaded at all.
+iOS: JamesDSP live only when `liveAudioIOS` (same opt-in as today's EQ) because of lock-screen suspension; for iOS saved songs offer the existing render-into-copy approach only if feasible via OfflineAudioContext + the same worklet (investigate; otherwise document unsupported and show the warning). Android/desktop: live.
+UI: Settings group (keep prototype controls but extended: preset list, 15-band EQ graph/sliders, bass, stereo, crossfeed, reverb preset, tube, limiter, output gain, compressor on/off, A/B bypass button, import preset JSON, IR upload, Reset), themed per layout, works in all 4 layouts + dark/light/contrast. No Liveprog/DDC text effects in v1 (trusted-code risk) — hidden behind an "Advanced" flag off by default.
+
+## Phases (single Codex thread, sequential, one commit per step, branch `codex/dsp-integrate` from main; port the useful prototype code, do not merge the old branch wholesale)
+1. Audit + design note: `plans/jamesdsp-audit.md` — map every `createMediaElementSource`/AudioContext/`liveAudioIOS`/EqRender/audio-continuity/Level/VocalReducer use; confirm single-source constraint; decide the AudioGraph seam; list iOS facts.
+2. Refactor: extract AudioGraph (single ctx/source owner) with the Standard chain unchanged; all existing EQ tests + new unit tests pass; behaviour identical (EQ harness, level tests).
+3. Assets: vendor files from the standalone repo (pin the commit hash in `frontend/vendor/jamesdsp/VERSION`), add `assets.json` group `feature:jamesdsp` (+contract), manifest entries, sw precache removal in favour of the manifest sync, offline-complete + lazy + migration harnesses green, first-paint bytes unchanged (assert `perf` cold-start numbers do not regress).
+4. Engine: JamesDSP controller on AudioGraph, lazy create, one node, error -> auto fall back to Standard + toast (never silence), 44.1/48 rate check (create the AudioContext at 48000 where allowed; otherwise fall back), suspend/resume handling, destroy on disable, pause-safe config changes (never while `Player._wantsPlaying` glitch-check: apply with short mute ramp).
+5. UI: Settings group + themed controls + preset library (Flat, Bass boost, Vocal, Loudness, Headphone crossfeed, Bright, Warm), A/B bypass in the Now Playing "more" menu; persisted `settings.jamesDsp`; included in profile sync + import/export; settings-persist harness extended.
+6. Verification: Playwright (Chromium; WebKit where Web Audio worklets work) — measured RMS change per effect with a real clip, bypass exact, switching tracks/elements/crossfade keeps processing, saved/OPFS + streamed + P2P sources all processed, offline reload with engine on, error injection (corrupt wasm) falls back dry, memory (one 128 MiB heap only), CPU timing for worst-case preset, `node --test`, manifest/hash, migration, ui-geometry/icons all green. Write `plans/jamesdsp-integration-report.md` with limits (iOS lock-screen, real-ear tuning).
+7. Beta: push branch only; deploy to beta.worship.hesed.sbs via app-deploy (branch codex/dsp-integrate) for owner listening tests BEFORE merging to main. Main merge only after owner OKs on beta.
+
+## Owner decisions needed (defaults chosen)
+- Licence: GPL-3 inclusion in a (possibly public) repo — default: keep ytplayer repo as is, vendor under its own LICENSE, add attribution; confirm you are OK shipping GPL code in this app.
+- iOS: default off unless `liveAudioIOS` opt-in (matches today's EQ). 
+- Default engine stays Standard until you approve on beta.