fix(highway_3d): recover from WebGL context loss instead of crashing on alt-tab (#790)

Switching the active window / alt-tabbing away (most often on Windows) can
trigger a GPU context reset. The 3D highway's WebGL renderer had no
webglcontextlost handler, so a lost context was left to escalate into a
render-process crash -- matching the intermittent "randomly crashes when I
change windows" desktop reports.

The renderer now binds webglcontextlost/webglcontextrestored on its own WebGL
canvas (ren.domElement): the loss is preventDefault()'d so the browser keeps the
context restorable, draw() bails while the context is down so no GL work runs on
a dead context, and on restore the viewport is re-applied and rendering resumes
(Three re-uploads scene resources on the next frame). Listeners are removed in
teardown.

Root cause is a strong hypothesis -- the crash is intermittent and
unreproducible -- but the fix is low-risk and additive and closes a real gap:
there was no context-loss handling anywhere in the renderer.

plugins/highway_3d 3.31.2 -> 3.31.3. Tests:
tests/js/highway_3d_context_loss.test.js (source-contract, like the other
highway_* tests). The sibling keys_highway_3d / drum_highway_3d renderers share
the same gap -- follow-up in their repos.


Claude-Session: https://claude.ai/code/session_01UR2Cr7GEu3yMY7SrfxH6c1

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -16,6 +16,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Player frame-time hotspots removed (trace-backed) + weak-hardware hardening.** A Chrome performance trace of a 3D-highway session surfaced two core per-frame layout-thrash sources, now fixed: the highway's visibility check read `canvas.offsetParent` every rAF frame (forces style/layout recalc — now sampled every 10th frame with a cached value, force-refreshed on init/canvas-replace/resize/override-clear), and the v3 player chrome loop called `matches(':hover')` per frame and unconditionally rewrote the Up-Next pill's `textContent`/bar width at 6 Hz (now hover-tracked via mouseenter/mouseleave, DOM writes only on value change, progress bar moved from `width` to compositor-only `scaleX`). The 3D highway pre-warms shader programs (`ren.compile`) and deterministic label textures at init — and chart-dependent chord/section label textures on first draw — so first-appearance shader-compile/texture-upload frame spikes move into the load spinner. For weaker hardware: the per-frame renderer bundle is now a single reused object instead of a fresh ~35-field allocation per frame (object identity is stable and meaningless; array fields still swap reference on chart changes), custom viz get `bundle.lowerBoundT`/`bundle.lowerBoundTime` binary-search helpers for visible-window culling, the default 2D highway's beat lines no longer scan every beat in the song per frame, and the 3D highway stops reading `localStorage` per frame (1 Hz poll) and caches its lyrics text-measurement layout per displayed line instead of re-measuring every syllable every frame. A second, throttled-CPU trace pass additionally removed: shader-program re-resolution churn from label texture swaps (`material.needsUpdate` is now only set on a null↔texture transition — swapping between two cached label textures never changes the compiled program), the 3D highway's per-frame `getBoundingClientRect` layout read in its canvas-size self-check (now every 10th frame, still immediate on backing-store change), and the core 60 Hz HUD clock rewriting `textContent` on every tick (now write-on-change, ~1/s). The dominant residual — steady `getParameters` shader-program re-resolution (~4% of throttled main thread) — turned out to be Three r158+'s transparent-DoubleSide two-pass rendering, which sets `material.needsUpdate` twice per object per frame; all 18 of the 3D highway's transparent DoubleSide materials are flat unlit quads (labels, rails, chord frames, lanes), so they now declare `forceSinglePass: true`, eliminating the recompile churn and halving those objects' draw calls.
 
 ### Fixed
+- **3D Highway: recover from a WebGL context loss instead of crashing on alt-tab.** Switching the active window / alt-tabbing away from the app (most often on Windows) can trigger a GPU context reset; the 3D highway's WebGL renderer had **no `webglcontextlost` handler**, so a lost context was left to escalate into a render-process crash — matching the intermittent "randomly crashes when I change windows" desktop reports. The renderer now binds `webglcontextlost`/`webglcontextrestored` on its own WebGL canvas (`ren.domElement`): the loss is `preventDefault()`'d so the browser keeps the context restorable, `draw()` bails while the context is down so no GL work runs on a dead context, and on restore the viewport is re-applied and rendering resumes (Three re-uploads scene resources on the next frame). Listeners are torn down with the renderer. `plugins/highway_3d` → 3.31.3. Tests: `tests/js/highway_3d_context_loss.test.js`. (The sibling `keys_highway_3d` / `drum_highway_3d` renderers share the same gap — tracked as a follow-up in their repos.)
 - **Guitar Pro 6 (`.gpx`) import no longer fails on every real file.** The GPX BCFS container reader (`lib/gp2rs_gpx.py`) rejected any file whose final sector wasn't a full `0x1000` block — but a real `.gpx`'s BCFZ-declared decompressed size isn't sector-aligned, so the last (small) container file always lands in a partial trailing sector. The bounds check *raised* `GPX BCFS sector pointer out of range (malformed file)` instead of clamping the tail read, so `_load_gpif` threw before `score.gpif` could be extracted and **no GP6 file could be imported into the song editor** (both real test files failed identically — this wasn't file-specific). GP7/GP8 `.gp` files were unaffected — they take the ZIP path, not BCFS, which is why prior GP-import work didn't surface it. The reader now **clamps the final sector read to the buffer end** (the per-file size field trims the padding anyway), matching canonical GPX readers (alphaTab / PyGuitarPro); a sector whose *start* is past the end still raises, preserving the malformed-file guard. Verified against two real GP6 files — both now unpack to valid GPIF with all tracks. Tests: `tests/test_gp2rs_gpx.py` (partial-final-sector round-trip, multi-file container, sector-aligned baseline, and the preserved out-of-range guard).
 - **v3 Songs grid: fixed the scroll stutter that "skips every so many scrolls," up or down.** The virtualized grid rebuilt its **entire** visible window (`grid.innerHTML = …` + a full `wireCards` pass) every time it slid by one row, so each row-boundary crossing was a heavy synchronous frame that stalled the main thread and buffered held-arrow key-repeats into a visible lurch (a tester's "super fast for a second then slowed down") at fixed scroll offsets — in **both directions and regardless of whether the page was already loaded** (the cost was DOM teardown, not fetching, which is why scrolling back up over cached songs hitched too). `renderWindow()` now **reconciles the window in place**: it reuses the card nodes that stay on-screen and builds only the row that enters/leaves (~6 nodes per slide instead of ~60), keyed by absolute index with a real-vs-skeleton + select-mode signature so hole-fills (after a page fetch) and select-mode toggles still rebuild exactly the nodes that changed. `wireCards`'s `data-wired` guard then wires only the freshly-built nodes, so per-slide listener churn drops with it. Follow-up to the stage-2 virtualized grid (got-feedback/feedBack#636 item 3). Frontend-only: `static/v3/songs.js`. Tests: `tests/js/v3_songs_window_recycle.test.js` (window stays `[start,end)` contiguous + in-window node identity reused across a down-then-up scroll; select-mode toggle and rail-seek jump rebuild correctly).
 - **Starter content seeds again (and now ships The Adicts' "Ode to Joy").** `_BUILTIN_STARTER_SOURCES` still listed `beethoven-ode_to_joy.feedpak` after that pack was deleted, and never wired up its replacement `the_adicts-ode-to-joy_vst_cover.feedpak` that landed on disk. The listed-but-missing file made the all-present gate never fire, so **no** starter songs seeded on first run. Synced the manifest to what's on disk (Für Elise, Star Spangled Banner, The Adicts' Ode to Joy). Tests: `tests/test_builtin_starter_seed.py` (the present/unlisted guards were red on `main`).