bug(e2e): Playwright runs but PixiJS canvas never attaches to #pixi-container in CI

[hyperpolymath]

Surface

After PR #114 fixed the Playwright infrastructure (vite optimizeDeps scoping + port alignment + testDir + DCO + symmetric path filters), the workflow finally runs end-to-end on all three browser matrices. Tests execute, but 17 of 19 fail with the same root cause:

TimeoutError: locator.waitFor: Timeout 15000ms exceeded.
Call log:
  - waiting for locator('#pixi-container canvas')

  20 | async function waitForPixiCanvas(page, timeoutMs = 15000) {
  21 |   const canvas = page.locator('#pixi-container canvas');
> 22 |   await canvas.waitFor({ state: 'attached', timeout: timeoutMs });

#pixi-container exists in the served index.html (verified via the SPA shell), but no <canvas> element is ever attached inside it. PixiJS bootstrap is silently failing or not running in the CI environment.

What we know

• vite dev-server boots cleanly (no unresolved imports — fixed in fix(e2e): Playwright fundamental — vite alias + ums sub-build (closes #113) #114)
• Playwright connects to http://localhost:1984 and serves the page (port aligned in fix(e2e): Playwright fundamental — vite alias + ums sub-build (closes #113) #114)
• #pixi-container div renders (DOM-loaded)
• One test passes ("[browser] › tests/performance.spec.js" line not shown — likely the simplest non-canvas-dependent assertion)
• All canvas-dependent tests fail identically across chromium / firefox / webkit
• Including tests/game-loads.spec.js:47:3 › Game loading › page loads without JavaScript errors — so there ARE JS errors during boot

Likely root causes (in order of probability)

1. PixiJS init throws in headless / WebGL environment — but webkit failing too is unusual; WebGL2 / WebGL fallback may be broken
2. Asset path mismatch in dev mode — wasm modules, sprite manifests, or audio files served with wrong content-type in vite dev (vs production build)
3. AffineScript-TEA bootstrap — vite.config.js has 'affinescript': '/src/app/tea' alias + optimizeDeps.entries: ['index.html']. If the SPA mount depends on TEA, the AOT compilation gap (per affinescript#57 status) may break dev-time boot
4. deno task res:build in workflow does compile-only — the workflow uses dev:vite (which compiles ReScript on-demand) instead of deno task build (full production build). If the asset pipeline (AssetPack) hasn't run, sprite manifests are missing
5. AFFINE_TEA_MIGRATION: true / USE_WASM_ROUTER: true flags in vite.config.js — if the wasm router fails to fetch its .wasm file (router.wasm), the boot pipeline aborts

Investigation recipe

1. Download a failed run's trace.zip artifact and inspect npx playwright show-trace
2. Check the browser console screenshot in test-results/*/test-failed-1.png
3. Run deno task dev:vite locally, point a browser at http://localhost:1984, observe DevTools console
4. If canvas works locally but not in CI: bisect by enabling/disabling AFFINE_TEA_MIGRATION flag
5. If canvas doesn't work locally either: bisect on commits since 2026-05-02 (when Playwright started failing per Playwright E2E: vite alias + idaptik-ums sub-build fundamental fix #113)

Cross-references

• Playwright E2E: vite alias + idaptik-ums sub-build fundamental fix #113 — original Playwright "make it run" issue (closed by fix(e2e): Playwright fundamental — vite alias + ums sub-build (closes #113) #114; this issue is the natural follow-up)
• fix(e2e): Playwright fundamental — vite alias + ums sub-build (closes #113) #114 — the PR that surfaced this; the infrastructure layer is now fully fixed
• PR fix(e2e): Playwright fundamental — vite alias + ums sub-build (closes #113) #114 is being admin-merged despite this issue because Playwright is NOT in the required-on-main checks list (per gh api repos/hyperpolymath/idaptik/branches/main/protection), so its red doesn't actually block merges
• affinescript#57 — Phase 3 migration assistant (relevant to root cause Quickdraw test #3 above)

Severity

P2 — Playwright never gave a useful signal anyway (it's been red since 2026-05-02), so this isn't a regression. But it's worth resolving so the workflow becomes the load-bearing E2E gate it was meant to be.

[hyperpolymath]

Closed by #121 — the original preload bundle 404 / canvas-mount issue is fixed (12 tests now pass).

Residual 6 failures (a11y, perf threshold, console errors) — pre-existing latent failures previously masked by the preload error — tracked in #122 as scope-creep follow-ups.