[MaximumTrainer]

Closes #584 FR1 / FR2 / FR3 (D1-a, D2-a, D3-a, D4-a, D6-c). FR4 (pixel-level tier differentiation) stays per the agreed D5-a — rely on today's content gates first, open a follow-up only if screenshots still look identical after SSAO lands harder.

The problem this fixes

Measured on Dan's Intel Gen-9 laptop against next/svelte-webgpu @ 43433ba: when the WebGPU device is lost, the engine sees it (deviceLostCount → 1, device.lost promise fires) but doesn't act on it. The rAF loop keeps submitting GPUCommandEncoder.finish + queue.submit against the dead device (45 of each in 1.5s), __ROWER3D_CONTEXT_STATE stays 'ready', no banner appears, and the rower sees a visually frozen scene with no sign of what went wrong.

What this PR does

FR1 — the engine actually recovers (D1-a + D2-a + D3-a)

• attach.ts now factors the per-device scene into src/lib/renderer/engine/rideScene.ts (mechanical extraction, same behaviour). attach keeps the device-session lifecycle, outer rAF loop and telemetry.
• On the first device.lost:
  • publishFallback('device-lost')
  • __ROWER3D_CONTEXT_STATE.kind → 'lost', backend stays 'webgpu'
  • banner (__ROWER3D_BANNER) → "Lost the graphics context — restoring…" (D2-a copy ported verbatim from the retired WebGL spec)
  • current scene disposed, rAF loop stopped (no more work to the dead device)
  • fresh device requested with isRetry: true; on success a new scene is built on the same canvas (D1-a) and the loop resumes; context state back to 'ready', banner cleared.
• On the retry's device also dying: requestDevice's existing isRetry path fires onDeviceLost('renderer-unavailable'), which maps to:
  • __ROWER3D_CONTEXT_STATE.kind → 'unavailable' with reason 'renderer-unavailable'
  • banner → "We couldn't get the graphics context back. Your row is still being recorded." (D3-a)
• /row reads __ROWER3D_BANNER on a 250 ms interval and renders an overlay at the top of the stage with --restoring (amber) and --failed (red) variants. The row keeps counting underneath because ride.currentSession is DOM-driven.

FR2 — forced loss, test-only (D4-a)

__ROWER3D_FORCE_DEVICE_LOST() is installed on window only when __PLAYWRIGHT_TESTING is set. Points at the current scene's device, calls device.destroy(). Replaced on every scene build so a reattach exposes the new device, cleared on dispose.

FR3 — timing (D6-c)

• device.ts now asks for the timestamp-query feature when the adapter has it (same shape as the existing BC-compression opt-in).
• __ROWER3D_FRAME_STATS.cpuMs is populated every frame — measured end-to-end around frameOnce with performance.now().
• gpuMs stays 0 until the frame-graph instrumentation lands. Follow-up: wire timestampWrites on each pass + a resolve/map readback. Called out in the issue so a reader can tell it apart from the no-timing case.

Tests

Flipped the it.todo('reattaches and flips contextState to lost on first device loss') from #585 to a real passing test; added:

• reattaches and flips contextState to lost on first device loss, then back to ready on retry success — covers FR1 success path, asserts the scene owns a different device after loss.
• drops to renderer-unavailable + failed banner when the retry device also dies — covers D3-a.
• exposes __ROWER3D_FORCE_DEVICE_LOST only under __PLAYWRIGHT_TESTING (FR2 D4-a) — covers the test seam gate.
• writes cpuMs on every frame as the gpuMs fallback (FR3 D6-c) — covers the always-on timing.

Test plan

• npx vitest run src/lib/renderer/engine/__tests__/ — 119 pass (was 116 + 1 todo)
• npx vitest run — 3652 pass + 8 skipped
• npx tsc --noEmit -p tsconfig.app.json
• npx tsc -p tsconfig.playwright.json --noEmit
• npm run lint
• Pre-commit + pre-push hooks green
• CI on next/svelte-webgpu (preflight / scaffold smoke / WebGPU engine smoke / e2e-shards 1/2 + 2/2 / GPU runner skipped)
• RC verification on wood-laptop-g5 after merge

---

Generated by Claude Code