[DennisSmolek]

Fixes #82. On Windows, synchronous compute-pipeline creation blocks the GPU main thread. Initialization now uses async compilation through a four-slot queue shared by GPUDevice, caches identical modules/pipelines, and prepares only the selected path. Debug and disabled shading-change passes remain lazy. TSL/presenter integrations render the input while preparation is pending; optional activation preserves the current input jitter and split-frame settings. API migration: raw callers must configure and await init() before dispatch, including after switching to an unprepared path. Adds prepare(), isReady, and a compile-free isSupported(device) check. MomentsPass and UpscalePass expose async initialization. Exposure behavior and shader math are preserved. Windows measurements (Chrome 154, RTX 3060 Laptop, three fresh profiles plus matching reloads): - Median GPU-main compute-creation time: cold 1647.425 → 2.077 ms; warm 152.924 → 0.958 ms. - Default temporal startup: 10 sync → 8 async pipelines; spatial: 10 sync → 3 async. - Two-instance split demo: 20 sync → 8 shared async pipelines. - Dawn traces place upscaler DXC compilation on ThreadPoolForegroundWorker. Residual render-shader stalls remain; this is not a total-startup speedup claim. Validation: - Node 22.23.3/npm 11.5.1: lint, typecheck, 472 portable tests, library/gallery builds, packed-consumer build. - Packed Chrome GPU check: 16 ordered split frames, two backing textures, zero monolithic fallbacks. - All 20 Chrome demos; six representative Edge demos; resize/DPR and 32 mode/debug selections; real MomentsPass and device-loss checks. - All three benchmark source bundles completed Q1/ratio-2 ABBA smoke with no browser validation errors. - Full Ubuntu verification passed on the initial implementation; final-revision CI is tracked on this PR. See Windows audit and roadmap and docs/windows-evidence/ for committed samples and coverage. Remaining findings: the existing POSIX release-workflow harness fails on native Windows (excluded from the portable result); demo 14 retains synchronous calls inside three-gpu-pathtracer; browser discovery/Windows CI need follow-up; S2 reads timings without opting in. Forced FXC and other adapter families were not tested. No package release or merge is performed.

Breaking initialization contract

Upscaler.init() now returns Promise<void>. Existing raw callers that ignore it can dispatch before preparation completes and throw; shader compilation rejection is asynchronous. MomentsPass requires a new explicit awaited init. UpscalePass and TSL nodes provide preparation fallback.

Configure first, await init/prepare, then start raw dispatch. This must be reviewed and released as a breaking API change with migration guidance; it is not compatible with the old immediate raw-dispatch pattern. No merge or release has been performed.