test(runtime): exercise in-flight accessibility deferral

[phall1]

Fix

Make the post-present accessibility batching regression exercise its stated premise. The fixture now provides a no-op request_gpu_surface_frame_fn, allowing requestCanvasFrameForView to mark the frame in flight. Batch end therefore keeps accessibility deferred until the explicit GPU frame completion.

Without the callback, the request returned UnsupportedService, no frame was pending, and batch end correctly settled accessibility immediately; the previous update_count == 0 assertion contradicted the fixture.

Evidence

• focused test now drives the real request/flag/defer/completion sequence
• consumer pin will be validated through the full Cockpit gate and live M4 first-frame measurement before merge