fix(gpu): drain resolved GPU timestamps at stop/teardown (#11)

[mledour]

Finding #11 (code review): GPU tail samples systematically blank

The per-frame DrainGpu() runs at different points on the two sides — pre drains after qpc_exit (outside its bracket), post drains inside its bracket — so when monitoring stops (ApplyToggle(false)) or the layer is torn down (~OpenXrLayer), each side can still be holding GPU timestamps the timer has already resolved but the writer never picked up. The sinks were closed without a final drain, so those samples were dropped and the last few frames' target_gpu_us went blank every session.

Fix

A one-shot, non-blocking DrainResolvedGpuOnce() called right before g_gpuCsv.Stop() on both stop paths. It recovers the resolved tail; frames the GPU hasn't finished yet are still not recovered (that would need a blocking fence wait → a stop hitch, deliberately out of scope). +26 lines, layer.cpp only.

Scope — the other GPU edge-cases are intentionally left as-is

After investigating #9/#10/#12, only #11 is a clean correctness win; the rest are non-bugs or accepted tradeoffs:

• feat(perf): wide post bracket + SPSC lock-free ring for ~12x less tar… #10 (D3D11 per-window vs D3D12 fixed timestamp frequency) — correct per-backend behavior, already documented in the README. The merge's freq-match blanking is right for D3D11 (disjoint can shift on power-state changes) and correctly never fires for D3D12 (fixed-rate counter).
• fix: lazy file-open + zero-frame merge guard close the parasitic-togg… #9 (bounded fence wait in Reset()/destructor) — a deliberate safety drain that returns as soon as the GPU catches up (~1 frame in practice); the destructor wait is required to free resources safely.
• feat(gpu): D3D12 backend behind IGpuTimer #12 (wedged writer after a failed thread join) — pathological, degrades gracefully, and is already logged; recovery would race a possibly-still-alive detached thread.

Verification

Like the rest of the stop/teardown glue (and MergeIntoOutput), this needs a live layer + GPU timer, so it's covered by the build + review rather than a unit test. No lines >120 cols.

🤖 Generated with Claude Code