HRX: wrong answers for every prompt longer than one ubatch since fork #82 (graph inputs written back to host) #319

Found 2026-10-05 while root-causing #313 (strixhalo, performance mode). On the llama.cpp fork tip (e57beb9 = the pin merged in #311), Qwen3-0.6B Q4_K_M on HRX0 answers a 4,798-token prompt by repeating the question — 28/28 runs, same hash (8d294327), with and without the private add-on. Short prompts (one ubatch, ≤512 tokens) are unaffected, which is why the e2e and "Paris." gates passed.

Cause. qwen_attention_metadata binds the positions, the K/V cache indices and the mask as read_write, so they are queued as deferred host writebacks. Since fork #82 (e0f637b) every upload first publishes pending writebacks — which overwrites the next ubatch's fresh inputs with the previous ubatch's values. Every multi-ubatch Qwen prompt is affected.

Fix. Fork commit "ggml-hrx: never write a device-rewritten graph input back to host memory": a graph_input flag on such bindings — uploaded, never downloaded. Same PR carries "wait for a non-replay command program that reads registered host memory" (graph replay off answered "123" instead of 7341: the non-replay path returned before its own device-side input copy ran, and ggml overwrote the host buffer).

The NaN in #313 itself is a separate, third bug (Loom compiler GFX11 wave64 SGPR lane-mask hazard; see #313).

After the fork PR merges the engine pin moves; until then --device hrx with prompts over 512 tokens is wrong on main.