Qualify native HRX Qwen prefill graph scheduling

Row: BACKEND-ROCM

The #3080 native HRX evaluation reaches Qwen3-0.6B generation after the separately tracked Loom repair (#3081), but longer prefill fails in the Qwen attention postprocess matcher.

Inputs: AMD-Ecosystem/llama.cpp 6319038132ed12f968ea68f37753f705da830ea8, HRX 6bcd5a4ff111fa5bf160ab9f4592ca8e7cc810b1 plus explicitly retained #3081 compiler patch, RX7900XTX gfx1100. Converted Qwen3-0.6B BF16 GGUF SHA256 2c7612c9a1d4f6d1f67fbee75ec88c089d1425696ac8361e9f73ca71023ea223. All 311 source tensors were checked against HF c1899de289a04d12100db370d81485cdf75e47ca.

Sixteen short prompts complete and match conventional HIP for all 256 output tokens. The retained 256-token prompt with context4096, batch512, microbatch512, F16 KV and flash attention enabled fails before generation: cannot cover value reshape node9, output f32[128,8,256,1]. The rejection originates in ggml/src/ggml-hrx/dispatch_registration/qwen/dispatch-qwen-attention-postprocess.cpp:693 through append_postprocess_node.

An explicit microbatch16 compatibility probe fails at the same value reshape, now f32[128,8,16,1]. Existing GGML_HRX_DISABLE_QWEN_DISPATCH=1 avoids that matcher but fails on unsupported per-head RMS_NORM f32[128,16,256,1]. These probes do not establish a hardware limitation or justify rejecting HRX. Standard HIP and its repaired HRX HIP-layer arm complete the same 256/2048-token requests and match all 512 generated tokens across the four qualification requests.

BACKEND-ROCM and the #3080 operator own diagnosis, a committed scoped spec before code, an upstream-native regression and minimal repair, fresh review and GPU qualification. Preserve the failed default configuration; do not silently replace it with disabled fusion or reduced context. The evaluation spec rocm-hrx-evaluation.md owns model/coverage qualification and will retain this issue under Owed in its final evidence update. Native broad prefill and model performance acceptance remain pending.

Start with ggml/src/ggml-hrx/dispatch_registration/qwen/dispatch-qwen-attention-postprocess.cpp at append_postprocess_node around line 693, using the retained context4096, batch512, microbatch512 configuration and the reported value reshape failure. Compare the default path with the microbatch16 and disabled-dispatch probes, then add an upstream-native regression and minimal repair. Done means the default native HRX prefill qualifies without disabling fusion or reducing context.