#1728 [Bug]: Windows gfx906 compatibility with native HIP 5.7

本文

### What happened

The opt-in `STRATA_HIP_GFX906` backend can be ported to a native Windows MI50 stack, but the Windows HIP 5.7 SDK needs compatibility changes beyond the current Linux recipe. This report concerns manual source builds, not setup detection or the ready-made Windows wave32 engine.

The remaining source blockers are the modern hipBLAS V2 datatype names versus the legacy 5.7 GemmEx API, two missing stream-priority aliases, Windows pinned-memory code reading a DXGI LUID absent from legacy HIP device properties, and HIP clang 17 parsing recent MSVC STL C++20 code in `remote_expert_opt.cu`. An anonymous-namespace prompt activation device symbol also needed external linkage for host `hipMemcpyToSymbol` on the Windows/PAL stack. The engine should identify gfx906 wave64 correctly and reject a wave32 card before launching kernels.

I searched the open issues and PRs and found no current Windows HIP 5.7 compatibility report covering these together. The fixes already merged for `cudaFuncSetAttribute` (#808) and Q6_K / event declarations (#1396) are excluded from the proposed patch.

### Strata version, GPU, OS

Upstream main `fb58e0dbc8399662c0e47c76578c6e878b14f6cf`, ggml pinned to `3cf03257f219afbe7334045ff7c6a06ac68c627d`; Windows 11; Instinct MI50 16 GiB, reported by the driver as AMD Radeon Pro VII, gfx906 wave64. Native HIP 5.7 / hipBLAS / gfx906 rocBLAS from AMD's 23.Q4 Windows HIP SDK, HIP clang 17, host clang 23 with MSVC 14.44, PRO 26.Q1 driver, code object V5. No ZLUDA.

The SDK was copied into a private prefix with the hip-lang import prefix, device-bitcode directory, and six non-inline BF16 free functions repaired. This is not a claim that the untouched SDK works. The proposed documentation records these prerequisites.

### Scope

Experimental opt-in compatibility only. No new optimized kernels, tuning tables, setup-card detection, vision, multi-GPU, or performance claim. I will link an accompanying draft PR with the build recipe and actual validation results. CUDA, wave32 HIP, Linux gfx906 and SYCL builds are not available on this host, so the draft will not request review yet.