AMD Windows fix: the Adrenalin OpenCL compiler rejects the GNU inline asm the kHeavyHash kernel used (CL_BUILD_PROGRAM_FAILURE on Radeon VII / gfx906 and other AMD cards). A -D NO_INLINE_ASM flag routes the JIT to a bit-identical scalar dot-product path on Windows (no-op on Linux/ROCm). Proven: preprocessor asm 1->0, and scalar == v_dot8_u32_u4 on real gfx906 across 100k random vectors (0 mismatches). AMD assets below.