HRX: intermittent MUL_MAT f32 m=1 n=1 k=2048 error on common.mul_mat.f32_f32_decode (about 1 in 40 runs at fork tip e44c9d0) #370

Seen 2026-10-09 on strixhalo while re-verifying the fork's quant PRs on the unmodified fork tip e44c9d01a4d5 (the engine's current llama.cpp pin; hrx-system 98d05d94; engine HRX configuration; performance mode). The fork has issues disabled, so it is filed here.

• test-backend-ops -b HRX0 -o MUL_MAT, 40 runs on the base: one failure (run 34), MUL_MAT(type_a=f32,type_b=f32,m=1,n=1,k=2048), ERR 0.0037, kernel common.mul_mat.f32_f32_decode.
• Full suite at the rebased quant stack tip (fork Registry: reviewed architecture gaps, and a guard against faking support with an alias #98-ONNX backend: 1bit serve --device onnx runs ONNX Runtime GenAI models (ryzenai-server) #100): 27 of 28 runs 1141/1141; the one failure is the same case (ERR 0.0102); 0 failures in 40 further MUL_MAT runs and 10/10 running the case alone.

Reproduces without the PRs, so it is pre-existing on the shared motifs/dequant.loom path. The magnitude and the 1-in-40 rate look like a race or an uninitialised read rather than rounding. Logs: strixhalo ~/wt/loom-rebase-scripts/out-*/.

🤖 Generated with Claude Code