- test_fmha_v3_scalar: direct acc_scale for C6 O-rescale (no vector)
- test_fmha_v3_vec_c9: TMEM vector for C9 row_sum transfer
- test_fmha_v3_noop_c9: hardcoded inv_row_sum=1.0 (no normalization)
- test_fmha_v3_debug: row_sum-based C9 normalization
- test_fmha_v3_proper: 11-warp correction warp group (in progress)
Key findings:
- QK and PV C-fragments map threads to same logical rows
- pv_row_sum (PV-based P read) gives cosine 0.993 for n=128
- row_sum (QK-accumulated) gives cosine 0.514 for n=128
- Noop (inv_row_sum=1.0) gives cosine 0.866 for n=128
- pv_row_sum is NOT 1.0 - it corrects PV MMA accumulator errors
- The C9 normalization is essential even for single-tile case
