Bug

PR #4 (merged in 065ef53, reverted in a52586e) introduced a turbo3 quality regression on Metal.

q8_0 baseline is unaffected at all context lengths — this is turbo3-specific.

Root cause

The kernel split from the shared kernel_set_rows_turbo template into separate kernel_set_rows_turbo3 / kernel_set_rows_turbo4 functions changed something in the turbo3 quantization path. The shared template worked correctly; the dedicated turbo3 function does not produce equivalent output.

Likely candidates:

• Template parameter difference (block type, QK constant)
• Quantization loop bounds
• Norm correction logic

Impact

turbo3 is completely broken on Metal when PR #4 is applied. Reverted.

References

• PR fix: turbo4 SET_ROWS, tail-block truncation, constant coupling, stack overflow (Issue #29) #4: fix: turbo4 SET_ROWS, tail-block truncation, constant coupling, stack overflow (Issue #29) #4
• Revert commit: a52586e