PR #4 turbo3 regression: PPL 6.18 → 181.6 after kernel split

[TheTom]

Bug

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

State	turbo3 PPL (c=512, 8ch, MoE 35B)
Pre-merge (7d1bd95)	6.1756 ✅
Post-merge (065ef53)	181.5955 ❌
Post-revert (a52586e)	6.1756 ✅

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

[TheTom]

The regression was reverted in a52586e the same day it was found. The kernel split approach was never re-introduced — turbo3 and turbo4 continue to share the SET_ROWS template. PR #3 (CUDA port) merged cleanly with no turbo3 PPL regression (verified: 6.1756, matching pre-merge exactly). No further action needed.