Interleave GPU tile variants: grouped count runs fabricated a Q4_K rows-per-SIMD win #865

Finding

An Apple M2 Pro Metal Q4_K sweep compared the production 2-row SIMD-group kernel with exact 4-row and 8-row variants. Both candidates were bit-identical to control, including odd tails, and mutated zero input bytes.

A conventional grouped Go benchmark (-benchtime=300x -count=4, all control samples followed by all candidate samples) made rows=8 appear roughly 1.19x faster at K2048/N2048 and about 1.05x faster at K5632/N2048. That apparent gain did not survive an order-alternated fresh-process protocol using the same binary, 7 pairs per campaign, 3 campaigns, and the second sample from each count=2 invocation.

Rows=8 results:

• K2048/N2048: 1.015x, 1.015x, 0.976x
• K5632/N2048: 0.963x, 0.984x, 0.982x

The dominant K2048/N5632 shape was already neutral-to-negative in the grouped screen. No candidate shipped.

Generalizable implication

GPU tile and occupancy variants can move thermal state and clock residency enough that benchmark subtest grouping creates a false shape-specific win, even with hundreds of iterations and multiple count samples. A detector or benchmark reviewer should flag performance claims where variant samples are grouped by sub-benchmark order. Require arm alternation at the process or invocation boundary, discard each arm first sample, and report every campaign rather than pooling the samples. This extends FIRST-BENCHMARK-SAMPLE-IS-NOT-COMPARABLE-001 to cross-variant GPU ordering.