[sady4850]

attention-sweep builds its graph with PagedAttention::prefillPlan / verifyPlan called without the tile argument, so it always takes the default AttentionTile::Mpp. On Apple7/8 the engine runs the register tile (ExecutionPlans picks AttentionTile::Register, the *_sgf pipelines), so on M1/M2 the sweep was timing the stock MPP kernels, not the ones that serve. Found it while looking at long-context verify on M1 Pro: decode-profile shows verify_attention_q8_split_sgf, the sweep printed verify_attention_q8_split.

The fix builds the case graph from ops::ExecutionPlans(backend.capabilities()), same as the runtime does, one per backend (so --compare-metallib also uses it). Apple9+ unchanged, its plans give Mpp as before. apple7-m1-1.3.2 has the same bug.

M1 Pro (16-core GPU), qwen3.8-27b shape, Page32 int8 KV, verify, 1 lane, median of 15 fused graphs:

history	before (times MPP verify_attention_q8_split)	after (times verify_attention_q8_split_sgf)
2048	0.449 ms	0.280 ms
8192	1.419 ms	0.803 ms
24576	3.843 ms	2.145 ms
49152	7.667 ms	4.280 ms

Prefill phase picks prefill_attention_q8_split_sgf / _kv2_g8 now too. dev/tests/attention_sweep_cli.sh passes, build clean with -Werror. Output digests change on Apple7/8, as expected (other kernel, other summation order).

Why I care: at agent contexts verify attention becomes a real part of the M1 decode cycle (decode-profile, 27B, B1: 3% of the cycle at 2K, 22% at 24K, ~50 ms of 226 ms), so I want to try some things in the register tile and need the sweep to A/B it.

[paperniuk]

Thanks, good catch! Confirmed on M1 Max: the sweep now times the _sgf kernels in both phases. 27B verify at 1 lane goes from 0.239 / 0.742 / 1.919 / 3.758 ms (MPP) to 0.162 / 0.451 / 1.100 / 2.147 ms at 2K / 8K / 24K / 48K history. CLI test passes, build is clean. Merging.