perf(gfx1201): split packed-Q8 verifier KV scans — +22.6% DFlash E2E at 21.5K context

[HUSRCF]

Summary

Stacked follow-up to #758. The base is intentionally fa2-prefill-gfx1201: #758 supplies Kaden's packed-Q8 GQA/LDS front end, while this PR only adds the split-KV verifier back end and product admission.

For small DFlash verifier batches at long context, partition the live causal KV prefix across 2/4/8 CTAs. Each split emits normalized Oacc plus LSE, then a small merge kernel combines the records with a stable max-LSE reduction. This restores parallelism without discarding #758's shared packed-KV loads inside each tile.

The route is opt-in through HIPFIRE_GFX12_FA2_SPLIT_VERIFY=1 (HIPFIRE_GFX12_FA2_SPLIT_COUNT={2,4,8}, default 8). It remains fail-closed to the existing multi-row admission plus exact gfx1201, H24/KV4/D256, Q8 KV, eager sequential verification. Tree, independent-batch, graph-capture, and retained/PM4 recording routes remain on the established path. Plain AR (n=1) cannot enter this route.

Performance

Radeon AI PRO R9700, gfx1201, Q8 VMM KV, Qwen3.8-27B MQ4XT, B16, context 20,676:

attention path	time	relative to split S8
incumbent WMMA	6.8854 ms	16.79x slower
#758 direct packed FA2	2.5556 ms	6.23x slower
packed FA2 + split-KV S8	0.4100 ms	1.00x

Fresh-process product DFlash A/B at 21,550 actual prompt tokens, 200 output tokens, HIPFIRE_VERIFY_GRAPH=0:

route	samples (tok/s)	median
existing R4/R8 verifier	39.3, 39.3, 39.3	39.3
split-KV S8	48.3, 48.2, 48.2	48.2 (+22.6%)

All six samples retained tau 1.80 and 71 cycles. A plain-AR smoke was unchanged within noise (32.6 tok/s with the env arm off, 32.5 with it on), consistent with the n=1 exclusion.

Fixture identities:

• prompt MD5: b4d0b63cddcac872648ddf3cdd92cac2
• target SHA-256: 9f91556f7e0431a077d03756a7102d0154108757289e6e5fe9a2d204c0c9eeb7
• DFlash draft SHA-256: d0a74a232a0e2166d889f823e91e0fbf778d21dd9668d7de055cdecb065401bc
• measured daemon SHA-256: 36576b9fc40345bbce6fa96c90906f56db81f2955c042c254432bc64d01437c1

Correctness and long-run E2E

The CPU-f64 oracle covered S8 tail/ragged/sentinel shapes. Worst relative L2 was 2.486e-4; minimum cosine similarity was 0.999999821. The partial kernel compiles at 243 VGPR / 27 SGPR and the merge at 18 VGPR / 10 SGPR, with zero spill and zero private scratch.

LongBench hard30, thinking enabled (temperature=1, top_p=.95, xhigh), Q8 VMM KV, max_seq=131072, output allowance 65,536:

• 30/30 completed, 0 errors, all natural stop
• prompt range 20,462–30,258 tokens
• 98,989 generated tokens total; median 1,568, maximum 21,876
• 20/30 scored correct (66.7%; the previous same-fixture checkpoint recorded 17/30)
• median decode 72.9 tok/s
• no empty output, open-think failure, replacement character, NUL, or repeated-line attractor across the 30 transcripts
• dataset SHA-256: 839a19be0b3b1c801a0ca58d388996faff34704e34dd196d54346adf17a1dce9

Five no-think long-decode fixtures (KV Cache article, SGLang article, story, Snake, Tetris):

• 5/5 completed, 0 errors, 15,954 generated tokens total
• three natural stop; the two 1,536-token technical articles reached their fixture-specific length cap
• longest output: Tetris, 8,178 tokens / 24,036 characters, natural stop
• generated Snake (290 lines) and Tetris (656 lines) Python code both pass AST parsing
• no repeated prose attractor or malformed output observed

The E2E binary was the equivalent current-beta port used to develop this stacked delta. The kernel source in this PR and that measured binary is byte-identical (SHA-256 7a006296173579377e69ec2a8a3947d3af40017922fb35d0d69eee91add2e394).

Memory

At S8/B16/H24/D256 the split records require about 3 MiB, but they reuse the existing flash_partials allocation. There is no new persistent VRAM allocation.

Static validation on this exact stacked commit

• changed-file rustfmt: pass
• cargo check -p rdna-compute -p hipfire-arch-qwen35: pass
• cargo test -p rdna-compute --lib: 246/246 pass
• cargo test -p hipfire-arch-qwen35 --lib: 196/196 pass, 17 hardware-fixture tests ignored by design

This improves the verifier cost only. It does not claim to repair low draft/target acceptance on prompts where tau itself collapses.