[merceod]

Stacked on #1090 (pr-24-t3-specdec). Picks the commits of #1091, #1096, #1099, #1101 first; they drop out of this diff as those merge.

Why

#1090's draft passes ran over the verify's (rows, k+1) tokens and threw k of them away after attention, and the step's bookkeeping was 88 small torch kernels plus 148 copies and fills. A 9B k=2 step at c=1 spent 886 us per drafted token on drafting (one GPU trace, shared node). After that, the 248k x 4096 vocabulary head is most of a draft pass.

What changed

Stacked on #1090, with #1091 (bringing #1085), #1096, #1099 (bringing #1092) and #1101 (bringing #1100) picked first: #1099 for 7915843, the fp8 op the draft head uses, #1096 for the compile_dynamic capture configs, #1101 because this PR's decode configs give a seeded start 22 compile-pool jobs instead of 1, which exposes the hang #1101 fixes.

• One query a row (Qwen3_5MTPHead.draft, SpecDecodeResource.attend_draft, attend_rows): pass 0 projects all k+1 tokens and attends one query a row, later passes run one token a row.
• mstar::qkv_norm_rope_cache: q/k norm, rope and the MTP layer's KV write in one kernel. mstar::spec_verify_inputs and mstar::spec_accept replace the bookkeeping with two Triton kernels.
• record_draft runs sample_unfiltered at temperature 0 in the graph, into the next step's token register. GDN verify rows run on FlyDSL's chain kernel. decode_rows=True compiles a verify bucket as decode by its rows, and the MI325X decode GEMV configs gain rows 3, 6, 12 and the fc.
• draft_head: fp8: an fp8 copy of the head with one scale for all of it, AITER wvSplitKQ at 1-2 rows, _scaled_mm from 3. 415 to 213 us at one row, +1 GB of VRAM on 9B. draft_head_variants.py measures acceptance per head variant.
• CUDA path unchanged: decode_rows defaults to False and only the ROCm inductor_overrides read it, the rest sits behind num_draft_tokens, which no CUDA deployment declares. rust/ unchanged. Docs updated.

Verification

• One GPU trace at 9B c=1: drafting costs 486 us a token at k=2 (886 before) and 428 at k=3 (789 before).
• Window I (job 551), 8x MI325X, exclusive, 9B text-chat-256, three timed runs, tok/s. plain and mtp2 ran f7122c3 (Speculative decoding as an engine resource, with Qwen3.5's MTP head as the draft #1090's path), the fused arms this branch's tree, all under MSTAR_PHASE_TIMING=20, so not comparable with window G.

c	plain	mtp2 (#1090)	mtp2 fused	mtp3 fused	fused k=2 / plain	fused k=2 / mtp2
1	224	331	408	440	1.82x	1.23x
4	766	1155	1331	1424	1.74x	1.15x
16	2586	3529	3542	3966	1.37x	1.00x
64	7139	7880	8560	7857	1.20x	1.09x

• k=3 fused is 1.96x plain at c=1. The c=16 fused k=2 cell is noisy: 4 of 6 runs stalled, range 2948-3933.
• Shared node, three launches per arm, clean ms/token (best repeat's median request), 9B chat: k=2 2.33 against plain's 4.36 = 1.87x (bf16 head 1.74x), k=3 2.17 = 2.01x (bf16 head 1.86x, so k=3 needs the fp8 head). long-1k k=2: 1.59x at c=1, 1.23x at c=4, wall 1.01x at c=64.
• Greedy correctness PASS with 0 non-tie divergences at k=2 and k=3 on 9B and 0.8B, fp8 and bf16 heads, and on every window I arm. Sampled k=2 against plain, 1600 samples over 8 prompts: p<0.01 in 0 of 32 tests on 9B and 0.8B. Head acceptance within 1 % of bf16 on chat, long and 12-language sets; tokens a step 2.44 / 2.85 (Speculative decoding as an engine resource, with Qwen3.5's MTP head as the draft #1090 2.47 / 2.85).
• No host sync: the whole step, every draft pass's attention included, runs under set_sync_debug_mode("error"), and the fp8 head at 1, 2 and 4 rows. Three defects found reading the diff (argmax buffers sized by max_requests, a crash on an ungated head, a context past the rope table) are fixed with tests in d1e608d.
• Tests, from the branch and integration trees, not this cut branch: ruff clean, GPU 1355 passed and 9 skipped (test/rocm plus test_kernel_ops_parity), CPU Core 10692 tests and 34 failing at 6ce8b1e, the base's set.

Not in this PR

• The speculative configs stay opt-in. The fp8 head is on in the four MTP configs; draft_head: bf16 turns it off.
• k=3 at c=64 only equals Speculative decoding as an engine resource, with Qwen3.5's MTP head as the draft #1090's k=2 (7857 vs 7880), so k=2 is the config where c=64 matters. 0.8B keeps speculation off. long-1k wasn't in window I; at long-1k c=64 every MTP arm's ms/token is 0.75x plain's, Speculative decoding as an engine resource, with Qwen3.5's MTP head as the draft #1090's mtp2 included.
• Slow repeats with 25-33 ms event_sync bursts show up on plain arms too. FlyDSL's standalone GDR test script segfaults in a bare process after an AiterVerify case, never in pytest or a server.
• TP and NVIDIA not run: the fp8 head refuses TP, the configs are one GPU, and speculation needs AITER's attention.