perf(draft): pack dynamic-convolution projections across lanes

[Graffioh]

Depends on #656. Review the top commit, 00ba24f1; the diff collapses to this mechanism after #656 merges.

Summary

• pack normalized hidden columns across active draft lanes before the shared attention and MLP dynamic-convolution coefficient projections
• slice projected coefficients back into lane views before every convolution apply
• keep convolution history, temporal shifts, positions, masks, cache writes, RoPE, attention, and apply steps lane-local
• retain the singular projection path for C1, preserving the existing graph
• add a real-model multilane isolation check with distinct inputs and positions

The production change is 25 additions and 6 removals in one file. It retains no legacy multilane runtime path.

R9700 fixed-C A/B/A

Qwen3.8 27B IQ4_XS target, Q8 DFlash2 width 8, Q8 KV, HumanEval raw code prompts, ten fixed-width waves, 256 forced output tokens, temperature 0. Normal runs provide latency and goodput; matched ROCprof runs provide dispatch and device-time evidence.

C	Baseline A draft p50	Candidate	Baseline A2	Improvement vs weaker baseline	Dispatch delta / fixed round
1	5.3470 ms	5.3441 ms	5.3488 ms	0.06%	0
2	6.8028 ms	6.5815 ms	6.7946 ms	3.14%	-10
3	8.6155 ms	8.2460 ms	8.6070 ms	4.19%	-20
4	10.2278 ms	9.6294 ms	10.2264 ms	5.84%	-30

The mean minimum A/B/A draft-compute improvement at C2-C4 is 4.39%. Complete output digests match at C1-C4, all 300 requests return 256 tokens, and captures report zero dropped records.

Against the A/A2 bracket mean, end-to-end goodput improves by 0.79%, 0.23%, 0.43%, and 0.47% at C1-C4. Those are marginal system gains. The keep decision rests on the small production diff plus two separately measured effects: materially lower draft compute and ten fewer speculative dispatches per added lane.

The launch accounting is exact per added lane: ten Q8 coefficient-input quantize launches and ten coefficient MMQ launches disappear, while ten packing concat launches are added. Fixed-C elementwise device time rises by 0.88%, 1.36%, and 1.56% at C2-C4, but coefficient GEMM time falls enough for the draft phase to win. TTFT is mixed and is not a justification for this change.

Strix Halo

The same normal A/B/A protocol passes output parity, fixed-token accounting, capture integrity, and lower-C drain accounting. Its timing is performance-inconclusive: the unchanged C1 path moves by 8.65% against the bracket mean and A/A2 draft drift reaches 15.47%. It is retained as correctness and harness evidence, not as support for the performance decision.

Verification

• exact HIP build
• test_inference_profile
• test_observability
• test_qwen35_roctx
• test_seq_engine_contract (15 checks)
• real Q8 three-lane dynamic-convolution isolation check; all lane outputs have max absolute difference 0
• R9700 C1-C4 normal A/B/A and matched ROCprof comparison
• Strix Halo C1-C4 normal A/B/A integrity comparison
• git diff --check

The device-resident selector handoff and the R9700 target MMQ C2-to-C3 cliff remain separate follow-ups.

[Graffioh]

Folded into #659. The surviving head e9103d7 contains this PR's exact dynamic-convolution packing commit (00ba24f), preserves its causal evidence, and passes the combined stack checks. Closing this draft to keep one #651 concurrency follow-up.