research(KERNEL-CUDA-DECODE-MEGAKERNEL): evaluate tile-scheduled persistent decode

Row: KERNEL-CUDA-DECODE-MEGAKERNEL

Evaluate a CUDA decode megakernel that schedules GEMM tiles, attention splits,
and routed-expert work across operation boundaries.

Cohere reports that its North Mini Code engine reaches 292 tok/s at batch size
1 on one H100. The engine reports 1.58x the decode throughput of vLLM v0.24.
The release uses one persistent block per streaming multiprocessor, fixed task
descriptors, counter barriers, and dynamic work queues.

Reference: cohere-ai/cohere-megakernel commit
67d0b9ca22ea3652796b715d1d1863459e0e2c3c.

Scope:

• Pin and inspect the complete executing path in the Cohere release.
• Trace an identical low-batch decode workload in vLLM and vllm.cpp.
• Measure launch gaps, wave tails, device bandwidth, and idle capacity.
• Specify a CUDA task ABI with descriptors, operation codes, barriers, and
  schedule variants.
• Build one synthetic persistent-kernel proof if the trace shows recoverable
  idle capacity.
• Preserve eager and CUDA Graph fallbacks for every unsupported shape.

The pinned vLLM oracle remains the correctness source. The Cohere release is a
performance reference only. Any implementation must route through
ModelRegistry::Forward, dense_attn::AttnBlock, vt::FusedChain, and
vt::MergedGemmGroup where those shared seams apply.

Out of scope:

• A performance claim before an identical-workload, same-device A/B test.
• A universal execution model for CPU, ROCm, Metal, Vulkan, or Tenstorrent.
• Vendoring ThunderKittens before the dependency decision has measured evidence.
• Treating the BF16 H100 release as completion of required quantized arms.
• Model integration in the initial spike.

Stop conditions:

• Stop implementation if the trace does not show material recoverable idle
  capacity after CUDA Graph replay.
• Leave H100-specific execution PENDING until a leased or otherwise authorized
  H100 can run the gate.
• Refuse a result if token output differs from the pinned vLLM oracle.

FOLLOWING_AGENTS_PROTOCOL

Start by pinning and inspecting Cohere commit 67d0b9ca22ea3652796b715d1d1863459e0e2c3c, then trace the identical low-batch workload in vLLM and vllm.cpp through ModelRegistry::Forward, dense_attn::AttnBlock, vt::FusedChain, and vt::MergedGemmGroup. Done requires measured launch, tail, bandwidth, and idle-capacity evidence against the pinned vLLM oracle, with implementation attempted only if recoverable idle capacity is shown.