## Summary
KDA (Kernel Design Agents) builds its kernel tasks from this dataset, with the rule that a kernel written against a definition must drop into a real inference engine. This PR curates the dataset to that rule. It also adds SGLang's own implementations as baseline solutions, and KDA's acceptance configs.
**Removed**
- `dense_gemm_bf16_nt_n64_k4096` and `dense_gemm_bf16_nt_n4096_k64`: the image output and input projections, memory-bound 8–45 µs GEMMs that run once per forward.
- `grouped_gemm_fp8_fp4_e384_n2304_k5120_g32`: cut from the columns of the concatenated gate/up GEMM. No engine runs a separate gate or up GEMM, and its baseline computes the full N=4608 GEMM.
**Replaced**
- `gated_delta_rule_{decode,prefill}_h48_dk128_dv128` gave Q/K pre-repeated to 48 heads and a dense `[B, H, K, V]` state, while engines keep a k-last state pool. A kernel written against them would not drop into an engine.
- They are replaced by the [MLSys 2026 FlashInfer contest](https://huggingface…
24 unchanged definitions and references, 1383 workloads with matching captures, 24 baseline solutions, and all raw safetensors grouped by definition. Include only capture, validation, model preparation and plotting tooling.
