Kimi-Linear on Wormhole

A correctness-first bring-up of Kimi-Linear-48B-A3B on four n300 cards (8 Wormhole chips). Each hot layer type (KDA, MoE, MLA) is one fused program built from Python through `ttnn.generic_op` and a `ProgramDescriptor`, with no tt-metal fork or C++ rebuild. Per-layer gates check each kernel against an fp32 torch oracle before a full load. Decode went from 1.44 tok/s eager to 13.5 tok/s at 16K context once it moved to paged flash-MLA on the compressed latent. The author notes that chunked prefill is 28x faster but wrong, so it is disabled. `FINDINGS.md` lists 15 faults, three of them silent, and how each was found.