DeepSeek-V4.1-Flash on Wormhole

A correctness-first bring-up of DeepSeek-V4.1-Flash (552B plus a 196B Engram) on four n300 cards (8 Wormhole chips). Every model FLOP runs on device, and the host serves as a 449 GB expert store of pre-packed `.tensorbin` tiles behind an on-device LRU pool. Every stage is gated against DeepSeek's official `model.py` run on the CPU, reaching 0.985 decisive top-1 over a 2048-token prefill, with traced decode, 128K context, persistent prefix state and vision input. Speeds are stated honestly: decode is 1.68 tok/s (1.42 served), prefill is 95 tok/s, and DSpark speculation measured 0.48x so it is off. `CEILING.md` attributes the gap to about 7,200 ops per token of device-side dispatch, and `FINDINGS.md` lists 20 traps.