nx_nofloat_prefill_gate.nx

nx_nofloat_prefill_gate.nx -- R0r (2026-09-17): the BATCHED PREFILL is bit-identical to the sequential per-token prefill it replaces, proven on a real model through the serve core's own generate path, and its speed is printed as values. Two runs of nsv_generate on one prompt long enough for two full NF_PREFILL_M blocks plus a partial one: A = g_nsv_prefill_batched 0 (the sequential control) B = 1 (the served default) Teeth: the fixture reached the condition (blocks and a partial block), every KV cache row of every layer for every prompt position identical, a planted cell read as a difference (the comparator is live), the generated bytes and the token count identical, the served default is batched. Prefill ms for both paths and the speedup are values. The model (2026-09-18): argv[1] > sp_engine_model -- the engine model THIS HOST has staged (NX_ENGINE_MODEL env > knowledge/stage.conf engine_model=), a valid subject because the identity holds for ANY model the engine runs -- > the stage's nx_coder_model.gguf > SKIP when none opens. A roster forks this gate BARE, so the host-declared key is how the heavy roster reaches a real model without staging another model under the coder model's name (a label that lies). The rung that answered prints on the MODEL line (model_source=argv|engine_model|stage|none) and as model_source_code. expect_exit: 0 license_tier: ORIGINAL