nx_blockfloat_generate_gate.nx

nx_blockfloat_generate_gate.nx -- R2 of the model-assembly arc: AUTOREGRESSIVE GENERATION on block-float weights. Composes R1's full LM forward (embed -> block -> final norm -> LM head -> logits) in a feedback loop: argmax the last position (greedy / temperature-0 decoding), append it to the sequence, recompute. This is the loop that turns a forward into an LLM (prompt -> generated continuation). Greedy decoding is fully deterministic = the determinism moat at the SEQUENCE level. Proves block-float generation == full-precision generation (lossless), while per-tensor INT8 CORRUPTS what the model generates. 1 the loop GENERATES n_new valid tokens (sequence grows, all ids in [0,vocab)) 2 DETERMINISTIC: block-float generation twice == identical sequence 3 block-float generation == full-precision generation (lossless at the sequence level) 4 per-tensor INT8 CORRUPTS the generated sequence (out_pt != out_fp) -- block-float's win is over real corruption expect_exit: 0 license_tier: ORIGINAL