nx_specdec.nx

nx_specdec.nx -- sovereign SPECULATIVE DECODING (prompt-lookup drafted, greedy-verified; census OPPORTUNITY "Speculative decoding", momentum 3). Decode is MEMORY-bound (weights re-read per token). Speculation turns k sequential m=1 forwards into ONE m=k+1 forward: DRAFT k tokens cheaply, VERIFY them in a single multi-token pass (weight bytes read ONCE for the whole block -- the same amortization that makes prefill cheap). With GREEDY acceptance (argmax match) the output is BIT-IDENTICAL to plain greedy decode: pure measured speedup, zero quality change. Draft source here = PROMPT-LOOKUP (n-gram continuation from the model's own context; Saxena 2023 "prompt lookup decoding", the draft-model-free member of the speculative family: Leviathan 2023 / Chen 2023; EAGLE-class trained drafts can replace it later behind the same verify loop). Acceptance is WORKLOAD-DEPENDENT: echo/quote/list-heavy text accepts big chunks; free-form text accepts little (then cost ~= plain decode plus the occasional rejected block). Cache contract: the verify forward appends k+1 rows; on partial accept the cache is TRUNCATED back to the accepted prefix (seq_len rewind -- rows are plain arrays, later appends overwrite). Positions/causality inside the multi-token pass are the offset-causal q_pos = total_k - n_q + i already implemented by nx_f32_attn_multi (proven by nx_specdec_gate EQUIV rows). genealogy_id: leviathan_2023_speculative + saxena_2023_prompt_lookup lineage_id: substrate_specdec_v1