nx_speculative_verify.nx

nx_speculative_verify.nx -- H3 speculative decoding bits-up. Per NISHI_ELDER_AI_OFF_DOCKER_2026_05_20.md §2.1 H3: small fast draft model proposes K tokens; large target model verifies them in one parallel forward pass; substrate keeps the longest accept-prefix + 1 bonus token (the target's own next-token prediction at the rejection boundary is free). Leviathan 2023 + Chen 2023 "Accelerating Large Language Model Decoding with Speculative Sampling." Practical wins: 2-3x throughput at equivalent quality when draft + target agree on common tokens. V1 mechanics: the substrate primitive does NOT run a model. It accepts the draft tokens (what the small model proposed) and a parallel accept_flag array (1 == target verified, 0 == target rejected) computed upstream by the verification kernel, PLUS the target's own next-token at the rejection boundary (the "free" bonus token). It returns the longest accept-prefix length and the count of tokens actually committed to the caller's output buffer. Composition path: - nx_speculative_session_new(K, request_id) on admit - For each verification round: nx_speculative_verify(session, draft_tokens, accept_flags, K_actual, target_next_token, out_committed) returns count_committed - Per-session stats (drafted / accepted / rounds) accumulate so the consumer can monitor the acceptance rate and back off speculative depth when the draft model is mispredicting - Acceptance rate exposed as Q16 fixed-point (rate * 65536) to keep the substrate primitive integer-only Pure substrate logic. No Linux features. Composes with shipped nx_batch_scheduler (each batch slot can carry a NxSpeculativeSession pointer for its decoding state). V1 honest scope: - Fixed max-K (NX_SPEC_MAX_K = 16) per round - Caller provides verification flags (substrate doesn't run the target forward pass; that's a tensor kernel one layer up) - Greedy verification only (no sampled-rejection-resampling;