nx_llm_sched.nx

nx_llm_sched.nx -- sovereign CONTINUOUS-BATCHING scheduler (Orca / vLLM serving core). Requests ADMIT at any time (chunked prefill onto their own paged sequence, forked-nothing: independent prompts) and JOIN the decode batch mid-flight; each ROUND advances every active request by one token via ONE batched forward (nx_f32_llama_v4b); finished requests leave the batch and free their pages. The socket seat wraps THIS; the correctness property gated by nx_llm_sched_gate is BATCH-INVARIANCE: a request's bytes are BIT-IDENTICAL to its solo run, regardless of co-tenants or admission timing. genealogy_id: yu_2022_orca_continuous_batching + kwon_2023_pagedattention lineage_id: substrate_llm_sched_v1