nx_mesh_batch.nx

nx_mesh_batch.nx -- WORKER MESH: DYNAMIC BATCHING scheduler + request QUEUE with admission control (the Triton dynamic-batcher analog). A GPU serves far more throughput if concurrent requests are COALESCED into one forward pass instead of run one-at-a-time. This organ is the SCHEDULER that sits in front of a worker: * QUEUE (bounded FIFO) + ADMISSION CONTROL: admit while qlen < cap, else REJECT (backpressure / 503) -- protects the GPU from unbounded pileup under load. * DYNAMIC BATCH COALESCER: form + dispatch a batch when EITHER the queue reaches the preferred batch size (qlen >= max_batch) OR the oldest queued request has waited >= max_delay_ms (bound tail latency) -- exactly Triton's {preferred_batch_size, max_queue_delay}. A dispatched batch of K maps to an n=K worker call. The scheduler is pure integer policy -> mechanically provable. A discrete-event SIM feeds an arrival stream and reports batches formed, avg batch size (the throughput win), rejects, and max queue depth. CLI: (no args) -> self-test GATE (policy table + sim + neg-controls) sim <burst> <max_batch> <max_delay> <cap> -> run the sim for a burst of N arrivals at t=0, print result Closes TWO mesh axes: dynamic-batching AND request-queue-admission. NO fake greens: the sim + neg-controls prove a batch NEVER exceeds max_batch, an empty batch is NEVER dispatched, and admitted NEVER exceeds cap. license_tier: ORIGINAL