nx_converge_lib.nx

nx_converge_lib.nx -- THE CONVERGENCE CONTROLLER: hit a RETENTION TARGET instead of guessing a denoise. THE PROBLEM. Reference-conditioned generation has one continuous knob (denoise strength) and one thing you actually care about (how much of the reference survived). Today an operator guesses a denoise, waits 25-50s for the GPU, eyeballs the result, and guesses again. The guess is the whole workflow, and it is neither reproducible nor auditable. THE INSIGHT THAT MAKES THIS EXACT, NOT HEURISTIC. nx_refbench_gate's T4 proves the ruler ORDERS correctly, and the measured curve falls monotonically with denoise (859 -> 647 -> 359). A monotonic function on a bounded interval is exactly the case where BISECTION IS GUARANTEED TO CONVERGE, and to do so in ceil(log2(range/tolerance)) steps. So this is not a tuned heuristic with magic gains -- it is a bracketing search with a proof behind it, and its step count is known BEFORE the first render. Retention DECREASES as denoise rises, so the bracket updates are inverted relative to a naive search: measuring ABOVE target means we kept too much and must push denoise UP. WHY IT IS STATELESS. The render happens on the GPU host; only the DECISION belongs in the sovereign organ. So `next` takes the whole bracket as arguments and returns the next probe -- an agent, a workflow, or a human can drive it one render at a time over MCP with nothing persisted between calls. State that lives in a daemon cannot be replayed; state that travels in the arguments can. NO-FLOAT: every quantity is permil (0..1000) integer, per the sovereign arithmetic law. license_tier: ORIGINAL