nx_domain_admit_lib.nx

nx_domain_admit_lib.nx -- THE PURE DECISION CORE OF nx_domain_admit, THE /compare ADMISSION GATE. WHY THIS FILE EXISTS AT ALL. nx_domain_admit decides whether a /compare domain may publish, and on 2026-08-25 the ship loop reported for it: "PROVE NO-GATE-FOUND -- a NAMED absence, NOT a pass: NOTHING about this target has been proven." The organ that refuses on everyone else's behalf had nothing proven about itself, on a day it shipped TWO changes (C3 both-tree gates resolution, C4 fresh-kill from the bite records). A GATE COULD NOT HAVE BEEN WRITTEN AGAINST THE ORGAN AS IT STOOD. nx_domain_admit is a top-level programme, so nothing can import it; and a gate that fork/execs the deployed elf is NOT mutation-provable -- the harness rebuilds the TEST and not the SUBJECT, so every mutant returns NOT-REACHED and the resulting GREEN reads exactly like a proof while proving nothing. Extracting the decision core into a library that sits INSIDE the gate's build closure is what makes the proof possible at all. THIS IS A MOVE, NOT A COPY. Every function below was CUT from nx_domain_admit.nx, which now imports this file. There is still exactly ONE word-boundary ruler, ONE row classifier, ONE bite-record reader and ONE gates-map resolver in the estate. A second copy would be the duplicate-ruler defect wearing a test's clothes, and the two would drift the moment either was fixed. WHY THIS FILE LIVES IN _hdl_build/ AND NOT IN runtime/ BESIDE THE OTHER LIBS -- MEASURED, NOT CHOSEN. The first draft was written to buildroot/runtime/ and the build FAILED with twelve "there are two definitions of dp_read/dp_len/dep_run/..." errors. nx_deploy_lib.nx exists TWICE in this estate -- buildroot/runtime/nx_deploy_lib.nx and buildroot/runtime/_hdl_build/nx_deploy_lib.nx (5,655 B) -- and imports resolve RELATIVE TO THE IMPORTING FILE, so the organ in _hdl_build/ and a library in runtime/ each pulled in a DIFFERENT nx_deploy_lib and the compiler saw every dp_* symbol twice. A SHARED DEPENDENCY THAT EXISTS AS TWO TWINS IS DEDUPED BY PATH, NOT BY NAME: a library and its consumer must sit in the same tree or their common import is textually included twice. The compiler refused loudly, which is the good failure -- had it silently picked one, the organ would have changed which dp_read it runs without a single line of its own code moving. TWO THINGS WERE EXTRACTED RATHER THAN MOVED, and each removes a real duplicate: da_read_2dir / da_gates_read -- C3 (in main) and C4 (in da_c4) EACH built the <domain>.gates path and EACH tried the two trees. That second copy is precisely how the 2026-08-25 defect survived: the 2026-08-20 repoint moved the MATRIX read to the publishing tree and did not move its sibling, so has_gates read 0 for every domain in the estate. ONE resolver, called twice, cannot half-move. da_c4_partition -- the bucket arithmetic, separated from path resolution so it can be measured against a fixture instead of against the live gate fleet. A gate that shares a fixture with a production beat measures the beat and not the code. 100 percent sovereign. No hardware writes (Rule 26). license_tier: ORIGINAL