nx_stubmain_gate.nx

nx_stubmain_gate.nx -- proves the stub-main ruler separates three states and does not invent a fourth. PLANTED FIXTURES ALONE WOULD NOT CALIBRATE THIS. The estate has already measured that a detector bite-proven on planted fixtures can carry a 941-permil false-positive rate on real data. So T5-T8 run the ruler against REAL SOURCES IN THIS TREE whose correct answer is independently known: nx_https_get.nx -> SM_STUB_MAIN (its own comment says `Compile-only smoke.`; measured 2026-09-03 to build 14,880 bytes against a 468,929-byte live binary) nx_https_fetch.nx -> SM_NO_MAIN (nx_catalog independently reports `KIND LIB -- no top-level main()`) nx_npy_gate.nx -> SM_REAL_MAIN (a program shipped and run GREEN on the NAS this session) nx_stubmain_lib.nx -> SM_NO_MAIN (this ruler's own subject is a library) A ruler that cannot classify the four sources whose answers are already known has no business classifying the other 12,063. THE FALSE-POSITIVE TOOTH IS T4, and it is the one that matters: a REAL main whose LAST statement is `return 0` and which contains a nested block must read REAL. A "first closing brace" body-finder would end the body early and report STUB -- turning a working program into a false ARMED row, and a false ARMED on this axis sends a seat to disarm something that is fine, which is how detectors get ignored. license_tier: ORIGINAL expect_exit: 0