nx_triangulate_test.nx

nx_triangulate_test.nx -- VALIDATES the reusable triangulation harness (nx_triangulate.nx) itself: proves it PASSES three agreeing legs and CATCHES a deliberately-wrong leg. A harness that only ever says "pass" is worthless; this test proves the catch path fires, so the harness can be trusted to gate real capabilities (Cardinal 7 test before ship; Cardinal 25 build intelligence). We exercise the harness on a REAL capability (integer division) using three genuinely independent legs over a battery, exactly as the divider proof does: LEG A -- Newton-Raphson reciprocal divide (the multiplicative frontier) LEG B -- restoring shift-subtract divide (independent algorithm) LEG C -- behavioral `/` (oracle is C; A,B are the legs judged against it) Then we run two adversarial scenarios THROUGH the harness: (1) all three legs correct -> harness must report PASS, all-agree. (2) one leg deliberately corrupted (off-by-one) -> harness must report a CAUGHT failure, naming the bad leg index + its wrong value. KNOWN ANSWER (FAIL LOUD), one line: "<good_passed> <good_total> <caught> <bad_leg> <bad_got> <bad_exp> <argguard>" good_passed == good_total (clean battery: every vector triangulates) caught == 1 (the corrupted leg WAS detected, not swallowed) bad_leg, bad_got, bad_exp (the 5W1H diagnostic the harness surfaced) argguard == 1 (boundary guard rejects a malformed request) Any deviation -> nonzero exit. Expected: "65536 65536 1 1 36 35 1 " run=0