nx_board_contract_gate.nx

nx_board_contract_gate.nx -- THE REFEREE FOR BUILD CONTRACTS ON THE COMPARE BOARDS (2026-09-03). THE DEFECT THIS GUARDS, measured on /compare/lang the day this shipped. A matrix row declares a build contract in its symbol field as _ABSENT_:<symbol>, and the emitter flips that cell the moment the symbol exists in the organ named by field 2. That is the hive's completion signal. But NOTHING checks that the named organ is even CAPABLE of carrying the symbol -- so a row whose field 2 points at the wrong file is a contract that can NEVER flip, and it reads exactly like honest work-in-progress forever. LR7 on the lang board sat in that state and cost a full re-derivation: its field 2 named runtime/nx_boundscheck.nx while bck_elide_dominated is declared in runtime/nx_bck_elide.nx. The capability had shipped; the cell still said WATCHING; the note still said the compiler lacked it. WHY A SUBSTRING TEST WOULD BE WORSE THAN NOTHING HERE. nx_boundscheck.nx DOES contain the characters bck_elide_dominated -- inside the DISTINCT symbol bck_elide_dominated_witness, and in a comment. A naive substring probe therefore reports that row as perfectly resolvable and the whole class stays invisible. This gate looks for a DECLARATION (func <symbol> followed by an open paren) and reports a bare textual hit as its own class, SUBSTRING-ONLY, because that is the shape the trap actually takes. CLASSES, and every one of them is NAMED, never merely counted: DECLARED the named organ declares the symbol -- the contract CAN flip. Healthy. MISPOINTED-CANDIDATE the symbol TEXT occurs in the named organ but no DECLARATION does. This is the LR7 shape and the ONLY class this gate treats as a defect. PENDING-IN-FILE the organ exists and does not carry the symbol yet -- a normal OPEN contract. PENDING-NO-ORGAN the named organ is not written yet -- also a normal OPEN contract. THE RATCHET COUNTS MISPOINTED-CANDIDATE ALONE, AND THAT WAS A CORRECTION MADE ON THE FIRST LIVE RUN. v1 called all three non-DECLARED classes "unflippable" and reported 1052 of 1169 contracts as defects. That is a false-positive machine: an _ABSENT_ contract naming an organ nobody has written yet is EXACTLY what an open build contract IS. Measured live: 468 PENDING-NO-ORGAN + 573 PENDING-IN-FILE are ordinary work in flight, and only 11 rows carry the shape that can never flip. ⇒ A DETECTOR WITH FALSE POSITIVES IS WORSE THAN NONE, and its own first census is where you find them. KNOWN LIMIT, DECLARED RATHER THAN HIDDEN. PENDING-IN-FILE cannot be split into "not written yet" versus "declared somewhere ELSE while this row points here" without a tree-wide declaration index. That index is a named next leg; this gate does NOT guess at it, and its absence is why PENDING-IN-FILE is reported as its own axis instead of being folded into the defect count. The partition SUMS over all four classes, and the sum is printed, because a partition you cannot reconcile is a leak. ADVISORY BY CONSTRUCTION ON THE RISE, RATCHETED ON THE FALL. A deliberate hold is legitimate (LR7 held its