nx_wat_refuse_gate.nx

nx_wat_refuse_gate.nx -- THE REFEREE FOR LN13a: THE WAT LANE MUST REFUSE WHAT IT CANNOT LOWER. WHAT THIS PROVES, AND WHY A GREEN HERE IS WORTH SOMETHING. Until 2026-09-04 the wat backend answered an IR opcode it had no arm for with a WAT COMMENT and `return 0`, without storing a result local. The module assembled, instantiated and RAN, silently computing zero at that instruction. Nothing in the build said a word. That is the worst shape a compiler defect can take, and it is invisible to every instrument that reads exit codes or artifact sizes -- the pre-fix compiler exits 0 and emits a perfectly well-formed module. So the only referee that can see it is one that compiles a KNOWN-UNLOWERABLE construct and demands a NAMED refusal. THE SUBJECT IS A BINARY, NOT A SOURCE. argv[1] is the compiler under test, so this gate can be pointed at a staged build before it is promoted (the nx_cc_equiv_gate / nx_langdiag_gate shape). It judges the compiler by the BEHAVIOUR of what it emits, never by inspecting its assembly. THE NEGATIVE CONTROLS ARE THE LOAD-BEARING HALF. A compiler that refused EVERY module would pass every refusal tooth here and be catastrophically wrong; the estate has bite-proved exactly that class on nx_langdiag_gate. So the direct-call control must still compile, still emit a module, and still NOT be refused. Without those three, this gate scores a guard-that-refuses-everything at 100%. FIXTURES ARE ASSEMBLED AT RUNTIME into a per-RUN directory from gk_fixture_dir. Two reasons, both learned the hard way: a gate that shares a production fixture tracks the FIXTURE and not the code, and two concurrent runs of a gate with a hardcoded /tmp path return false SKIPs. The old silent marker is built from two halves at runtime for a third reason -- written whole it would be a source literal, and a detector that scans sources finds its own fixture. license_tier: ORIGINAL. Forks the subject compiler, writes only under its own per-run /tmp dir. No hw writes (Rule 26).