nx_lostset_gate.nx

nx_lostset_gate.nx -- THE REFEREE for nx_registrymine `lostset` (DM3, devmgmt rung). END-TO-END: it forks the DEPLOYED subject elf, so what it grades is the artifact callers actually run, not a re-compiled copy of the source. Pass the subject path as argv[1] to grade a candidate. WHY EACH TOOTH EXISTS -- every one of these is a way this miner could report a confident wrong answer, and three of them are defects that were REAL in the input before the miner was written: * the harness ships `deferred_tools_delta` attachments that LIST tool names as structural JSON, so a bare-prefix scan counts listed-but-never-called tools as execution evidence; * prose in a tool result can DISCUSS mcpMeta, so a naive needle counts a memory note as a fork; * a resumed session replays earlier records, so a per-file count double-counts one call. A tooth that only ever sees correct data has not been shown to fire, so the fixture carries the bad case and the good case AT ONCE and the gate asserts they are told apart -- two runs that each isolate one signal do not prove discrimination between them. FIXTURES ARE ASSEMBLED AT RUNTIME under /tmp/nx_lostset_gate/ and never in a production tree: a gate that shares its scratch with a live path reports on the FIXTURE, not on the code (measured estate-wide, the gate_fixture ratchet). Idempotent: every fixture file is opened O_TRUNC, so the second run grades the same bytes as the first. license_tier: ORIGINAL Read-only outside /tmp. No hw writes (Rule 26).