nx_swebench_local_gate.nx

nx_swebench_local_gate.nx -- grade our autonomous fix loop by SWE-BENCH'S OWN CONTRACT (operator 2026-07-16), NOW PUBLISHING THROUGH THE RIGOR ENVELOPE (autodev AD1, 2026-08-27). SWE-bench = 2,294 real GitHub-issue problems (500 human-verified) across real Python repos. The contract: an instance is RESOLVED iff, after the patch, FAIL_TO_PASS tests pass AND PASS_TO_PASS tests still pass. The test suite is the only judge. Our nx_autofix_auto loop already IS this contract sovereignly: a candidate's per-function grader rows are FAIL_TO_PASS (the located failing function) + PASS_TO_PASS (every other function). HONEST SCOPE: a sovereign NishiLang ANALOG under SWE-bench's contract and metric over OUR foreign-bug set -- NOT a run of SWE-bench's Python instances. WHAT CHANGED (2026-08-27): the D001 migration that shipped here set ctr[1] = the EPISODE COUNT of the newest batch, so wherever the ledger existed this gate printed "passed 4/14 verdict=RED" -- FOUR of its own teeth against FOURTEEN episodes, a denominator mismatch that read RED on a healthy loop. ★A GATE'S PASS/TOTAL IS ITS TOOTH COUNT, NEVER ITS SUBJECT'S POPULATION. It is now per-tooth gv_check, and the resolve rate is published as a gv_envelope STRUCT (k, n, Wilson + hierarchical bootstrap, seed, harness) rather than a bare percentage -- the same ruler /compare/autograde and /compare/gen now share. SKIP when the ledger is unreachable (a host without the fix-loop stage is not a failure). expect_exit: 0 license_tier: ORIGINAL