nx_codegen_bench_gate.nx

nx_codegen_bench_gate.nx -- OBJECTIVE code-gen grader (eats the anti-navel-gazing debt, 2026-07-15). nx_capability_triage graded code-gen = MEASURE(bench=none): we had NO objective way to grade generated code, so any "our coding is improving" claim was navel-gazing. This is the missing instrument: coding tasks with OBJECTIVE (input->expected) test cases; a candidate solution (a function = "generated code") is run against every case; PASS iff ALL cases match. The grade is not our opinion -- it is whether the code COMPUTES THE RIGHT ANSWER. The load-bearing ANTI-NAVEL-GAZING property (T2): the grader must FAIL wrong code -- a grader that passes broken solutions is exactly the navel-gazing the operator warned about. T1 CORRECT solutions PASS (the tasks are gradable; reference code scores 100%) T2 SOUND / anti-navel-gazing: BROKEN solutions FAIL (the grader rejects wrong code -- the whole point) T3 SCORE computed = pass-rate over candidates (the metric), persisted to a durable bench ledger (living) T4 DETERMINISM: re-grading is byte-stable The candidate here is a compiled-in function (proves the GRADER sound); the real GENERATOR (forge e1 coder -> candidate SOURCE -> nx_cc compile -> this grader) + a standard suite (HumanEval-style) = the next rungs that turn this into the external number proving "we reduce Claude coding". expect_exit: 0 license_tier: ORIGINAL