nx_forge_engine.nx

nx_forge_engine.nx -- THE ASSEMBLED AUTONOMOUS-CODING ENGINE v1 (pure core; CLI = nx_forge_engine_run, gate = nx_forge_engine_gate). One invocation turns a task spec into judged NishiLang code using every proven lever, in measured-lift order: STAGE 1 GREEDY + REPAIR: the gated A1 closed loop (nx_forge_loop) -- deterministic generation, byte-exact judge, verbatim-stderr re-prompt, bounded cycles. STAGE 2 BEST-OF-N SAMPLED (the Large-Language-Monkeys / dossier-N3 lever): if the loop exhausts, fan N independent sampled candidates ("temp">0, DETERMINISTIC per-sample seeds FE_SEED_BASE+k so engine runs are reproducible) against the FREE byte-exact verifier -- first GREEN wins. No repair ctx between samples (independence). Relationship to the sibling nx_forge_bestof (2026-07-16): that organ is the IN-PROCESS 1.5B bench arm (imports serve core, prefix-cache); THIS is the API-native engine stage over HTTP /gen per the operator's no-one-off-shell directive -- the serve already parses temp/top_p/ top_k/seed (gp[8..11]) so sampling needs zero serve surgery. rc: 1..7 GREEN at loop cycle N · 100+k GREEN at sample k · -40 both stages exhausted · hard errors pass through from the loop (-2/-3/-4/-22/-30/-31/-33/-34). Engine clients do NOT claim the model lane -- the DAEMON launcher owns the RAM claim (M6 law). license_tier: ORIGINAL