nx_forge_model.nx

nx_forge_model.nx -- FORGE F3 rung 2: the MODEL BACKEND for the driver. Turns a model serve's HTTP /gen response into a candidate .nx file, and composes the prompt the model sees. Same loop as rung 1 -- the backend just PRODUCES the candidate that fg_run then judges byte-exact. fm_compose_prompt(pack, task, repair, out, cap) pack v1 + task statement + (optional) the verbatim repair context from a prior failed cycle -> prompt file. fm_build_gen_body(prompt, plen, max_new, mode, out) the serve's POST /gen JSON body. fm_post_gen(a,b,c,d, port, body, blen, resp, rescap) TCP POST /gen to a.b.c.d:port over the proven nx_http_client primitives; returns response bytes (header+body). fm_extract_text(resp, rlen, out, cap) pull the JSON "text" value, unescaping \n \t \" \\ \/ \r and \uXXXX (low byte) -> the raw model output. -1 if absent. fm_strip_fences(src, n, out) drop a leading ```lang fence line + trailing ``` (the pilot's hand extraction stage, now an organ). Idempotent on unfenced text. fm_stub_serve_once(port, respfile) an in-process single-shot serve (gate fixture): binds, accepts one client, replies with respfile's bytes as a /gen JSON -> lets the gate prove the REAL loopback round-trip with no external model. Rung 2 proves the model-stage mechanics sovereignly; pointing fm_post_gen at 127.0.0.1:11434 (the no-float serve) or an API egress is a live address swap, not new logic. license_tier: ORIGINAL