nx_bright_rewrite_lib.nx

nx_bright_rewrite_lib.nx -- R0l: the SOVEREIGN reasoning-rewrite run for BRIGHT as a LIBRARY (search R0l, 2026-09-16). WHY. The referee replay (R0o) proved the rewrite lever is real at 1.5B: the authors' own TongSearch-QR-1.5B rewrites score 229 permil through the estate's plain BM25 against the 149 floor. That number was produced by a query another model wrote, so it can never place the estate on the board. This library produces the rewrites on the estate's OWN engine (nx_nofloat_serve_core, the Q8_0 GGUF twin of the MIT weights) with the paper's own prompt, so the queries it writes are the estate's system and nx_beir_eval may score them under the pre-declared accept rule: within 30 permil of the 229 replay, or the port is not the arm. TWO ENGINE DOORS, ONE RUN (operator 2026-09-16: "through mcp and apis and primitives and workflows and agents not wsl ... from the first byte up"): in-process gguf=<path> the model is loaded into this process (nsv_init_i8) -- the laptop lane. remote engine=a.b.c.d:port every prompt is POSTed to a sovereign nx_nofloat_serve on the LAN (POST /gen, the same JSON contract the app page and nx_engine_dispatch use), so a NAS plan row drives a compute node's engine and the NAS never runs the model. The remote door composes nx_http_client (request), nx_http_response_parse (reply) and the serve core's own JSON helpers (nsv_jesc / nsv_jkey / nsv_jstr / nsv_jint) -- one escaper, one key finder, no second ruler. A per-prompt REFUSAL (the engine answered 400 with an err code: too long for its window) is written as the ORIGINAL query and counted as passthrough, exactly as in-process. An ENGINE FAILURE (unreachable, no reply, an unparseable reply, or err=1 which means the request itself was malformed) is NOT a refusal: it ABORTS the split (state ENGINE-DOWN) so a dead node can never manufacture a "complete" file of original queries. IDEMPOTENT RE-RUNS, so a clock agent may call this every beat: * a split whose final queries.<out>.tsv already exists is skipped as COMPLETE-EXISTING (no engine call); * a split is run under a lease (nx_lease_lib, name brightrewrite-<out>-<split>, ttl lease_ttl); a second runner on the same tree reports LEASED-ELSEWHERE and moves on, so two runners take disjoint splits; * rows already in the .part file are skipped by count (the same resume as before); * a run whose every written row was a passthrough does NOT rename to complete (ALL-PASSTHROUGH): an engine that refuses every prompt has produced nothing, and a file of originals must never be scored as rewrites. Per split it appends <root>/<split>/queries.<out>.tsv.part one row per query as each rewrite lands and renames the part to queries.<out>.tsv only when every query of the split has a row. limit>0 is a SMOKE run: it stops after that many queries per split and never renames, so a smoke file can never be mistaken for a full one. Progress: <root>/../../status/bright_rewrite.<out>.progress (or progress=<path>) is truncate-written after every query (a ts that stops moving while the process lives is a hung run, not a slow one). license_tier: ORIGINAL. No hw writes (Rule 26). LIB (no main).