nx_goalmap_lib.nx

nx_goalmap_lib.nx -- THE GOAL MAPPER (seat-system, 2026-09-05). Operator: "build our mapping setup to exceed yours so we dont need to reinvest heavy tokens and instead the nishi ecosystem can do the lifts." MEASURED the day it was built: two Claude workflows spent 5.7M tokens and 1,272 tool calls to map ONE operator goal onto the boards. This lib does that lift from DATA, deterministically, with no language model in the path: goal WORDS -> per domain, the rung| rows whose title / done-rule / symbol carry them (score = DISTINCT words matched, published as a COUNT and never thresholded -- the capsearch law: a ruler that scores a novel thing highest cannot be thresholded, publish the list), each rung's STATUS derived from the matrix (a present symbol = PRESENT, an _ABSENT_:<sym> watch = OPEN, neither = UNMAPPED) and from the operator's own plane (knowledge/store/accept-: a REJECT whose targets name <dom>:<sym> = PRESENT-REJECTED, an ACCEPT = ACCEPTED), the log| ledger size and last epoch per rung, the matrix WATCH rows the words match (gaps, with SOURCE / BUILT / PROMOTED probed by open at the three roots), and a dependency ORDER over the matched rungs (Kahn over the deps column; ties by score then file order; a cycle is ANNOUNCED, never hidden). Output is ONE capped JSON digest (GM_OUTCAP declared in the output, clipped=1 when the budget bound it) so a seat reads the decision surface, not the boards. COMPOSES nx_sovjson_lib (emit + spans) and nx_store_seed_lib (plane load sized from the plane -- never a partial board). Every read declares its true size beside the window it took. A duplicate rung id is REFUSED BY NAME (the ranker's closure hides a rung behind one; measured 'rung|GE50|' x2 on gameengine.plan on 2026-09-05). No hw writes (Rule 26). gm_run(base, domains-csv, words, nwords, out) -> bytes written, or -3 NO-DOMAIN-READABLE, -4 DUPLICATE-RUNG-ID base = path prefix for a fixture tree ("" in production = the nishihost cwd).