nx_bench_ingest.nx

nx_bench_ingest.nx -- THE GENERIC BENCHMARK INGEST (rule 15 DRY; the second half of nx_bench_fetch). WHY: nx_swebv_ingest and nx_gmmlu_ingest are ~90% the same body -- an escape-aware JSON row walker whose only real difference is WHICH FIELD NAMES it pulls. Baking field names into source means a new organ per benchmark, and (per the ~119 *_research_fetch family) every copy inherits the original's latent bugs. Here the field names are ARGUMENTS, so a new card axis is a DATA row, not a new organ (rule 11). nx_bench_ingest <rawpath> <first_field> <id_field> <q_field> <opts_csv|-> <a_field> <taskpfx> <goldpfx> first_field = the FIRST feature of the row object; the row anchor is "row":{"<first_field>":" opts_csv = e.g. option_a,option_b,option_c,option_d (or `-` for a non-multiple-choice set) emits taskpfx : id <TAB> question [<TAB> opt]... (NEVER the answer -- contamination split) goldpfx : id <TAB> answer Same correctness core as the bespoke organs it replaces: escape-aware value end (terminating '"' = one preceded by an EVEN run of backslashes, so text containing \" cannot mis-terminate a field), row search BOUNDED to [row,next-row) so a malformed row cannot steal a sibling's fields, values stored JSON-escaped verbatim (TAB/NEWLINE-safe by construction), truncations COUNTED not silent, and fail-closed: task==gold and >0 or NEITHER store is touched. expect_exit: 0 (GREEN) | 1 RED | 2 usage license_tier: ORIGINAL No hw writes (Rule 26).