nx_bench_receipt_lib.nx

nx_bench_receipt_lib.nx -- THE ONE READER, ONE VERDICT AND ONE SERIALISER for a /compare <dom>.bench receipt (2026-09-01, lang leg: "meet and exceed gcc and rust, independently verified, reproducible, visible"). WHY A LIB. A benchmark receipt is read by THREE organs -- the writer that measures (nx_lang_h2h), the two /compare generators that render it (bench_pass in nx_swcompare_lib), and the gate that judges it (nx_lang_h2h_gate). WHEN THREE ORGANS MUST AGREE ON WHAT "VALID" MEANS, MAKE DISAGREEMENT IMPOSSIBLE BY CONSTRUCTION: there is exactly one classifier (br_verdict), one row grammar (br_parse / br_write) and one ratio rule (permil of the reference arm's median). The writer calls br_verdict BEFORE it serialises and the reader calls it AFTER it parses, so a receipt whose written @verdict disagrees with its rows is detectable. THE RECEIPT (knowledge/compare/<dom>.bench), a data file the emitter renders for ANY domain: @title <text> @workload <text: what every arm computes and the checksum it must print> @host <hostname>|<cpu model>|<kernel>|nproc=<n> @runs <n> @ref <arm name that is the 1.00x baseline> @asof <epoch> @repro <the exact command that regenerates this file> @writer <organ> @verdict VALID|VOID|UNMEASURABLE|EMPTY (written by the writer, RE-DERIVED by every reader) arm|<name>|<toolchain>|<version>|<source>|<source_sha256>|<bin_bytes>|<bin_sha256>|<runs>|<min_us>|<median_us>|<checksum>|<status>|<note> STATUS per arm: VALID (measured, checksum equals the reference arm's), VOID (measured, checksum DIFFERS -- the arm did not do the same work, so its time is not comparable and is never ranked), UNMEASURABLE (the toolchain is not declared on the measuring host -- an absence, never a loss), BUILD-FAIL, RUN-FAIL. VERDICT for the receipt: EMPTY (no arm rows), UNMEASURABLE (the reference arm is not VALID, so no ratio exists), VOID (any measured arm disagrees on the checksum), VALID otherwise. A ratio is permil of the reference median: 1000 = 1.00x, 2020 = 2.02x slower. Only VALID arms carry one. NOTHING HERE IS A THRESHOLD: the lib says what was measured; the gate says what is acceptable. license_tier: ORIGINAL No hw writes (Rule 26). LIB (no main).