- **WHERE** (conformance vectors): extends `conformance/vectors/gen_all_formats.py` with dedicated reference codecs that promote 6 packs from structural (`bitexact: false`, placeholder) to bit-precise (`bitexact: true`, `abs_error = 0` for every recorded vector): `ibm_hfp32/64/128` (IBM Hexadecimal Floating Point, base-16 exponent, layout S1:E7 excess-64:M, value = 0.M(2) * 16^(E-64)); `x87_fp80` (Intel 80-bit extended with the explicit integer bit as the MSB of the 64-bit significand field, S1:E15:SIG64, bias 16383); `nf4` (QLoRA/bitsandbytes NF4 16-entry quantile table over [-1,1], exhaustive 16-code lookup); `gfternary` (2-bit {-phi, 0, +phi}, exhaustive 4-code enumeration). Adds `make_ibm_hfp_decoder`/`ibm_hfp_encode_exact`, `make_x87_decoder`/`x87_encode_exact`, `make_nf4_decoder` + `NF4_TABLE`, `make_gfternary_decoder`, and routes them in `build_decodable`/`build_bitexact_pack`. Regenerates `INDEX_all_formats.json` (`total_packs=83, bitexact_packs=55, structural_packs=28`) and the 6 affected `*_conformance_v0.json` packs. `README.md` index, counts, SHA-256 block, and a new changelog entry updated. The catalog itself is unchanged at 83 formats; only round-trip depth increases. Follows #1070.
0 commit comments