v4 union split-section dictionary: both full dicts, zero trimming
Format v4 (docs/FORMAT_V4.md): one union CHD MPHF over 226,791 stroke
keys from Plover (147,424) + Lapwing (114,885), sections distributed
across halves. Left blob 387KB (displacements log-class Huffman 59KB,
membership 57KB, fingerprints 113KB, conflicts 9.6KB, value-index
slice 157KB); right blob 515KB (strings front-coded+deflate 193KB,
value-index remainder 324KB, conflicts dup). Shared string table:
73,464 unique translations, Lapwing 93% subset of Plover.
- tools/compile_v4.py: encoder + full verification (all 262,309
entries byte-exact through cold decode path; hard error if either
blob exceeds budget — never trims)
- src/dict_v4.c/.h: decoder — canonical Huffman displacement decode,
conflict binary search, FC block walk, 2-block LRU cache
- src/behavior_steno.c: Plover-style sliding longest-match with
retrace replaces multi-stroke timeout + prefix checks
- src/split_dict.c/.h: protocol v4 — GET_STRING(string_id) /
RESOLVE(slot, dict); decisions are 100% left-local, BLE only for
translation text
- CMake/Kconfig: STENO_DICT_BOTH default, one compiler run emits both
half blobs; legacy MPHF path kept behind STENO_DICT_V4=n
- Host round-trip test: 42,000 vectors left→right split path,
0 mismatches (tests/test_dict_v4.c)
afiqzudinhadi
9bf21c9db2