OpenCorpora/libmorph hand a noun reading to Stage 2 above the РАН note, so
substantivized-adjective false nouns a dictionary misreads (нёбный, акцизный,
велярный, …) reached scrabble.txt with no veto path — the earlier abbreviation
filter only carved out Abbr+Fixd.
ru_stage2.py now subtracts sources/scrabble_ru/manual_reject.txt last, after
every admission path (OC seed / libmorph / note / manual_confirm / variant),
mirroring manual_confirm.txt. The list seeds 62 hand-reviewed words — 43 pure
adjectives plus 19 marginal (slang/archaic/jargon) substantivizations; genuine
substantivized nouns (больной, знакомый, учёный, участковый, …) stay.
scrabble.txt -62 (83206->83144); erudit.txt re-folded -62. The DAWGs are
gitignored and rebuild from these lists in CI.
OpenCorpora tags indeclinable abbreviations as common nouns (Abbr+Fixd), and
Stage 2 seeded the result with its whole noun lexicon, so non-words leaked into
scrabble.txt: ндс, ст, ср, кпд, чп, гибдд, днк, …
ru_stage2.py now drops Abbr+Fixd nouns from the OpenCorpora seed and lets the
orthographic dictionary decide instead: a lowercase РАН headword whose note is a
noun is kept (the lexicalised сельпо, под, ска, роно, врио, тв, фио, суперэвм),
everything else is dropped. Function words (CONJ/PRCL, e.g. зато) and OpenCorpora
inflected forms (ан = род. мн. от «ана»; proper «Ан») are excluded too.
--trace WORD now prints a detailed per-signal audit (all.txt, OpenCorpora POS /
Abbr / Fixd / CONJ/PRCL, libmorph, РАН note + classify, outcome) for auditing a
word or repeating the analysis; tools/README.md documents the rule and method.
Net: 179 words removed from scrabble.txt (83385 -> 83206) and erudit.txt
(83343 -> 83164); verified to remove exactly those 179 and add none. DAWGs
rebuild clean.
Consolidate the scattered build inputs (dictionaries/english/, dictprep/russian/)
into one sources/ tree keyed by the variant labels (scrabble_en/scrabble_ru/
erudit_ru), and move the Russian prep pipeline to tools/. The dawg outputs and
their filenames are unchanged — rebuilt byte-identical (en_sowpods/ru_scrabble/
ru_erudit) — so the release artifact and the backend are unaffected.
ru_stage2.py OUT_DIR and the ruwords flag defaults are repointed to
sources/scrabble_ru/; Makefile / CI / cmd/builddict default / README updated;
pipeline intermediates git-ignored. Verified: make dawg byte-identical to the
committed baseline, py_compile + go vet of the moved tools. The full Russian
regeneration pipeline (pymorphy3/libmorph/orfo PDF) was not run here.
Ilia Denisov