A step now counts a rule, a condition class, a split's suffix lookup,
16 flags of a set looked through and a group visit, so hostile rules
whose steps were cheap to count but costly to run hit the budget in
time. Retuned: 2^13 steps per candidate, 2^25 per suggest call; the
81 packs' typo sets suggest exactly as at d0c341c8.
Condition character sets are sorted and looked up by binary search.
A stem's sets are worked out once per distinct entry list when its
entries hold 256+ flags (16 kept-apart homonyms of 8,190 flags on
17,576 stems: 55 s load, now 57 ms). A near-word candidate whose
steps ran out counts as a word, so suggest never offers a compound
check() would refuse.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
Every TRY/KEY/typo candidate got a fresh 65,536-rule check, so a hostile
.aff with a 128-letter TRY line took ~20 s per suggest call. Now a
suggest call shares 2^23 affix rules across its simple-word checks, each
capped at 2^11; prefix-rule collection counts against the budget and
mixed-case lookups try at most MAX_HOMONYMS stems. Suggestion lists are
unchanged on the 77 packs and the recall sets.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
- simple_word sorts the grouped prefix flags (non-cross then cross flags were unsorted, so
`redo`/`rework` from a no-cross prefix beside cross ones were missed).
- One budget of 65,536 affix rules per simple-word check, shared by every prefix group and
suffix walk; past it the word is taken for no word. Prefixes that each strip their own
text (h2.aff: 1,000 roots) no longer cost a full walk each: 1.7 s → 1.1 ms per two
words, suggest 14.9 s → 65 ms. The 77 fetched spelling packs answer as without a budget
on 6,000 generated words each.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
- simple_word groups the prefix rules that leave the same root (circumfix and NEEDAFFIX
alike) and answers them in one stem lookup, one chain walk and one suffix walk with all
their flags (FlagProbe): 1,000 or 16,000 prefixes adding nothing no longer cost one walk
each (h1.aff: 1.6 s → 17 ms per two words; 16k prefixes 22 s → 19 ms).
- Deferred suggestion candidates are not looked up again before their compound check.
- A word in capitals matches a mixed-case stem as a simple word only (`JQUERYKABEL`,
`MSCFOO` wrong, as in the CLI).
- to_bytes writes the flags only the kept-apart homonyms use (NEEDAFFIX beside a compound
flag survives a round trip).
- MAGIC WCAFF2: compact data from before homonyms were kept apart is refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
- A prefix with two chained suffixes (hu `meg`+`épít`+`és`+`e`, eo `enjugo`): every rule
cross, the prefix flag from the stem or a continuation class (CLI answers).
- Homonyms with different flags are kept apart (and written apart in the compact form):
a prefix and a suffix, or a compound flag and a suffix, must come from one homonym
(nl `afstopje`, `schuurpoederden` wrong in the CLI). nl now 100% in both comparisons.
- Hungarian syllable rule leaves out a last suffix with no continuation classes;
COMPOUNDROOT counts on the stem only.
- Near-word check: a word listed as an ONLYINCOMPOUND/NEEDAFFIX stem counts only with a
suffix, not a prefix (de `Fahrradverkehr` right).
- A word in capitals may be a mixed-case stem (`KB`/`KBS` from `kB/S`, hu `DNY`, nl `BSC`).
- Mixed-case retry keeps KEEPCASE for the first part; KEEPCASE kept apart from the other
exclusions (NOSUGGEST sharing its flag is no longer cleared); overflow of the homonym
groups counts as dropped.
- Suggestions: candidates are checked as simple words first; the rest get the compound
check cheapest first, those changed where the word stops being parts before others, in
one step budget per call, with piece analyses shared between candidates. The suggestions'
stem scan makes forms in one shared budget (mn).
- affix_check example: `suggest` command; AffixDict::suggest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
- ONLYINCOMPOUND homonyms beside others are kept apart (compound parts only), so their suffixes
make no standalone words (de verfilmungs, sv norges); round-trips as a second stem line.
- CHECKCOMPOUNDTRIPLE also checks a SIMPLIFIEDTRIPLE join (busssen, booob wrong).
- KEEPCASE holds for the first part only; a capitalized stem is a part (FooQux).
- CHECKCOMPOUNDPATTERN `0`: no suffix (a prefix doesn't count): unfoobar wrong.
- COMPOUNDFORBIDFLAG on a prefix: no part anywhere (bardefoo wrong).
- FORCEUCASE: from the stem only; wrong only without any capital (fooQuxbar right).
- Suggestion candidates: 2,048 compound steps each, compounds for the first 128 only.
- A piece's ways are hashed, capped at 256 and counted (with the patterns) against the budget.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
Black-box: every rule pinned to hunspell 1.7.5's answers on synthetic dictionaries and on
the Language packs' own files. Bounded work per word (16 parts, 65,536 steps). Also:
ONLYINCOMPOUND belongs to its homonym like NEEDAFFIX; COMPOUNDRULE reads the Swedish `)k`
and the Mongolian `[..]`, `%?`, `.` as the CLI does; COMPOUNDMORESUFFIXES is refused only
with flag-based compounding (Korean keeps loading).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
The total flag-set bound (MAX_SET_DATA, 8.4M) stays the hostile-input guard: LibreOffice's
tr_TR loads with 5.2M of it in 0.6 s, and abart, abartmıştım, savaşıyor, inandım, aranıyor,
açılamadı and bilinmiyor are right, as hunspell 1.7 answers. The merged-homonym cap is now
4 × the entry cap (32,768), under the 65,536 distinct flags.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
hunspell::candidates_exact (folder → exact-locale pack → system, own names only) for English without SCOWL; packs::exact_spelling_dirs. Other languages keep the region-ordered cross-region tiers. Drops has_british_source().
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
- build.rs: unset CRAFT_PROOF_EMBED embeds de, pl, pt, it patterns and ru-extra as before (DEFAULT_EMBED); `en` embeds English and our supplements only; `de,pl` and `all` as before. Making `en` the default is CP4.4, in the release that publishes and pins language-packs-1. The default-build data tests cover pl/it/de/pt again; CI's extra step is now the English-only build.
- licence_texts: one Moby entry (the data's, else the built-in en-words.dic) and en-extra.txt (WordCraft contributors, MIT OR Apache-2.0); the deleted core-list notice is gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
The goitre failure was not a race: GitHub's Ubuntu runners have a system en_US Hunspell dictionary, which en-GB uses without the Language data (by design), and it marks goitre. Reproduced locally with a fake system en_US; the test now checks British words only when en-GB reads the Moby list or a British dictionary.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
P2:
- Every test whose English results depend on installed dictionaries takes DICTIONARY_TESTS (the en-GB goitre CI failure: tiny en_US fixtures raced unlocked en-GB tests).
- hunspell::candidates/find order: dictionaries folder, the first name's region pack (or the language alone), the system's exact-region dictionary, any other region's pack, the system's other names. pt_BR system beats a pt-pt pack; de_DE / de-ch, nl_NL / nl-be, es_ES / es-mx, en_US / en-gb likewise.
P3:
- Pack hyphenation lookups: a map, at most 256 languages (cleared when full); 3000 tags stay bounded and fast.
- A decomposed word (a + U+0301) is hyphenated composed (NFC) and the breaks mapped back; combining marks stay in their word.
- licence_texts checks each file's size against what is left before reading it; never reads past 16 MB.
- German, Bulgarian and pack languages: spelling message in the language where we have it; missing_dictionary/unreadable_dictionary report them like the others.
- build.rs: a missing file no longer says nothing is embedded; ru ignore message says CRAFT_I18N_DIR is enough. Built-in English list notice names the whole Moby list (en-words.dic).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
hunspell::candidates (which English now loads from) searches the pack tier between the dictionaries folder and the system, as find() does; pack-loaded English goes through the same reloads guard.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
Languages without tools of their own (nl, fr, sr-Latn…) are spell-checked and given suggestions
through the shared affix engine with default tuning, and hyphenated from their pack's patterns;
no grammar. At most 8 such dictionaries stay loaded. Never another script's pack.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FypgCX8LsYYE8fsj4byBm2
