feat: bump suggestions to 7, filter exact input from fuzzy, ignore-ca…

…se word check; bump v1.16.0

- SuggestionStrategy: add DEFAULT_SUGGESTION_COUNT=7 constant; filter exact
  input word (ignore-case) from fuzzy candidates so it doesn't consume a slot
- CodeKeyboardIME: use DEFAULT_SUGGESTION_COUNT instead of hardcoded 5 at all
  3 suggest() call sites
- HinglishIntegrationTest: use DEFAULT_SUGGESTION_COUNT in proximity test
- bump versionCode 26→27, versionName 1.15.3→1.16.0

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EBK4zLARnqALg3z2aUxma8

chore: bump to v1.14.3

v1.14.2 tag's build failed CI before the PhraseCompleterTest skip
landed; re-cutting as v1.14.3 rather than rewriting the published
v1.14.2 tag.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QrHkvaupGo88j5CykWYAtY

fix: abandon composing when cursor moves off the composing tail

ComposingBuffer only supports append/backspace at its own tail, so a
click or cursor move landing inside an active composing region (but
not at its end) left the buffer editing at the wrong position — the
next setComposingText(word, 1) call would then snap the real cursor
to the end of the word, ignoring the user's tap.

Tighten onUpdateSelection's cursorOutsideComposing check to require
the cursor sit exactly at the composing region's end; anywhere else
inside it now abandons composing instead of silently keeping stale
state.

Bump to v1.14.2 (patch: bug fix, no new features).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QrHkvaupGo88j5CykWYAtY

fix: weighted edit distance ranking for proximity-based fuzzy correct…

…ions

- FuzzyResult carries weightedDistance (half-steps/2) alongside integer editDistance
- BevaTrieSearch fills weightedDistance=halfDist/2f per match
- SuggestionStrategy dedup and sort use weightedDistance; adjacent-key errors
  now rank above uniform edits at the same integer threshold
- FuzzyIntegrationTest awarxh test uses real LayoutAdjacency with Sofle-approximate
  coordinates instead of QwertyAdjacency; asserts search>award ranking
- aws_ngrams: 4gram-full variant on c5.4xlarge with combined OpenSubtitles+SwiftKey
- bump v1.14.1 (versionCode 20)

feat: raise fuzzy correction threshold to 3, fix BEVA boundary bug, a…

…dd LayoutAdjacency

- Fix BEVA pruning: ev[k]==0 was too aggressive, now ev.none{it!=0} so
  branches with active lower-cost positions are not cut prematurely
- Raise FuzzyThreshold to 3 for words len>=5 (was 2); adjacent-key subs
  cost 0.5 half-steps so 6-key hand-shift errors now fit within budget
- Add LayoutAdjacency: derives key adjacency geometrically from actual
  PositionedKey center positions (Sofle/Ferris Sweep layouts), wired
  into WordDictionary via updateAdjacency() called from CodeKeyboardIME
- Update tests: brute-force comparison tests use Int.MAX_VALUE cap;
  add QWERTY right-shift integration tests (adwrxg, drstvj → search)
- Bump version 1.13.1 → 1.14.0 (versionCode 18 → 19)

feat: CKLM language pack as primary prediction path (v1.13.0)

- Ship SwiftKey-WDP n-gram model in a single binary .cklm pack (22 MB)
  replacing three separate JSON files; APK shrinks 302 MB → 94 MB
- LanguagePack: mmap reader for vocab, char-trie, context-trie, phrases
- PackNgramModel + PackBackedBigramModel replace JSON-backed adapters
- Ngram cascade: Gboard-style weights (static 0.7 / user 0.3), MIN_USER_DEED=3,
  self-referential continuation filtering
- Legacy JSON path, Trie.kt, Trigram.kt, BigramModelAdapter, FuzzyTrieSearch
  (BaseTrieAdapter), kn/ + kz/ assets all removed
- Three n-gram build variants (KN, Katz, SwiftKey-WDP) + AWS runner scripts
- Autocomplete eval harness (ADR-007), absolute-discounting bigram seed (ADR-008),
  CKLM compiler (ADR-009/010)
- Sentence Case arm-on-space fix, snippet expansion via Enter/Space, RN autosave

Merge pull request #3 from ikouchiha47/features/improvements

feat: Sentence Case, remove dead caps state, comma/period Shift fixes (v1.12.0)

chore: bump version to v1.11.2

Needed a fresh tag push to actually exercise the CI permissions fix (a
re-run of the failed v1.11.1 run reuses that run's original workflow
snapshot, not the corrected one).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

chore: bump version to v1.11.1

v1.11.0's CI build never actually succeeded (pre-existing broken commit
compile + stale test, both fixed in the prior two commits) so nothing was
published under that tag. Patch bump, no new feature in this fix-forward.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

chore: bump version to v1.11.0

New feature (Colemak-DH keymap) alongside the Dvorak table fix and Ferris
Sweep padding fix — minor bump per semver.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>