feat(android-ime): keyboard data export, fuzzy pinyin and separate learning stores (depends on #1230)

[DepengWang]

Depends on #1230. This branch is stacked on #1230's head (272be3ba), so until #1230 merges the diff here also shows its commits. Only the last commit (abb0969e) belongs to this PR. I'll rebase onto beta and mark it ready once #1230 is in.

Summary

Android keyboard only (openless-all/app/android/kotlin/ and the shared mobile-ime/ Rime wrapper); no Rust core, desktop or frontend changes, and the migration bundle format is unchanged.

1. Keyboard data export/import on the keyboard settings page (the haptics page), independent of the migration bundle and cloud sync.
2. Fuzzy pinyin switches.
3. Separate stores for stroke picks and association picks, with larger caps and batch trimming.
4. LitePinyin's leftover backup adapter removed.

1. Keyboard data export / import

Under "个人偏好数据", two actions write and read one JSON file (openlessKeyboardDataVersion: 1). A checklist picks the parts:

Part	Content	On import
Keyboard settings	keyboard height and raise, haptics (on/amplitude/duration), debug features, cloud pinyin, weekly public dictionary, fuzzy pinyin, English suggestions	replaced
Pinyin user dictionary	Rime pinyin_simp snapshot (.userdb.txt text)	merged by Rime
Stroke ranking	stroke code → picked character	merged
Association learning	text before → picked association	merged
Cloud notes	webhook URL and token (plaintext, labelled)	replaced; unchecked by default on import

• Merging keeps the higher count and the later use per entry and never deletes. Rime's own restore_user_dict merge does the same for the pinyin dictionary.
• Every part is validated before anything is written: known keys only, value types and ranges (e.g. height 220–420 dp, amplitude 1–255), stroke keys only in the stroke part, a snapshot whose header says pinyin_simp / userdb. One invalid item rejects the file.
• Settings that are also in the migration bundle (haptics, English suggestions, debug) are written through OpenLessMigrationState.settings. A changed fuzzy-pinyin set triggers a redeploy.
• This is deliberately not part of the .omb bundle or encrypted sync; app settings keep travelling through those.

2. Fuzzy pinyin

Seven switches under "拼音云候选": z c s = zh ch sh, n = l, f = h, r = l, an = ang (covers ian/uan), en = eng, in = ing. They are rime-ice's own commented-out rules, appended to speller/algebra through openless_pinyin_{learning,private}.custom.yaml in the user data dir. Both schemas get the same patch because they share one compiled spelling index. Toggles made while a redeploy runs are folded into the next one.

Native Rime (mobile-ime/rime_engine.cpp, shared with HarmonyOS)

• user_dict_export / user_dict_import: librime levers backup_user_dict / restore_user_dict on pinyin_simp.
• redeploy: removes the compiled pinyin schemas from build/ (so a same-second rewrite isn't seen as up to date) and deploys both again.
• All three first destroy every session. That commits the user dictionary transaction and releases the LevelDB lock that the levers tools need. OpenLessRime.sessionGeneration is bumped, and RimePinyinController starts a new session when its generation is stale instead of failing with "session is stale".
• A second installation_update run at initialization: on a fresh profile the first run only writes installation.yaml and leaves Rime's sync dir as the relative "sync" (the read-only process cwd on Android), so a backup failed until the next process start. Export also refuses a non-absolute sync dir with a clear error. Found by CI's fresh emulator.
• rime_require_module_levers() keeps the statically linked levers module from being dropped by the linker.
• rime_levers_api.h is copied unchanged from librime 1.17.0, because its install step only ships rime_api.h.

3. Separate learning stores

StrokeUserFrequency used to hold stroke-code picks and association picks in one store with a 3000 cap, so heavy use of one evicted the other. Now:

• openless_stroke_frequency keeps stroke picks (cap 5000), and the new openless_association_frequency keeps association picks (cap 10000).
• Existing data is split once by key shape (stroke contexts are spelled only with hspnz*), inside the migration fence.
• Past the cap a store trims to 90% in one pass instead of a full copy and sort for every new entry once full. Lookups go through FrequencyStampCache, so query cost doesn't depend on store size.
• The migration bundle and legacy JSON keep one combined strokeFrequency map (15,000 at most, within the 20,000 limit). On import it is split by key.

4. LitePinyin leftovers removed

LegacyPinyinBackup, its settings row and its legacy-JSON import category are gone. Its two SharedPreferences stores are deleted on first Rime initialization. Bundle schema 1 requires pinyinLearnedPhrases, so it is kept, always written empty and ignored on import. That keeps bundles readable by both older and newer apps without a schema change. local-bundle-contract.test.mjs, ImeLearningMigrationTest and OpenLessMigrationBundleInstrumentedTest are updated to match.

Testing

• :app:compileArm64DebugKotlin, :app:testArm64DebugUnitTest: 123 pass, including 10 new in RimePinyinDataTest (fuzzy patch, snapshot validation, key classification, entry validation and merging, trim target, setting ranges, webhook URLs, default checklist).
• :app:compileArm64DebugAndroidTestKotlin.
• node --test --test-concurrency=1 scripts/*.test.mjs: 92 pass.
• arm64 debug APK installed on a OnePlus CPH2573 and checked by hand: export and import of every part, fuzzy pinyin on and off, the split counters, and normal pinyin/stroke/English typing afterwards.
• New instrumented tests in RimeEngineIntegrationTest, not run locally (CI emulator): userDictionaryExportsAndMergesSnapshots, fuzzyPinyinPatchIsDeployedAndRemoved, controllerStartsANewSessionAfterSessionsWereClosed.
• HarmonyOS build with the rime_engine.cpp changes.

Worth a look

• Opening the keyboard settings page reads the dictionary size through an export, which closes the open Rime sessions; a half-typed pinyin composition is dropped (the keyboard is hidden at that point).
• The keyboard data file is plaintext and contains learned words, plus the token when that part is exported.

🤖 Generated with Claude Code