Merge upstream LeanBitLab/LeanType v4.1.2

[AsafMah]

Merges 200 upstream commits from LeanBitLab/LeanType, covering v4.0.9 → v4.1.2, pinned at tag SHA 8720abebd944eee38dcea0bee71bbd755fb51bb0.

Conflict decisions

File	Resolution
app/build.gradle.kts	ours — fork keeps 4200 / 0.2.0, never upstream's 4102 / 4.1.2
README.md, docs/FEATURES.md	ours — fork-identity docs (our badges, two-thumb as namesake, our sponsor links)
docs/badges/*.svg	ours, per the fork invariants
fastlane/.../changelogs/4100.txt	ours — upstream reused the same versionCode; this is our 0.1.0 note
LatinIME.java	both — fork's SwipeGestureEngine init/cancel plus upstream's sInstance tracking and voice-manager release
SettingsValues.java, strings.xml	both — fork's flag-unknown-words + graduated-trust, plus upstream's personal-dictionary learning threshold
DictionaryFacilitatorImpl.kt	both — fork's blacklist guard against auto-promotion, plus upstream's auto-capitalization decapitalize handling
ClipboardHistoryManager.kt	ours — both define updateClipText with different signatures; ours returns the deleted entry so the caller can offer undo
SuggestionStripView.kt	both, with upstream's recycled-view restore cancellation ordered before the fork's uncurated-word dialog so it can't be skipped by the early return
ClipboardHistoryView.kt	theirs per hunk — upstream adds a clipboard edit mode; the fork side was only reformatting
MainSettingsScreen.kt	@Preview arity updated to the merged 15-parameter signature

Two fork regressions found and fixed

Both came from automatic conflict resolution, and neither was caught by the compiler — only by the fork's own tests. This is the concrete argument for the "always audit Mergiraf resolutions" rule.

1. Whole-word backspace left partial words behind (5 failing tests)

Upstream 0cf92cf1b ("prevent single-click backspace from bulk deleting numeric sequences") split the end of the composing-word backspace branch on wasBatchMode. The fork's two-thumb whole-word delete also ends with no composing word but is not batch mode, so it fell into the new else — which calls deleteTextBeforeCursor, exactly the path its own comment warns against. That produced the "This is pretty cool" → "precool" corruption the fork had previously fixed.

Fixed with an explicit wholeWordDeleted flag so whole-word and batch both clear the composing span, while upstream's numeric-sequence fix still applies to the ordinary last-character path.

2. Custom layouts stopped restoring after symbol mode (1 failing test)

Upstream added lastCustomIndex = 0 when switching symbols → alphabet. This fork ships persistent custom layout slots meant to survive symbol mode, orientation changes and reloads (shipped in 0.1.0), and KeyboardStateTest asserts it. Dropped the upstream line.

Both inherited upstream defects are fixed — guards removed

HANDOFF §7 recorded two defects reproduced on a pristine upstream checkout and guarded with runTests skips. This merge fixes both, verified twice rather than assumed:

• On the upstream check-out worktree re-pointed to v4.1.2, SubtypeTest passes (3 tests, 0 failures) — including subtypeStaysEnabledOnEdits.
• In this merged tree, both subtypeStaysEnabledOnEdits and immediateRegexExpansionTriggersForSymbolPrefixedRegex pass.

Both runTests guards are removed. Worth calling out: those two tests were previously returning early on CI, so runTests was green on them for reasons unrelated to correctness. CI now genuinely covers them.

Verification

• compileOfflineRunTestsKotlin — passes
• :app:testOfflineRunTestsUnitTest (the CI variant) — 324 tests, 4 failed: exactly the documented Windows-only ParserTest baseline that passes on Linux CI. Unchanged from origin/dev.
• :app:testOfflineDebugUnitTest (full suite, both sides on the same machine, minutes apart, 2026-08-20):

	tests	failures
origin/dev (6ac372de3)	320	12
this branch	324	5

Remaining on this branch: tapOnlyCombiningWordDoesNotShowAutospaceIndicatorWhenGestureGateEnabled and ParserTest ×4 (the same four the CI variant reports).

Seven tests stopped failing. Split by how confidently that can be attributed to this merge:

• Attributable — subtypeStaysEnabledOnEdits and immediateRegexExpansionTriggersForSymbolPrefixedRegex (both independently confirmed fixed on a pristine upstream v4.1.2 checkout), and insertLetterIntoWordHangulFails (deterministic, no external inputs).
• Plausible but unconfirmed — StringUtilsTest ×2 and ParserTest > backgroundType. Both depend on bundled data/assets that this merge changes, so a real fix is likely, but it wasn't isolated.
• Not attributable — XLinkTest > otherLinks reaches Codeberg over the network and can flip with no code change at all. It is not counted as a fix.

Two corrections to earlier revisions of this description: it once quoted "4 failures → 1", which came from a run filtered to *InputLogicTest* + *SubtypeTest* and was never a whole-suite number; and it once claimed the merge "fixes seven", which overstates attribution for the three items above.

Failure names were diffed against an origin/dev baseline run, not pass counts.

Fork invariants re-verified after the merge: applicationId com.asafmah.leantypedual with its flavor suffixes; version 0.2.0/4200; INTERNET declared only in the standard and standardfull manifests; SwipeGestureEngine wired into LatinIME; the two-thumb module still registered in SettingsContainer; a single InitializationProvider.

Also fixed here

A latent test-harness bug: setText accepted a requireIdle parameter and then called handleMessages() without it, so reset()'s setText("", requireIdle = false) still asserted an idle queue. Harmless until this merge shifted JUnit's hash-based method ordering, after which it failed corpus - default mode char-by-char backspace during setup rather than in its own logic. (The same one-line fix appears independently in #134.)

Merge order

dev still carries the guards this PR removes, and #138's §7 documents them as gone — so merge this before #138. #134 and #136 both touch InputLogic.java and will want a small rebase afterwards.

Follow-ups

• ⚠️ Not yet verified on a device. Worth a smoke test of clipboard edit mode, two-thumb whole-word backspace, and custom layout switching before release.
• The emoji accelerated-delete bug fixed in fix(input): correct the accelerated backspace emoji measurement #134 still exists upstream and is reported as LeanBitLab/LeanType#423.