CI compiles tests but never runs them — 3 unit tests fail on main unnoticed

[AsafMah]

Summary

While reviewing #11 I ran the JVM unit tests (:app:testOfflineDebugUnitTest) on both main (origin/main @ c63c688f) and the PR branch. 3 tests already fail on main, independent of #11. They go unnoticed because CI never runs the test suite.

Pre-existing failures on main

Test	Failure
InputLogicTest.tapOnlyCombiningWordDoesNotShowAutospaceIndicatorWhenGestureGateEnabled	Mockito ArgumentsAreDifferent (verifies setCombiningMode(false, …) called and setCombiningMode(true, …) never)
InputLogicTest.insertLetterIntoWordHangulFails	org.junit.ComparisonFailure
InputLogicTest."revert autocorrect on delete"	org.junit.ComparisonFailure

Run on JDK 21 / Windows: 79 tests completed, 3 failed on main; the exact same 3 fail on #11's branch (87 tests completed, 3 failed). So #11 adds 8 passing tests and introduces zero regressions — these are old.

Why they're invisible

.github/workflows/build-test-auto.yml runs only compileOfflineRunTestsKotlin (compile-only). Tests compile but are never executed, so failures don't break CI.

Likely nature (needs triage)

• The two ComparisonFailures are classic LatinIME tests (Hangul insertion, autocorrect-revert) that may depend on dictionary/locale data not available in the JVM unit environment — possibly environment-sensitive rather than genuine product bugs.
• The Mockito failure is a LeanType two-thumb combining-mode expectation; could be a brittle verification or a real logic drift.

Proposed

1. Add a CI step that actually runs :app:testOfflineDebugUnitTest (at least non-blocking at first) so regressions surface.
2. Triage the 3 failures: fix, or quarantine/@Ignore with a tracking note, so the suite goes green and future regressions are detectable.

Minor side note (from #11 review, not a blocker)

The opt-in live-converge path (tryLiveConvergeTap) makes a synchronous, GET_SUGGESTED_WORDS_TIMEOUT-bounded recognizer call per tap-after-swipe. It mirrors the existing performUpdateSuggestionStripSync pattern and is gated behind a default-off setting, so it's acceptable — but if per-tap latency ever shows up, making it async is the follow-up.

[AsafMah]

CI now runs the unit suite on every PR (PR #54, :app:testOfflineRunTestsUnitTest, non-blocking until this is triaged). First real Linux-CI run shows 9 failures to resolve before flipping it to a hard gate:

Network (should skip on CI — they hit external URLs):

• XLinkTest.readmeLinks, XLinkTest.otherLinks

Known pre-existing (#12 debt, env/data-dependent):

• ParserTest.backgroundType
• InputLogicTest.tapOnlyCombiningWordDoesNotShowAutospaceIndicatorWhenGestureGateEnabled
• InputLogicTest."revert autocorrect on delete"
• StringUtilsTest.isEmojiDetectsAllAvailableEmojis

New from the v3.8.5 upstream merge (#52) — need triage (likely Robolectric screen-width config):

• MiscTest.isWideScreen_true_whenWidthGreaterThan600
• MiscTest.isWideScreen_false_whenWidthIs600
• MiscTest.isWideScreen_false_whenWidthLessThan600

Note: the existing runTests self-skip mechanism (if (BuildConfig.BUILD_TYPE == "runTests") return) doesn't cover these. To make CI a hard gate: extend that skip to the network tests + quarantine the genuinely-known-failing ones here, and triage the new MiscTest trio. (4 more ParserTest failures are Windows-only and pass on Linux CI.)

[AsafMah]

CI now runs the unit suite as a hard gate (#57, merged, green on dev/main).