A3d: Native gesture-recognition replay harness + run native tests in CI

[AsafMah]

Part of #13. The gap #22 does not cover.

Problem

All automated tests are JVM/Robolectric and tap-based. The real swipe→word recognition runs in the native C++ engine (app/src/main/jni/), which JVM/Robolectric tests cannot exercise. So gesture recognition has zero automated coverage — only on-device manual testing catches a recognition regression (e.g. a scoring/geometry change that makes swiping produce the wrong word). This is the long-standing coverage gap noted in AGENTS.md / IMPROVEMENT_PLAN.md.

What already exists (so this is tractable)

• A native C++ gtest harness is already in the repo, just not run in CI: app/src/main/jni/tests/** (dictionary / suggest-core / dicnode / geometry tests), run-tests.sh, HostUnitTests.mk (host build — runs in CI with no device/emulator), TargetUnitTests.mk. None of the existing tests cover gesture recognition.
• TraceRecorder (A3a: Pointer-trace recorder & capture format #20) already captures replay-ready golden fixtures. Per trace: pointers = flattened InputPointers (xCoordinates / yCoordinates / pointerIds / times), keyboard = {width=mOccupiedWidth, height=mOccupiedHeight, layout, locale}, committedWord = the recognized word. i.e. the swipe + its expected output.

Deliverables

• Run the existing native host unit suite in CI. Wire run-tests.sh / HostUnitTests.mk (host build) into a CI job. Instant coverage of the dictionary/suggest/geometry engine the JVM can't reach. (Low effort, high value — the tests already exist and pass.)
• Native gesture-replay test. A new C++ test (under app/src/main/jni/tests/suggest/) that: loads recorded trace fixtures (a small committed corpus of TraceRecorder JSON), reconstructs the proximity/keyboard geometry, feeds the InputPointers through the gesture/glide scorer, and asserts the top suggestion == recorded committedWord (or within top-N). Deterministic, no device.
• Commit a seed fixture corpus — a handful of recorded traces for common words across the default layout (recorded via the A3a: Pointer-trace recorder & capture format #20 recorder on-device, sanitized), as test resources.

Acceptance

• CI runs the native host test suite and fails on a real native regression.
• The gesture-replay test catches a deliberate recognizer break (sanity: perturb scoring → test goes red).
• Adding a new golden trace is a drop-in (copy a recorder JSON into the fixtures dir).

Notes / open questions

• Confirm the gesture scorer is reachable from a host build (some native paths are target/NDK-only). If the host build can't link the gesture path, fall back to a TargetUnitTests.mk (emulator) job for the replay test only, keeping the rest host-side.
• Keep the fixture corpus small + curated; this is a regression net, not a benchmark.

[AsafMah]

Progress update: #96 merged.

What #96 adds:

• TraceRecorder-style JSON fixture parser for native host tests.
• Seed hello_qwerty.json fixture.
• 9 enabled parser/fixture tests in the native host suite.
• 1 disabled recognizer replay scaffold (DISABLED_GestureReplayTest.ReplayHelloQwerty).

Why the real recognizer assertion remains disabled (no fake green):

1. ProximityInfo currently requires a live JNIEnv* because its constructor calls env->GetArrayLength(...) unconditionally. Need a non-JNI raw-array constructor or a tiny host JNIEnv shim.
2. Dictionary needs a compiled binary .dict asset; source tree does not ship one. Need to bundle/fetch a small compiled en-US dict or generate a tiny v4 dict in-test.

So #78 is partially done: the fixture/replay path exists and is CI-covered for parsing; the next concrete engineering step is the host-testable ProximityInfo seam, then a test dictionary asset, then enabling the replay assertion.

[AsafMah]

Proof spike result (#104):

• ✅ Removed the JNI blocker for keyboard geometry: ProximityInfo now has a host-test constructor that accepts raw geometry arrays.
• ✅ Added GestureReplayHostSeamTest.BuildsProximityInfoWithoutJNI, proving the hello trace can be paired with host-built QWERTY ProximityInfo.
• ✅ Native host suite: 77/77 enabled tests pass (plus the known FormatUtils quarantine excluded), replay assertion remains disabled.
• ❌ Real gesture replay still cannot run in this open-source tree because there is no gesture suggest policy implementation here — only GestureSuggestPolicyFactory, a nullable hook. In host builds the factory returns null, and Dictionary::getSuggestions(..., IS_GESTURE) crashes at TRAVERSAL->getMaxSpatialDistance().

So the remaining blocker is not .dict assets; I tested an in-memory dict and ASAN showed the crash occurs before dictionary contents matter. To enable real replay assertions, we need an open/host-buildable GestureSuggestPolicy implementation (e.g. future NLnet recognizer) or a test policy with the same interface.

[AsafMah]

Closing as far as we can take it without an open gesture recognizer implementation.

Delivered:

• test(native): host CMake build + CI for native C++ engine tests (#78) #81 — native C++ host tests in CI (dictionary/suggest/geometry suite), already blocking native changes.
• test(native): gesture replay fixture loader + disabled recognizer scaffold (#78) #96 — TraceRecorder-style fixture parser + seed hello_qwerty.json + disabled replay scaffold.
• test(native): host-constructible ProximityInfo for replay harness (#78) #104 — host-testable ProximityInfo constructor + proof test, removing the JNI geometry blocker.

Proof result:

• fixture → native arrays path works;
• host ProximityInfo construction works;
• enabled native suite passes;
• actual Dictionary::getSuggestions(..., IS_GESTURE) replay cannot run because this open-source tree has no GestureSuggestPolicy implementation — only the nullable GestureSuggestPolicyFactory hook. In host builds it returns null and the recognizer path crashes before dictionary contents matter.

So a real replay assertion is blocked on an open/host-buildable gesture policy (e.g. future NLnet recognizer) or a deliberate test policy double. Track that separately when the recognizer exists; do not keep this issue open as an implementation task we can finish today.