[IgnacioLD]

Play Console (2.5.4, Android vitals) warns that the R8 setup may use more memory because optimized resource shrinking is off, and says to move to AGP 9. Google starts enforcing these memory and code optimization thresholds in February 2027. This PR upgrades the build, removes keep rules that are no longer needed, and lowers the memory that on-device scoring holds. There are three commits, one per concern.

1. AGP 8.9.1 -> 9.3.3, Gradle 8.11.1 -> 9.7.1

• Why 9.3.3 and not 9.5.1: 9.3.3 is the newest AGP inside Kotlin 2.4.20's tested range (AGP 8.5.2-9.3.1, Gradle 7.6.3-9.7.0). 9.4.x and 9.5.x are outside that range, and 9.5.x was released less than two weeks ago. AGP 9.3 needs at least Gradle 9.5. I picked Gradle 9.7.1, the version deploy/Dockerfile already builds the server with. Both picks are patch releases of tested versions. The wrapper jar matches the official sha256.
• KMP modules: :engine and :shared now use com.android.kotlin.multiplatform.library, because AGP 9 rejects com.android.library next to the KMP plugin. These stay as they were: the opt-in Android target (-Pglosso.android), onnxMain, kotlin.mpp.applyDefaultHierarchyTemplate=false and JVM target 17.
• :androidApp: uses AGP's built-in Kotlin, so the kotlin-android plugin is gone and the JVM target follows compileOptions. resourceConfigurations is replaced by androidResources.localeFilters. Both flavors, the licensing gate per flavor and the release -P inputs are unchanged.
• Test moves:
  • shared/src/androidUnitTest is now androidHostTest, so the task is :shared:testAndroidHostTest.
  • BottomActionBarTest, GlossoArtworkTest and PlayerPointerTest stopped compiling in :androidApp once the UI moved to :shared: they use internal API, and one still imported Coil 2. That broke connectedAndroidTest and tools/score-parity/run-parity-test.sh on main. They now live in shared/src/androidDeviceTest (:shared:connectedAndroidDeviceTest).
• Gradle 9.7 deprecations fixed: delegated by getting/creating/registering in engine, shared and androidApp, and Task.project read at execution time by the licensing tasks.
• Licensing: the locks and licenses.json are regenerated with tools/licensing/collect.py. kotlin-stdlib-common:2.4.20 now resolves.
• Docs: README and CONTRIBUTING list :shared:testAndroidHostTest.

2. R8

• Optimized resource shrinking is AGP 9's default. isMinifyEnabled and isShrinkResources were already on for release.
• Rules removed from proguard-rules.pro:
  • the kotlinx.serialization block, which the library already ships;
  • -keep class androidx.credentials.playservices.** { *; }, because the AAR ships a narrower rule;
  • the Coil 2 rules, which named classes that no longer exist;
  • every blanket -dontwarn, because R8 reports no missing class without them;
  • the duplicate native-method rule;
  • -renamesourcefileattribute, so R8 9 writes its mapping id as the source file.
• Rule kept: ai.onnxruntime.**. The JNI layer looks up classes, fields and constructors by name, and the AAR ships rules only for its telemetry classes.

Release, unsigned	before (main d70a496)	after
playstore APK size	138,753,248 B	138,421,983 B
playstore referenced methods	30,511	26,178 (-14%)
playstore dex	5,119,168 B	4,621,976 B
playstore res/ files / resources.arsc	277 / 113,228 B	13 / 44,704 B
fdroid APK size	138,232,808 B	138,227,046 B
fdroid referenced methods	25,514	23,963
fdroid dex	4,463,520 B	4,291,760 B
playstore AAB	60,310,866 B	59,993,510 B

Methods were counted with apkanalyzer dex references. Most of the APK is the ONNX Runtime .so for 4 ABIs (about 135 MB), so the total size barely moves.

3. On-device scoring memory

• Model load: the model was already loaded from the file path, not through a ByteArray.
• Session options: the session is still created once and kept. It now runs with setCPUArenaAllocator(false) and session.use_device_allocator_for_initializers=1. These change only where tensors are allocated, not the math.
• Background release: OnDeviceScorer releases the model on TRIM_MEMORY_BACKGROUND and reloads it on the next take. A take holds the load lock while it runs, so the session cannot be closed under it.
• Not changed: the model stays loaded when the learner leaves practice but stays in the app, to avoid reloading it on every screen.
• Bit-exact check: a SHA-256 over the recognition (IPA, per-token confidences, timings) of score-parity/fixtures/good.wav was 1f3c5e77... before and after. tools/engine-bench/run.sh output is identical line for line between main and this branch.

Memory on emulator-5558 (arm64, 1080x2340 at 480 dpi), debug build. A scratch instrumented probe (not committed) walks through the phases while adb shell dumpsys meminfo me.shirobyte42.glosso is sampled every 0.5 s. Values are MB at the end of each phase:

Phase	PSS before	PSS after	Native heap before	Native heap after	Java heap before -> after
idle	183	187	4	6	9 -> 10
model loaded	589	592	400	400	9 -> 10
after 3 scored takes	610	574	416	378	13 -> 13
second recognizer loaded (peak)	1018	918	826	724	11 -> 10
second recognizer closed	639	588	447	394	10 -> 10
onTrimMemory(TRIM_MEMORY_BACKGROUND)	638	230	447	36	10 -> 10

Scoring latency for the fixture was unchanged within noise (321/291/268 ms before, 337/310/265 ms after).

Verification (rebased on main d70a496)

• Unit tests:
  • ./gradlew :engine:jvmTest: 463 passed
  • LC_ALL=en_US.UTF-8 ./gradlew :server:test: 347 passed
  • :androidApp:testFdroidDebugUnitTest / testPlaystoreDebugUnitTest: 296 / 295 passed
  • :shared:testAndroidHostTest: 146 passed. The same test files ran as testDebugUnitTest before the move: 142 on the old base, plus the tests that came with feat(app): let the learner point the app at a self-hosted server #15.
• Lint: :androidApp:lintFdroidDebug and lintPlaystoreDebug report 0 errors. Warnings went from 73 to 11 per flavor. The one new warning is OldTargetApi: the newer lint knows API 37, and raising targetSdk is a separate product change.
• iOS: :engine:compileKotlinIosSimulatorArm64 and :shared:compileKotlinIosSimulatorArm64 pass.
• Server image path: ./gradlew :server:installDist -Pglosso.android=false -Pglosso.ios=false passes without the Android SDK in the environment. Docker is not available on this machine, so docker build was not run. The Dockerfile already uses Gradle 9.7.1.
• Release builds: assemblePlaystoreRelease, assembleFdroidRelease and bundlePlaystoreRelease pass.
• Release on device: I used a temporary local setup, not committed: testBuildType = "release", debug signing, and keep rules for the test harness only. ScoreParityTest passed on the R8-minified playstore release on emulator-5558: good.wav scored 100 / PERFECT with delta 0, and silence, noise and low were not evaluable. ONNX Runtime loads and scores under the new rules. OfflineReferenceParityTest skipped because thirty.wav is not staged anywhere.
• Debug on device: tools/score-parity/run-parity-test.sh --serial emulator-5558 passed. :shared:connectedAndroidDeviceTest ran 6 tests and all passed.
• Engine bench: GLOSSO_MODEL=<abs>/wav2vec2_phoneme_int8_dyn_pertensor.onnx tools/engine-bench/run.sh gives native >= 98 for 30/45 and REFSCORE 22/28 (>= 98), identical on main and this branch. The bench's own 0.9 rate gate already fails on main; this PR does not change that.
• gitleaks: gitleaks dir . --redact reports only the 2 known false positives.

Notes

• AGP prints an informational warning that :engine's commonTest has no Android host test. Turning one on fails to compile, because ParityPlatform has no Android actual. The old androidUnitTest variant never compiled either.
• No NDK r28 is installed locally, so "Unable to strip libonnxruntime.so" is printed, as it was before. The libraries are already stripped.

https://claude.ai/code/session_01FcVpZoyr8VdRLp1QecX1Ng