# R-A07 Record Row Flake Investigation ## Status Active flaky-test investigation. Do not apply or merge the suspected fix until a diagnostic run captures the failure and confirms its cause. The suspected fix is currently stored in: ```text stash@{0}: On master: wip-rA07-record-row-wait ``` The working tree contains diagnostic-only changes in `AttachmentRealOsRegressionTest.kt`. The original immediate record-row click behavior remains active. ## Failure Summary The scheduled regression run failed in: ```text AttachmentRealOsRegressionTest rA07_cancelledCameraCaptureShowsErrorLeavesNoPartialAndRetrySucceeds ``` Failed GitHub Actions run: ```text https://github.com/yorvana/android/actions/runs/27553797634 commit: 3443923e1a2a6ace919b44af3a3f749e317e6bac ``` The failure occurred in `openRecord` when the test immediately clicked the newly saved record: ```kotlin composeRule.onNodeWithTag(TestTags.recordItem(record.id)).performClick() ``` Compose reported that no node with the expected `RecordItem_` tag existed. The failure was at `AttachmentRealOsRegressionTest.kt:345` in the original CI source. The CI result also contained a suppressed retry failure: ```text IllegalStateException: Only a single call to runTest can be performed during one test ``` That indicates `RetryRule` cannot safely retry this Compose test, but it does not explain why the record row was initially absent. ## Current Hypotheses The missing record-row node may be caused by: 1. A Compose/UI rendering race after save and navigation. 2. Save navigation not completing before `openRecord`. 3. Record persistence or repository visibility lag. 4. State interference from another test in the full suite. The proposed stashed fix waits for the target record-row tag before clicking. That may be correct, but applying it before capturing the failure would hide other possible causes. ## Diagnostics Diagnostic logging is implemented in: ```text app/src/androidTest/java/com/yorvana/regression/attachments/AttachmentRealOsRegressionTest.kt ``` Logcat tag: ```text AttachmentRegression ``` The instrumentation logs: - Before and after the save click. - After `composeRule.waitForIdle()`. - Persisted record IDs. - Whether the repository can read the record. - Immediately before and after the original record-row click. - On click failure only: target-node count and visible navigation markers. Semantics queries are intentionally performed only after click failure. Querying semantics before the click added enough delay to risk masking the timing race. ## Reproduction Commands Run managed-device tests sequentially. ANGLE works locally and is slower than host rendering, making it the preferred reproduction renderer. Targeted forced run: ```bash ./gradlew :app:pixel2api33DebugAndroidTest --rerun-tasks \ -Pandroid.testoptions.manageddevices.emulator.gpu=angle \ '-Pandroid.testInstrumentationRunnerArguments.class=com.yorvana.regression.attachments.AttachmentRealOsRegressionTest#rA07_cancelledCameraCaptureShowsErrorLeavesNoPartialAndRetrySucceeds' \ -Pandroid.testInstrumentationRunnerArguments.annotation=com.yorvana.testsupport.tiers.RegressionFull ``` Full extended suite: ```bash ./gradlew regressionFullExtendedCheck --rerun-tasks \ -Pandroid.testoptions.manageddevices.emulator.gpu=angle ``` Do not count an `UP-TO-DATE` invocation as a reproduction attempt. ## Log Locations R-A07 device log: ```text app/build/outputs/androidTest-results/managedDevice/debug/pixel2api33/logcat-com.yorvana.regression.attachments.AttachmentRealOsRegressionTest-rA07_cancelledCameraCaptureShowsErrorLeavesNoPartialAndRetrySucceeds.txt ``` JUnit XML: ```text app/build/outputs/androidTest-results/managedDevice/debug/pixel2api33/TEST-pixel2api33-_app-.xml ``` Extract the diagnostic timeline: ```bash rg -n "AttachmentRegression" \ app/build/outputs/androidTest-results/managedDevice/debug/pixel2api33/logcat-com.yorvana.regression.attachments.AttachmentRealOsRegressionTest-rA07_cancelledCameraCaptureShowsErrorLeavesNoPartialAndRetrySucceeds.txt ``` Search for captured failures: ```bash rg -n "click failed|AssertionError|No node|Only a single call" \ app/build/outputs/androidTest-results/managedDevice/debug/pixel2api33/ ``` ## Evidence To Capture When the failure reproduces, preserve the complete managed-device result directory before another run overwrites it: ```bash cp -a app/build/outputs/androidTest-results/managedDevice/debug/pixel2api33 \ /tmp/ra07-failure-pixel2api33 ``` The decisive diagnostic line should resemble: ```text openRecord click failed recordId= targetNodes= saveNodes= addRecordNodes= detailNodes= ``` Interpretation: | Evidence | Likely cause | | --- | --- | | Persisted ID exists, repository record is present, `addRecordNodes=1`, `targetNodes=0` | UI record-list rendering race; waiting for the row is justified. | | Persisted ID exists and repository record is present, but `saveNodes=1` | Save/navigation did not complete; wait for vehicle-detail navigation instead. | | Persisted ID absent | Persistence/save failure; investigate record save flow. | | Repository record absent but persisted ID exists | Repository visibility/cache problem. | | `detailNodes=1` before click | Test is already on an unexpected detail screen. | Also preserve the surrounding logcat lines, especially frame-skipping, Choreographer, OpenGLRenderer, and lifecycle events. ## Successful Diagnostic Baseline Two full extended ANGLE suites passed after adding diagnostics: ```text 18 executed, 3 skipped, 0 failed ``` The latest successful R-A07 timeline was: ```text rA07 after save click rA07 after waitForIdle +622 ms rA07 persisted record ids=[] +35 ms rA07 repository record present=true +2 ms openRecord before click +1 ms openRecord after click +17 ms openRecord detail visible +329 ms ``` Earlier local attempts also passed: - Three original-code targeted runs using host GPU rendering. - One original-code full extended suite using host GPU rendering. - Three original-code targeted runs using ANGLE. The issue has not reproduced locally yet. ## Next-Agent Checklist 1. Confirm `stash@{0}` still contains `wip-rA07-record-row-wait`. 2. Confirm the diagnostic-only changes remain active. 3. Run the full extended ANGLE suite sequentially with `--rerun-tasks`. 4. On failure, immediately copy the result directory to `/tmp`. 5. Compare persistence, repository visibility, navigation markers, and target node count using the table above. 6. Apply the stashed wait-for-row fix only if the captured evidence shows a healthy persisted record on vehicle detail with `targetNodes=0`.