P6.3 Android phone-use runtime lane

[Harzva]

Summary

• Adds MobileCode Phone Use Accessibility service wiring and an in-app Tools card for dry/action probes.
• Adds P5.7/P5.8/P6.3 non-counted phone-use runner scripts, including a robust P6.3 adb runtime verifier for install, launch, Accessibility state, Tools navigation, dry/action probe, Back/Home focus state, screenshot/UI XML/logcat capture, and validator-compatible run output.
• Adds P6.3 Android emulator evidence under docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/ and updates roadmaps/contracts with the new evidence boundary.

Non-counted Boundary

This is Android emulator phone-use runtime QA only. It is not a formal benchmark and must not be reported as a counted strategy ablation result.

• run_kind=strategy_pilot_not_counted
• counts_as_experiment=false
• counts_as_strategy_ablation_result=false
• score_boundary=pilot_android_phone_use_runtime_score_not_counted
• no model callback, provider claim, raw transcript, secret material, or private local path evidence is included
• the same runtime score is mirrored across strategies, so this does not prove strategy quality differences

Evidence

• Run package: docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/
• Verifier JSON: docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/phone_use_runtime_verifier.json
• Run JSON: docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/run.json
• Summary: docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/summary.md
• Key screenshots: 01-launch.png, 02-tools-phone-use.png, 03-dry-probe.png, 04-action-probe.png, 05-home-after-action.png
• Result: status=passed, runtime_score=100.0, action_acceptance=4/4

Verification

• python3 -m py_compile scripts/run_mobile_harness_strategy_real_pilot.py scripts/run_mobile_harness_strategy_p57_phone_use_eval.py scripts/run_mobile_harness_strategy_p58_android_phone_use_runtime.py scripts/run_mobile_harness_strategy_p63_android_real_device_lane.py scripts/run_mobile_harness_strategy_p60_p62_runtime_benchmark_upgrade.py scripts/validate_mobile_harness_strategy_ablation.py
• python3 scripts/validate_mobile_harness_strategy_ablation.py --registry docs/mobile-harness-benchmark/strategy-ablation/strategy_registry.json --run docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/run.json
  • PASS, 6 strategies, 6 results, strategy_pilot_not_counted
• flutter analyze lib/services/phone_use_accessibility_service.dart lib/widgets/phone_use_mode_card.dart
  • No issues found
• flutter test test/services/phone_use_accessibility_service_test.dart test/widgets/phone_use_mode_card_test.dart
  • 6 tests passed
• flutter build apk --debug
  • built build/app/outputs/flutter-apk/app-debug.apk
• python3 scripts/run_mobile_harness_strategy_p63_android_real_device_lane.py --apk mobile_agent/build/app/outputs/flutter-apk/app-debug.apk --device emulator-5554
  • status=passed score=100.0
• git diff origin/main-dev..HEAD --check
• staged secret/private path scan: no secret findings; only generic redaction regex mentions /Volumes and /Users
• boundary assertions scan: no true counted flags and no non-empty transcript_paths

Review Focus

• mobile_agent/android/app/src/main/kotlin/com/mobilecode/app/PhoneUseAccessibilityService.kt
• mobile_agent/lib/services/phone_use_accessibility_service.dart
• mobile_agent/lib/widgets/phone_use_mode_card.dart
• scripts/run_mobile_harness_strategy_p63_android_real_device_lane.py
• docs/mobile-harness-benchmark/strategy-ablation/runs/p63-android-real-device-lane/
• roadmap updates for P6.3 scope and P6.4 WebView/localStorage follow-up

Scope Notes

• This evidence was collected on an Android emulator lane, not external Android hardware.
• WebView/localStorage/generated artifact interaction assertions are intentionally left for P6.4/P6 counted promotion gates.
• Merge should wait for PR checks, then the next lane is external Android real-device Accessibility authorization and repeated runtime samples.