Fix Shizuku double-bind cascade that orphaned virtual displays

[Suprhimp]

Summary

• VirtualDisplayManager and ShizukuSetup were both calling Shizuku.bindUserService. VDM also treated every onServiceConnected after the first as a binder death and recreated the VD — so duplicate connects spawned 6+ orphan VirtualDisplays per session and zombie :privileged processes.
• Result: apps launched from the web app-list targeted a different display ID than the one being encoded, the browser canvas stayed empty, and the JS-side 5s timeout fired "Launch timed out, try again" while the activity appeared on the phone.

Changes

• VirtualDisplayManager: strip the embedded Shizuku.bindUserService ServiceConnection. New attachPrivilegedService(svc: IPrivilegedService?) API receives the binder from outside. release() is now local-state only.
• MirrorForegroundService:
  • ensureShizukuSetup() — lazy singleton creation at first browser connect; ShizukuSetup now lives for the foreground-service lifetime.
  • startReconnectObserver(setup) — singleton coroutine (reconnectJob) collects setup.serviceConnected and classifies transitions via BinderConnectionTracker. Only Reconnect recreates a VD; FirstConnect is owned by trySetupVirtualDisplay.
  • trySetupVirtualDisplay rewritten: waits for serviceConnected.first { it } with timeout, reuses long-lived setup, gated by shizukuSetupInProgress against concurrent rebuild races. Single-delivery semantics via safeResult.
  • releaseShizukuSession → tearDownVdSession — no longer releases ShizukuSetup mid-session. performCleanup cancels reconnectJob before shizukuSetup.release().
  • displayId < 0 guards on launchExternalBrowserTarget and launchSplitExternalBrowserTarget.
• ShizukuSetup: removed dead attachPrivilegedService(svc) API (had no callers after the refactor).
• BinderConnectionTracker (new, pure logic, 7 unit tests): state machine that distinguishes a fresh connect from a reconnect after a real disconnect. Key correctness rule: onDisconnected from New stays in New (so MutableStateFlow's initial false emission to a fresh collector doesn't misclassify the next real connect as Reconnect).

On-device verification (Samsung Z Flip 6, open state)

Before:

13:17:50.967 VD_CREATED id=26
13:17:51.051 VD_CREATED id=27
13:17:51.150 VD_CREATED id=28
13:17:51.239 VD_CREATED id=29
13:17:51.488 VD_CREATED id=30
13:17:51.981 VD_CREATED id=31

(6 VDs in 1s; 14+ zombie :privileged processes accumulated across sessions)

After:

10:09:46.694 Shizuku connection transition=FirstConnect connected=true
10:09:46.939 VD_CREATED id=64 864x720
10:09:50.458 Launched com.nhn.android.nmap/com.naver.map.LaunchActivity on virtual display 64

(1 VD per session; canvas renders the launched app in browser as expected.)

Out of scope

When Z Flip is closed, Samsung's FlipLargeCoverScreenSizeCompatMode intercepts am start --display <vd> for non-whitelisted apps (e.g. Naver Map) and reroutes them to the cover screen, bypassing the virtual display. This is per-app OS policy — Samsung Calculator launches on the Castla VD under the same code path. Not addressed in this patch; tracked separately.

Test plan

• ./gradlew :app:testDebugUnitTest --tests "com.castla.mirror.shizuku.*" — green
• On-device verification on Samsung Z Flip 6 (flip open): 1 VD per session, transition=FirstConnect, single Shizuku bind, app launches render in browser canvas
• Verify on additional non-foldable devices to ensure no regression in the open-flip / non-foldable launch path

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
website	Ready	Preview, Comment	May 16, 2026 0:38am