Prevent recursive dex2oat execution after daemon restart

[XR-stb]

When the daemon starts in a mount namespace containing an earlier dex2oat wrapper bind mount, Dex2OatServer can cache the wrapper itself as the stock compiler. The wrapper then execs that FD through the linker, re-enters main in the same process, and repeatedly acquires descriptors and copies the hook library.

This change:

• Validates the opened compiler descriptor against the corresponding wrapper inode/device before caching it.
• On compiler initialization failure, invokes the existing wrapper-disable/fallback path, closes cached descriptors, reports DEX2OAT_MOUNT_FAILED, and does not start the socket server or SELinux observer.
• Adds a PID-scoped environment guard before any socket requests so recursive exec terminates even if the daemon-side identity check misses a replaced wrapper inode. Child processes inheriting the environment remain allowed.
• Adds a Linux host integration test using the actual wrapper source and a mock SCM_RIGHTS server, and runs it in CI.

Validation:

• All 3 host tests pass: normal compiler execution, wrapper-as-stock recursion rejection, and inherited parent marker.
• The same recursion test against the unmodified source times out after 5 seconds.
• Wrapper cross-compilation passes for Android ARM64 and ARMv7/API 27 using NDK 26.1.
• Dex2OatServer compiles against Android API 35 with local project/hidden-API stubs (a limited syntax/type check, not the full project build).
• Full Gradle build, regression tests, and Release/Debug artifact generation passed in fork Actions run 38583368959. The build-only branch at 7a01e11 has the same application/test sources as this PR head (6b8a273); only release-publishing workflow steps were removed. The upstream PR workflow still awaits maintainer approval.

Device validation (2026-10-14):

• Xiaomi 14, Android 16 / HyperOS 3.0.305.0, KernelSU temporary-root mode (32680), Zygisk Next 1.5.2.
• Installed Release build 3112 from the run above and activated it with one KernelSU soft reboot. Root remained available; the module reported enabled with no pending update. The fork labels it v1.0 because release tags were absent, not because it uses the historical v1.0 source.
• Exactly one vectord process was observed. Cached stock compiler descriptors had device/inode pairs 65067:20 and 65067:21, distinct from wrapper files 65078:6021722 and 65078:6067890. Framework injection into WeChat and HyperCeiler module loading were logged.
• A 60-second background sample showed 90.5–94.8% aggregate CPU idle. No repeat of the previously observed sustained compiler high-CPU loop was observed during these samples.
• During 73.435 seconds of user-driven WeChat activity, gfxinfo counters from the same PID and unchanged stats-start time increased from 1,667 to 6,884 frames, and from 8 to 91 janky frames: 5,217 new frames, 83 janky (1.59%). The separately reported legacy jank counter increased by 821 (15.74%); these are different metrics. Aggregate CPU idle in the concurrent samples ranged from 46.1% to 90.5%.
• The device owner reported that WeChat felt much smoother and Honor of Kings no longer felt laggy. No game frame-time trace was collected.

Interpretation and limits:

• This is an installation/runtime smoke test and a short workload observation, not a controlled before/after performance benchmark. Stale wrapper mounts had already been removed during recovery before installation, and the tested build also moves from the older 88f8e1f Debug build to current-source Release. The observed improvement cannot be attributed solely to this patch.
• The stale-mount fault was not deliberately recreated on the phone. The controlled recursion rejection evidence comes from the host regression test; the device checks confirm normal startup with correct stock compiler descriptors. Repeated soft-reboot lifecycle behavior and long-term stability remain untested.
• Compatibility warnings remain in the device log (including an unavailable systemui_is_cached field); this does not establish compatibility of every module or feature.

Motivation: a Xiaomi 14 / Android 16 / KernelSU temporary-root setup running Vector 2.2 (88f8e1f) exhibited repeated wrapper entry in one PID, with stock descriptors increasing beyond 10,000 and the next DEX2OAT_CMD referring to the previous stock FD. This is consistent with recursive exec; the exact event that left the stale mount was not captured.

Related to the daemon lifecycle concerns in #878, but this does not fix duplicate daemon ownership or claim to resolve that entire issue. The existing unmount/fallback operation is best-effort; automatic stock-FD recovery and complete soft-reboot lifecycle handling remain outside this change.