Add a local replay harness for the Emulator Preview multi-run job

[jpcottin]

What

scripts/replay-preview-multirun.sh replays the Emulator Preview multi-run CI job locally. Same boot/snapshot cycles, a few minutes instead of a ~25 minute push round trip:

scripts/replay-preview-multirun.sh            # 4 cycles, throwaway AVD, cleaned up after
scripts/replay-preview-multirun.sh -n 2 -k    # 2 cycles, keep the AVD
scripts/replay-preview-multirun.sh -h         # options

It reports the app pid per cycle — identical from cycle 2 on means the app survived every snapshot restore, which is the property the experiment exists to check.

Written for a dev machine, not a runner

CI can assume it owns the host. A laptop cannot, so:

• adb calls are pinned to the emulator the script launched, identified as the device that appears after launch rather than assuming emulator-5554. Another emulator may already hold the default ports, and adb may still list a stale entry for one that has exited. If ours never appears and the log shows the console port was taken, it reports that instead of driving someone else's device.
• Shutdown is scoped to the emulator's own process group.
• A pre-existing AVD is reused and never deleted — only one the script created is cleaned up, and -k keeps that too.
• An existing console auth token is left untouched.

README

Updated the CI/CD section: it still described three jobs (there are six), and the smoke matrix had grown to API 37.0 + 37.1 across two GPU backends and both channels. Added the three experiment jobs, the shared composite action, and the local-replay instructions.

Verification

• Full 4-cycle run end to end: app pid identical across cycles 2–4, quickboot snapshot restored each time, qemu-img snapshot table populated.
• Input guards rejected -a "", -a ../evil, -n abc, -n 0 without touching any AVD.
• Pre-existing-AVD protection tested against a sentinel AVD in a sandbox HOME: reused, reported, and left intact.
• Serial selection unit-tested across four cases (foreign emulator present, stale entry only, clean machine, two foreign emulators). The previous tail -1 approach hijacked a foreign device in the stale-entry case; the new logic errors cleanly.

The last one was not hypothetical — a concurrent emulator on 5554 during testing is exactly how the bug surfaced.

Not included

avd pause before the exit screenshot, which would have removed the few seconds of drift between the screenshot and the snapshot: the console advertises it under help avd but the current preview build answers KO: not implemented, and the guest keeps running through it. Worth revisiting when it lands.