[easyhooon]

Summary

Fixes the reconnect after a cable replug reported in #31: the Android app stayed out of Necto until it was restarted.

Cause. NectoSdkRuntime.serve handled one session to the end before it called accept() again. After an unplug the phone kept the old connection established, so the runtime stayed blocked reading it, and the Mac's probes after the replug piled up in the backlog. handle() already closed a previous session when a new one arrived, but a second one never got that far.

Fix.

• Each accepted connection is handled in its own coroutine, so accepting never waits on a session.
• A connection replaces the current session only once its handshake is accepted. Until then the working session, its requests and its status are left alone, so a connection that never answers, or is refused, cannot end it.
• stop() also closes connections that are still handshaking, since their reads block and are not woken by cancellation.

Tests

• aNewerHostReplacesOneThatWentQuiet: a connected host goes silent, a second host handshakes, the first is closed and the second answers test.echo. On the old runtime this test hangs.
• aHostStillSayingHelloLeavesTheConnectedOneAlone: a second connection that has only read the hello, then is refused, leaves the first answering and the status Connected.
• SDK/Android/gradlew -p SDK/Android test passes.

Device check

Galaxy S25+ over USB, script/android-sample R3CY70P3F4H, then unplug and replug without touching the app:

Time	Necto lists the app	ADB sees the phone
23:47:08	yes	yes
23:47:24 (unplugged)	no	no
23:47:32 (plugged in)	no	yes
23:47:34	yes	yes

After the replug, plugin send plugin-sample host.info answered, the app kept the same process, and /proc/net/unix showed only the listener and one connection.

Not in this PR

• iOS has the same loop shape in NectoSDKRuntime.swift. I have not unplugged an iPhone, so I don't know whether usbmuxd closes the socket on the phone and hides this there.
• Two Neccos on one app still don't work together. Before, the second waited in the backlog; now they would take the session from each other on every probe. This is no worse, but it isn't solved.
• Possible follow-ups: a protocol level heartbeat so either side notices a dead connection without a new one arriving, or serving several hosts at once with per session state.