perf(port): replace systeminformation with native TCP bind probe

[ubadineke]

Summary

isPortOpen called si.networkConnections(), which shells out to lsof/netstat and scans every connection on the machine just to check one port's state. Replaced with a native bind probe (net.createServer().listen()) — same "check state, don't connect" safety property, roughly 1/700th–1/4000th the cost, and drops the systeminformation dependency, which existed for this one call.

This isn't just a during-a-session optimization — StateMonitor polls it every 1-2s the whole time the Gimlet pane is visible, not only while actively debugging.

Why not net.connect()?

A connect-based probe was actually the original implementation here, and it was deliberately removed (7fea373) — but for a narrower reason than "avoid net": it was being used to monitor whether an already-connected debugger session was still alive, which a fresh connect can't do. anza-xyz/sbpf's gdbstub accepts exactly one client and drops its listener right after, so any later connect attempt gets refused whether the real session is healthy or dead. netstat's ESTABLISHED-state check fixed that correctly.

When si.networkConnections() later replaced the plain netstat call (2488386), it was for an unrelated reason (cross-platform parsing), not because of the connect-probe bug above — but it kept the same "don't use net for port checks" habit anyway. That habit doesn't actually apply to the check this PR touches: isPortOpen only checks LISTEN state before a session starts, and a bind probe is safe for it the same way netstat/si are — it never calls .accept() on the target, just without the subprocess. Verified against a fixture mimicking the single-accept gdbstub: 0/10 probes reached its accept handler.

Benchmarks

30 iterations per case, against the actual isPortOpen in this PR vs. si.networkConnections() reconstructed from the removed implementation:

Scenario	This PR (net bind-probe)	Before (si.networkConnections)
Port closed	0.19ms mean (p95: 1.32ms)	86.58ms mean (p95: 95.02ms)
Port open	0.02ms mean (p95: 0.05ms)	89.21ms mean (p95: 99.51ms)
30 sequential ticks (StateMonitor-realistic)	0.6ms total	2562.7ms total — ~4267x

[failfmi]

This method is used to wait for a fresh gdbstub listener while the current LLDB connection may still be established on the port. A bind probe can return EADDRINUSE for the established connection even when no listener exists, which will resolve the outer loop to start another debug session immediately.

[ubadineke]

Didn't consider that, I was mainly focused on the optimization aspect, didn't take note that the same check served for fresh listeners.

Thanks, I guess you can close now.