perf(sidecar): net.poll clamp 3ms->1ms — non-TCB mitigation for multi-guest X-latency

Shorter round-robin quanta so a polling guest holds the single per-VM sync-RPC service thread for less time, reducing starvation of sibling guests (X server + WM + app) under contention. The M8.0 note that 1ms regressed the xinitthreads-probe predates the libxcb timed-cond fix; the probe now passes cleanly at 1ms (verified). Partial non-TCB mitigation only — the clean fix (release the per-VM lock during net.poll's pure wait) alters the security-critical serialization lock and is reserved for the human TCB sign-off gate (WASM-THREADS-SPEC §10).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: NathanFlurry

crates/sidecar/src/execution.rs

@@ -19438,7 +19438,7 @@ where
 // serviced, so a chatty client (e.g. an Xt app's main loop polling its X connection) would starve the
 // window manager and other clients sharing the VM. A few ms keeps each guest's poll responsive while
 // letting the service thread round-robin quickly across all guests. Shutdown latency stays bounded.
-const JAVASCRIPT_NET_POLL_MAX_WAIT: Duration = Duration::from_millis(3);
+const JAVASCRIPT_NET_POLL_MAX_WAIT: Duration = Duration::from_millis(1);
 const EXITED_PROCESS_SNAPSHOT_RETENTION: Duration = Duration::from_secs(2);
 
 fn resolve_http2_file_response_guest_path(process: &ActiveProcess, path: &str) -> String {