Symptom

A client launched with no stored servers sits on the ClientScreen showing “No Nodes Yet” and the avatar bubble “Searching for Krill servers…”. A server that comes online after the client never gets discovered — node download and SSE connection never start. Restarting the app makes the same server appear immediately.

Root cause

Peer discovery was one-shot. Each side (ClientBeaconSupervisor, ServerBeaconSupervisor) fired a single startup beacon and then only listened. An idle client could only learn about a later-starting server from that server’s single best-effort UDP multicast startup beacon; one dropped packet (or IGMP-membership / interface timing on the just-joined listener) meant permanent silence, because neither side ever re-announced. The connection chain (ClientBeaconWireHandler → ClientServerConnector.connectWire → cert → node download → SSE) was intact — it simply never got triggered. Restart “fixed” it only because it made the client emit a fresh port=0 probe that the now-listening server replied to (ServerBeaconWireHandler answers any new peer’s probe). The BeaconSupervisor / Multicast KDocs already claimed beacons “fire periodically” — the implementation had drifted to one-shot.

Fix

ClientBeaconSupervisor now runs a reprobeJob alongside the listener: while AppLifecycle.isForeground is true it re-broadcasts the presence probe every REPROBE_INTERVAL_MS (5s) via the existing BeaconSender (itself rate-limited, so the cadence is an upper bound). Because the server replies to a new peer’s probe, a server coming online after the client is now discovered within one interval — no restart. Backgrounded apps skip the re-probe to respect battery.

Prevention

Discovery over best-effort UDP multicast must never depend on a single packet arriving — periodic re-announce/re-probe is mandatory (this is why mDNS/SSDP poll). The regression test (ClientBeaconSupervisorReprobeTest) advances virtual time and asserts the supervisor keeps emitting beacons after the startup one, and that backgrounding pauses them. When an interface KDoc says a behavior is “periodic,” a test should pin that contract so an implementation can’t silently regress it to one-shot.