perf: known_routes retention, digest identity, and a working batched reconfirm (#98)

* fix: cascade channel_messages on packet delete

* rebuild known_routes keyed on (iata, path_key)

* add route retention delete, batch reconfirm, path_key upsert

* compute route path_key in store, add DeleteOldRoutes

* add routes retention config

* wire route retention into cleanup, batch reconfirm

* prune routes in reconfirm task, not cleanup

* expand routes retention docs in example config

config.yaml.example

@@ -75,6 +75,25 @@ telemetry:
 packets:
   retention: 720h   # 30 days
 
+# Known-route retention. Routes are distilled packet history (the path a packet
+# took, hop by hop), so they may outlive packets.retention; there is no required
+# relationship between the two.
+routes:
+  # How long a route is kept after it was last observed. Every new observation
+  # of the same path resets this clock, so routes the mesh still uses never
+  # expire -- only paths nothing has traveled for this long age out.
+  retention: 336h   # 14 days
+  # Shorter window for rarely-seen routes: one observed fewer than
+  # min_observations times in total is dropped once it goes unobserved for
+  # this long. Must be shorter than retention, or it never applies.
+  grace: 168h       # 7 days
+  # Lifetime observation count a route must reach to earn the full retention
+  # window. Below it a route is treated as flood noise -- a path recorded once
+  # or twice and never confirmed -- and only kept for the grace window. The
+  # count never resets, so once a route crosses this bar it stays in the
+  # retention tier for good.
+  min_observations: 3
+
 websocket:
   max_connections_per_ip: 5 # default: 5