The 10s-after-healthy failure pattern was reproducing even with
the connection pool disabled. Smoking gun: the inbox loop opens
GET /messages?timeout=30 right after start_daemon returns, and
every subsequent peers/routes call timed out exactly when our
client-side reqwest timeout (10s) fired.
Concluded mycelium 0.6.1's HTTP server serialises requests: while
the long-poll connection is held, no other admin endpoint can
respond. The sidecar process kept logging routes the whole time
(seen in the in-app log buffer) — proof the daemon was alive,
just unable to serve concurrent calls.
Switch to short-poll: timeout=0 returns immediately, sleep 2s
between iterations. Per-iteration server hold time is now
millisecond-scale instead of 30s.
939565b88a