The trigger condition that decides whether to fire the data-plane
recovery flow in startPingCheck was AND-ed with `currentInterval <
maxInterval`. That clause was meant to throttle the *backoff ramp*
(don't widen the interval past 6s), but it also gated the recovery
trigger itself — a conflation that became invisibly load-bearing
once commit 8161fa6 (March 2026) bumped the default pingInterval
from 3s to 15s while leaving maxInterval at 6s. Under the new
defaults `currentInterval` starts at 15s and `15 < 6` is permanently
false, so the recovery branch never executed. Pings just kept
failing and the failure counter climbed forever, with no
"Connection to server lost" log line and no newt/ping/request
emitted on the websocket. Real-world recovery only happened when
the underlying network came back fast enough that a periodic ping
naturally succeeded again — which doesn't happen if the WireGuard
state on either end has rotated, so users were left stuck until
they restarted newt.
This is the proximate cause of the user reports in
fosrl/newt#284 (and dups #310, fosrl/pangolin#1004). Logs in
those issues all show ping-failure counters growing without ever
emitting "Connection to server lost", which is exactly the
fingerprint of this gate being false.
The fix is to extract the trigger decision into shouldFireRecovery
and remove currentInterval from it. Backoff is now computed in a
separate `if` in the caller, still gated by `currentInterval <
maxInterval` so the ramp is a no-op under default settings (which
is the existing behaviour, just no longer entangled with the
recovery trigger). Fixing the backoff ramp itself — making it
useful when pingInterval >= maxInterval — is a follow-up: the
priority is restoring recovery, not improving the dampening
schedule.
The new shouldFireRecovery helper is unit-tested. Its signature
intentionally omits currentInterval, so a future refactor that
re-introduces the interval-dependent gate would need to change
the function signature, which makes the historical bug harder
to reintroduce silently.
Preserve optional ResponseWriter interfaces through statusCapture so httputil.ReverseProxy can hijack upgraded websocket connections. Add a regression test covering websocket traffic through the HTTP handler path.
Owen