daemon: address watcher follow-ups — own retry backoff, Stop not held by a redial, no duplicate heartbeat recovery

[TeoSlayer]

Three follow-ups to the address watcher from #495, found in its review and confirmed by probes on main.

(a) Registry retries no longer push back the next address move

Retries shared the local-path backoff, so retries at +10 s, +30 s and +70 s left the next real move waiting 80 s, up to the 2-minute cap. They now have their own backoff.

Test: TestAddrWatchRegistryRetriesDoNotHoldBackTheNextMove fires at [1 11 31 71 81], against [… 151] on main.

(b) Stop no longer waits on a registry redial or a heartbeat backoff

The registry dial backoff and the heartbeat's re-register sleep ignored Stop. With a refused registry, a reconnect outlived Stop by 47 s, and Stop logged "leaked". A new sleepOrStop helper returns early on stop, and every attempt checks stopping().

Test: TestForceReconnectRegistryGivesWayToStop.

(c) The heartbeat no longer repeats a recovery that just ran

• The problem: the heartbeat's serialised reconnect and re-register never skipped a duplicate and never recorded success. When its call on the old connection timed out during a slow recovery, it replaced the recovery's fresh connection and ran a second full re-registration: 2 registers and 4 report_trust.
• The fix:
  • Shared helpers reestablishedRecently and noteReestablished.
  • The reconnect skips if the connection was already replaced.
  • The re-register skips within 5 s of a full run and records its own.
  • The reconnect still has the heartbeat re-register when the recovery that replaced the connection did not register (its own re-registration failed): nothing registered after the heartbeat's call began.
  • A full run counts as full only if every visibility, hostname and trust write succeeded. Otherwise it counts as endpoint-only, so the heartbeat's full re-registration is not skipped.
  • RegisterWithBeacon in reestablishTransport now runs before the "done recently" check, so the beacon registration happens even when the registry half is skipped. An endpoint-only re-registration does not register with the beacon, and the rx watchdog's soft recovery counts on the beacon's reply as inbound traffic.
• Tests:
  • TestHeartbeatDoesNotRepeatARecovery (1 register, connection kept; both orders).
  • TestHeartbeatReregistersWhenTheRecoveryDidNot.
  • TestPartialRestoreDoesNotCountAsFull (the heartbeat re-syncs both trust pairs after a recovery whose trust writes failed).

Each test fails on main. Results:

• go test -race ./pkg/daemon passes;
• go test -race -run 'Addr|Recover' ./tests -count=2 passes.

AdvertiseEndpoint is untouched.

🤖 Generated with Claude Code