[oluciano]

Summary

ServerHeartbeatService created its heartbeat timer only after a successful RegisterServerAsync. A storage error at start left the node out of the dashboard and the cluster tracking for the life of the process (the jobs still ran). The timer now always runs: until the node is registered each tick retries the registration (every ServerHeartbeatInterval, default 15 s) and no heartbeat is sent. A retry that completes after the host stopped deregisters the node again, so it cannot leave a ghost node. No storage contract change.

Type of change

• Bug fix

Red-first proof

• Red commit 14de1af: three tests fail with System.TimeoutException against the unfixed service (the node must register once the storage answers, a node that is not registered keeps trying, heartbeats must start once the node is registered). The 'no heartbeat before registration' half of the third is a guard.
• The review of the retry found a race: a retry in flight at shutdown could register after StopAsync deregistered. Its own red commit 9492808: StopAsync_WhileARegistrationRetryIsInFlight_DeregistersAgainWhenItCompletes times out waiting for a second DeregisterServerAsync.

Verification Checklist

• dotnet format --verify-no-changes passed
• dotnet build -c Release passed, 0 warnings
• NexJob.Tests 1215 passed on net8.0 and net10.0; core integration subset 157 passed on both; ServerHeartbeatServiceTests run repeatedly without flakes
• 3N: N1 registers when the storage returns, N2 keeps retrying without throwing and stops at shutdown, N3 no heartbeat before registration and no ghost after shutdown
• No assertion removed or weakened; HeartbeatAsync_WhenStorageThrows_SurvivesAndLogs still passes because HeartbeatAsync is unchanged (the registered-or-not decision lives in the timer callback)
• No storage interface or public API changed; protected core file changed (src/NexJob/Internal/ServerHeartbeatService.cs) by the core owner
• DocsTruth --strict: 0 drift; delivery-guarantees and troubleshooting list the new warning
• CHANGELOG.md updated under [Unreleased] / Fixed

Notes

• RegisterServerAsync is an upsert on every provider (SQL Server MERGE, PostgreSQL ON CONFLICT, MongoDB upsert, Redis hash set, InMemory overwrite), so overlapping retry ticks are harmless; the worst case is a duplicate Information log.
• StartedAt is now fixed at construction, so retries send the same value; the difference from before is milliseconds.
• Patch scope: no new option, no contract change.

Closes #646

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud