[jeremydmiller]

Closes #5430

Two bugs left released inbox rows unrecovered under Balanced mode on SQLite. Durability chaos round three (#5414) lost about 4,700 messages per run.

1. Every node was node 1

SqliteNodePersistence.PersistAsync wrote back the node's default AssignedNodeNumber (1) instead of assigning one, as every other store does with a sequence or identity column.

It now assigns max + 1 over the node table and the node event records, in a single INSERT ... SELECT, which is atomic under SQLite's single writer. Including the records makes the number only move forward: a node ejected as stale keeps its number when it resumes and re-registers, so a newcomer must not get it.

2. Health checks were read as local time

datetime('now') is UTC text with no offset, and DateTimeOffset.Parse gave it the machine's offset. West of Greenwich, every health check looked hours in the future, with three effects:

• the leader never ejected a stale node
• the stale nodes' rows were never released
• the database agent kept being assigned to a dead node

Node and record timestamps now parse as UTC. CI runs in UTC, which hid this.

Tests

• New compliance test: NodePersistenceCompliance.persisting_nodes_assigns_each_its_own_number. It was red on SQLite, and it passes on PostgreSQL, SQL Server, MySQL, Oracle, RavenDB and Cosmos DB.
• Bug_5430_node_numbers_are_assigned_by_the_store: covers the no-reuse rule and the UTC read-back. Each test was red without its fix; the UTC test can only fail on a machine outside UTC.
• SqliteTests: 498 passed, 2 skipped.
• Chaos rig leg 2 (seed 55, SQLite queue, Balanced, 14 kills and 4 SIGTERMs) went from 4,727 lost to clean:
  • with fix 1 alone it still lost about 4,900
  • with both fixes it settled in 73 s

Balanced mode stays supported on SQLite, since refusing it would be a breaking change.

🤖 Generated with Claude Code

https://claude.ai/code/session_013nKJrqsJw3j23Xd46q37ob