Release delayed workflows against the database's clock, and stamp updated_at

[devhawk]

Stacked on #612.

A DELAYED workflow was released when its delay_until_epoch_ms was at or before the clock of whichever JVM ran the release. That is often not the JVM that enqueued it. transitionDelayedWorkflows now compares against the database's now(), the one clock every executor sharing the system database shares, so whichever executor runs the release, a delay ends at the same instant.

• updated_at: the release also stamps it from the database's clock. The statement used to leave it unchanged.
• Not changed: the end of a relative delay is still resolved from the enqueuing JVM's clock, once, outside the write's retry loop. Moving that into the statement too would let a retried insert push a delay later by its backoff, which would have to be settled for every SDK first. So the enqueuer's clock still sets the delay; only the releasing executor's clock is gone. Go releases delays the same way.
• Javadoc: the NOW_EPOCH_MS javadoc now lists the release among the times kept on the database's clock.

Tests

DelayReleaseClockTest (new, Postgres only):

• shifts the database's clock an hour either way by shadowing now() and clock_timestamp() on the connections' search_path;
• delays a workflow by a day, then moves its delay to end a second from the database's now();
• checks that it is released a second later, with updated_at on the database's clock.

The executor in the test doesn't dequeue the workflow's queue, so the released row stays ENQUEUED and keeps the release's own updated_at. On the old statement the test fails both ways: with the database an hour ahead the workflow is never released, and with it an hour behind it is released at once.

Run locally on Postgres: spotlessCheck, compileTestJava and javadoc, plus:

• DelayReleaseClockTest, DatabaseClockTest, WorkflowTimeoutSweepTest, TimeoutTest and DebounceDelayedWorkflowTest on this branch;
• DebouncerTest, DebouncerClientTest, StartWorkflowTest, ClientTest and SystemDatabaseTest before it was stacked.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

The focused implementation is consistent with existing database-time handling and is adequately tested.

0 open findings

What changed in this PR

Uses the shared database clock to release delayed workflows consistently across executors and stamp release time.

Changes:

• Compares delayed-workflow deadlines against database now().
• Updates updated_at when releasing workflows.
• Adds PostgreSQL clock-skew coverage and updates clock documentation.

File	Description
DelayReleaseClockTest.java	Tests release timing and timestamping under clock skew.
SystemDatabase.java	Documents database-clock use for delayed releases.
WorkflowDAO.java	Uses database time for delayed transitions and updated_at.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.