[milansvitlica-codaxy]

Pins orkes-conductor-queues to 2.0.0.rc2. This is a CxODEV-2125 follow-up to the 3.32.5 release (server-3.32.5-rmq-kafka). The image from this change would be server-3.32.5-rmq-kafka.1.

Why

• Conductor 3.31.0 and later bundle orkes-conductor-queues 2.0.0.rc3. Its new background poller has a race (fix: prevent permanent per-queue poller stall in QueueMonitor orkes-io/orkes-queues#30, open, not released):
  • wakePoller() can cancel a backoff poll that is already running, so two poll chains share one slot.
  • When demand drops, both release it and the slot count goes to -1.
  • From then on that queue never polls Redis again until the JVM restarts. No error is logged, and /health stays up.
• The JOIN queue is hit first, so workflows stay RUNNING on a JOIN whose branches are done. We saw exactly that in one of our e2e runs on server-3.32.5-rmq-kafka.
• rc2 has no background poller: its QueueMonitor fetches on the caller's thread. It is the version 3.30.0–3.30.2 shipped.

Compatibility (checked)

• 3.32.5 uses only ConductorQueue, QueueMessage and the three-argument ConductorRedisQueue / ConductorRedisClusterQueue constructors. rc2 has them unchanged.
• JedisCommands (redis-api), which rc2 calls, has the same methods in 3.30.2 and 3.32.5.
• There are no lockfiles, and the server config has no rc3-only settings.
• Cost: rc3's speed work goes away (dedicated poller, push wake-ups, doorbell). Queue pickup goes back to 3.30.x behaviour.

Tested

• A local image built from this branch: its server jar contains only orkes-conductor-queues-2.0.0.rc2.jar.
• The CxO end-to-end suite on that image: 254 passed, 26 failed (all known pending), 0 flaky. All 286 scenarios have the same status as on server-3.32.5-rmq-kafka, with the same suite commit.
• :conductor-redis-persistence:test passed 63/63, and spotlessCheck is green.
• Not run locally: :conductor-test-harness:test. It hardcodes localhost for its Redis container, which my containerised Gradle can't reach. This PR's CI runs it.
• No new regression test. A JOIN-like delivery test against real Redis never stalled rc3 in 4 × 60 s, so it couldn't show the bug. The PINNED comment and the AGENTS.md row are the guard.

🤖 Generated with Claude Code