Summary

Fixes critical bug where broken connections were blindly reused during cluster
topology refresh, causing persistent "wrong shard/slot" errors requiring pod restarts.

Root Cause

During _refresh(), connections were reused without health validation. Network
issues could break connections, which would persist in error state indefinitely
across refresh cycles.

Changes

1. Critical, Connection health validation and reuse:
   Validate connection health with Error() before reusing

• The cluster client now checks the health of existing connections before reusing them during a topology refresh. Connections reporting errors are replaced with new ones, preventing broken connections from being reused.

2. Slot mapping resilience:

• When the cluster topology information is incomplete (e.g., missing slots), the client preserves old slot-to-connection mappings for slots not covered by the new topology, ensuring no slot is left without a connection.

3. Replica slot fallback improvements:
   Add replica-to-master fallback for missing replica slots

• In scenarios where replica slot information is missing, the client now falls back to using the master connection for those slots, ensuring continued availability. [1] [2] [3]

Impact

• Broken connections now auto-detected and replaced within seconds
• No more manual pod restarts needed for this issue
• Graceful degradation during incomplete topology updates
  This pull request improves the reliability of the Redis cluster client by ensuring that unhealthy (broken) connections are not reused after a topology refresh, and by making slot mapping more robust when the cluster topology is incomplete. It also adds comprehensive tests to validate these behaviors.

Testing enhancements:

• Added tests to verify that broken connections are not reused after a refresh, and that all slots are properly updated to use healthy connections, ensuring robust failover and recovery.