[DO NOT MERGE][MLI-7219] feat(k8s-cache): log + metric on cacher Redis write failure

[lorenzo-norcini-scale]

Summary

When the k8s cacher fails to write endpoint status to Redis (bad auth, network partition, expired credentials), the failure was observable nowhere — no log, no metric. Entries then expire (60s TTL) and the Gateway reports endpoint status as unknown.

This wraps the write loop in ModelEndpointCacheWriteService.execute() to emit a logger.exception (traceback names the cause) and a new scale_launch.k8s_cache.redis_write_failure counter, then re-raise (visibility, not recovery — crash-loop behavior unchanged). Only the Redis write is wrapped; the k8s read and image cache stay outside.

Notes

• Scoped to the cacher only. The Gateway's read-through fill and the builder's create-time seed are left alone.
• Because the Gateway falls back to reading k8s on a cache miss, this metric is a cacher-refresh early-warning signal, not a direct measure of user-facing unknown.

Testing

New failure-branch test (raising repo → metric increments and exception propagates); happy-path test asserts the counter stays 0. Cache-service + redis-repo unit tests pass.

Resolves MLI-7219.

🤖 Generated with Claude Code