Switched EventHorizon's shared text-inference model from inference-deepseek-v32 to inference-glm5 in the repository-managed local and Kubernetes configuration.
This changes persona decisions, narration, rule checking, item/effect evaluation, quick actions, player metrics, and persona-memory extraction. The embedding model remains inference-bge-m3; speech-to-text remains inference-whisper-large-v3.
EventHorizon/EventHorizonfcd30a583c33941a1f8cee8371c111036e9393e1gitea.app.lan/eventhorizon/eventhorizon:sha-fcd30a583c33gitea.app.lan/eventhorizon/eventhorizon:sha-2191ec0e93bb/v1/models advertised the exact inference-glm5 model identifier.cargo test --workspace: 345 backend tests and 8 frontend tests passed; doc tests passed./app/config.toml and the live ConfigMap both report model = "inference-glm5"./healthz and /readyz returned HTTP 200 after deployment.A naturally queued background persona_memory_extraction request (about 10 KB) has repeatedly exceeded its existing 15-second background budget. Five provider attempts recorded failure_kind=timeout, zero response bytes, and approximately 15-second duration between 14:12 and 14:17 UTC. The durable job is requeued with increasing retry backoff; optional background work is also protected by the provider circuit breaker.
Foreground operations bypass that optional/background circuit and retain their longer phase budgets, but no natural post-deployment gameplay turn had yet validated GLM-5 persona, narration, and transactional item-state phases at documentation time. Service readiness remains healthy at 2/2 replicas with zero restarts.
Monitor llm_usage_events by model and operation after natural gameplay, paying particular attention to structured response validity, narrator quality, phase duration, timeouts, and item-state completion.
Restore model = "inference-deepseek-v32" in config.toml and k8s/02-config.yaml, then deploy through the normal repository workflow. The previous immutable image is sha-2191ec0e93bb.