v3.43.4

Fixed

• A model provider that failed its health check when Home Assistant started stayed disabled until the integration was reloaded. On a busy boot the checks' 2-second timeouts can fire late and together, so working cloud providers were marked down, and with the primary (for example a local Ollama) really down, every turn answered "I'm sorry, I was unable to respond in time" for the whole session. A provider that a model category uses (as primary or fallback) and that failed its check is now checked again once Home Assistant has started, then after 30 seconds, 1, 2 and 5 minutes, and every 5 minutes while it stays down; when it answers, the integration reloads itself to use it. This also switches back to a primary that comes back online later, which used to need a manual reload. These reloads happen at most once every 5 minutes, backing off to once an hour if a reload does not fix the provider, so a provider that answers only some of the time cannot reload the integration in a loop. A provider whose key is rejected is checked once more and then left alone. The health-check warnings now name the cause, such as TimeoutError or HTTP 503, instead of ending in an empty message (#711).

• An Ollama model provider used only as a fallback, at an address other than the default http://localhost:11434, was never checked or set up, so its fallback did nothing: if the primary (for example OpenAI or Gemini) failed, there was nothing to fall back to, and nothing was logged above debug level. Every Ollama provider in a fallback chain is now health-checked at its own address and used for chat, vision, summarization and embeddings like a primary, and one that is down at startup is re-checked like any other provider. A fallback Ollama server embeds with the configured Ollama embedding model, so its vectors stay comparable with the ones already stored. The startup Ollama diagnostics (loaded models, shared servers) now cover only servers a model category uses and that answered, instead of probing the unused localhost default when Ollama is only a fallback (its health check still runs and logs a warning, as before); Ollama addresses in these and other log lines no longer show a username or password; and a malformed saved Ollama address fails its health check instead of stopping the integration from starting (#749).

Patch release on top of 3.43.3; it does not include the features waiting for the 3.44.0 bundle.