Azure gpt-5.x REST: max_completion_tokens, no temperature, urllib RemoteDisconnected retry

Calling Azure OpenAI gpt-5.x deployments (5.4-nano/mini, 5.5, 5.6) over plain chat/completions works with api-version 2024-10-21, but you must send max_completion_tokens (NOT max_tokens) and must NOT send temperature (reasoning models reject non-default temp); response_format {type:json_object} works. A stdlib urllib client must catch http.client.RemoteDisconnected in its retry clause — it subclasses ConnectionResetError/OSError, NOT urllib.error.URLError or TimeoutError, so except (URLError, TimeoutError) misses transient disconnects and crashes. Broaden to (URLError, TimeoutError, ConnectionError, http.client.HTTPException, OSError).