v1.4.0

• Added --native-chat-streaming to route streaming /api/chat through LM Studio's native /api/v1/chat, so streamed responses carry real token stats and per-event reasoning (non-streaming stays on the default path).
• Added an optional inbound API-key gate (--api-key / OLLAMA_API_KEY): off by default, but when set every request must send Authorization: Bearer <key> (constant-time match, rejected requests logged).
• Added --auto-evict to unload other models before loading a requested one, matching Ollama's single-model default; it runs proactively before inference and warns at startup about the multi-client effect.
• /api/ps now reports a real expires_at from each model's tracked keep_alive: a finite deadline for a positive keep_alive, a far-future timestamp for keep_alive:-1, and the default placeholder only for models the proxy did not load.
• Fixed the native chat path (--use-native-chat) sending an input shape LM Studio rejected, which made native chat return empty or errored responses; the /api/v1/chat request is now built correctly.
• Fixed model loading when LM Studio's just-in-time loading is disabled, and cold-loading embedding models: the explicit load now uses the resolved model key instead of the :latest name LM Studio rejected.
• Fixed the native streaming path silently returning an empty success when LM Studio errored; it now surfaces the upstream status and message.

Full Changelog: v1.3.0...v1.4.0