MN Uncensored v0.2.0

Released 2026-07-16.

Added

• API-triggered cold starts through the shared gateway.
• Automatic mode with a ten-minute idle GPU shutdown.
• Authenticated /wake route.
• Hermes custom provider with long cold-start timeouts.
• Explicit hard stop for running and pending model containers.

Fixed and changed

• The gateway recognizes Modal’s empty cold-start 503 response and waits for
  the backend instead of returning it directly to clients.
• Compressed upstream streaming chunks are decoded before forwarding.
• vLLM uses the pinned model snapshot and reproducible lifecycle configuration.
• Automatic mode no longer creates a permanently pending server.
• Recovery deployments require a clean worktree and verified signed HEAD.
• Stop transitions fail closed before a backend rollover.