Problem

On a real box driving the TUI over ssh, four things read as regressions:

• "Keep engine warm" re-warmed on every return. The idle TTL unloaded the model after engine_idle_ttl_minutes (default 5) regardless of the setting, and nothing kept the engine resident while the TUI was open.
• Switching chat models locked the input with no indication of why, and the new model only loaded on the next prompt, so the input came back before the model was ready and typing silently did nothing.
• Cold model load felt slower than it should. The chat server defaulted to --no-mmap on local disk, forcing a full upfront read of the weights instead of mmap's lazy paging.
• The first-start splash left a cursor blinking on the wordmark.

Solution

• _warm_ttl_seconds returns 0 (weights resident) whenever an interactive session owns the process, keep_engine_warm is set, or a bulk ingest is holding the fleet. A TUI session keeps the engine warm for its lifetime; closing lilbee releases it. The intent is recorded on the services state before the container builds and threaded through create_provider -> RoutingProvider -> FleetProvider, so it ends up as provider instance state rather than a process-wide flag.
• The model swap keeps the chat input disabled until the new model actually serves, with a placeholder naming the target and saying why the box is locked, and eagerly warms the new model instead of deferring its cold load to the next prompt.
• The warm line is computed regardless of task activity and outranks the task summary. It was previously computed only when the task bar was idle, so the document sync that starts on launch suppressed it entirely — leaving a locked input with no explanation for the whole load.
• The indeterminate phases use the sweeping bar (a spinner glyph read worse in use); the line now also carries the model name and phase.
• Local-disk chat weights use mmap; --no-mmap is kept only for a network filesystem, where it avoids wedging the loader in uninterruptible I/O.
• The onefile bootstrap hides the parked cursor at handoff so it stays hidden until Textual takes over.

Verified on hardware

Measured on a GTX 1070 Ti box running Jan v3.5 4B and Qwen3 0.6B, from source:

• Keep-warm: with keep_engine_warm = false, the live llama-swap configs wrote ttl: 0 and both servers stayed resident continuously past the 5-minute idle window (GPU held at 4917 MiB). Under the previous ttl: 300 they were evicted at 5 minutes.
• Model switch: instrumented end to end. The input is disabled at +0.00s and released at +37.1s, exactly when wait_chat_ready reports the model can serve — so it never re-enables in front of an unloaded model.
• mmap: the chat llama-server argv no longer carries --no-mmap on this local-disk host.
• Splash cursor: not verifiable from source, since it lives in the C onefile bootstrap that only exists in a built executable. The patch hunk arithmetic was corrected and validated; the visual check needs a build.