[sosidudku1]

What

Prepares the terminal for a new local model launch, ahead of the model itself.

• Catalog entries get an optional spotlight flag. No entry sets it, so nothing changes for users until a launch adds one.
• When a flagged model fits this machine's memory and none of the flagged models is on disk, the TUI prints one startup line. It appears at most once per model id, never during first-run setup, and the seen list lives in the state dir, so there is no config migration.
• atomic-agent models pull <id> --run downloads the model, makes it active and starts it. On a vision model it also fetches the projector, as the TUI does by default. --run with --background is refused.

Why

Today models pull, use and start are three commands, and nothing tells a terminal user that a new model exists. A launch should only need catalog rows.

Notes for review

• --run stops a running chat model first, because models start refuses otherwise. A TUI open in another terminal loses that model.
• If the new model fails to start, it stays active, same as the TUI.
• Fit is decided from RAM only. GPU memory could be added once Local engine: recommend the big model for this machine and fetch it in the background (ATO-1043, ATO-15) #1001 lands.
• Companion desktop PR: feat(desktop): one-click card for a newly launched local model #1101.

How it was verified

• Second Mac: build and lint clean; vitest for src/cli/models-command, src/cli/models-pull-run, src/local-llm/catalog and src/tui/local-models: 45 files, 440 tests green.
• Not yet exercised live with a flagged model or a real download.

[sosidudku1]

Merged as part of #1106.