Default /v1/models to downloaded-only, leave /models unchanged

[alytaphoenix]

Summary

/v1/models (the OpenAI-compatible model list, used by external clients/agents/integrations) was advertising every model card exo knows about, including ones nobody on the cluster has downloaded yet. Agents/tools that discover available models via this endpoint would see — and could try to call — models that don't exist locally.

/models (the dashboard's own model browser, used for finding new models to download) still needs to list everything by default. Both routes shared a single get_models handler, and a query-param default can't diverge per route, so this adds a thin get_v1_models wrapper registered only on /v1/models that defaults status="downloaded" and delegates to the existing (already-shipped) status=downloaded filter. /models keeps its unchanged default. ?status=all still returns the full list from either route — nothing that needs the full list is stranded.

Known limitations

• "Downloaded" is a loose proxy for "usable," not a guarantee. A model downloaded to disk but not yet placed as a running instance still 404s ("No instance found for model X") on first use — see API._validate_model_has_instance. This change only stops agents from being offered models that don't exist on the cluster at all; it doesn't guarantee a listed model is actually ready to serve on the first request.
• A multi-node model shows as "downloaded" once any single node has completed its download — not necessarily every node needed to actually serve it.

Out of scope (separate follow-up)

This is a default-filter flip only. A more granular "advertise this specific downloaded model to agents" checkbox — letting users curate a subset of downloaded models rather than all-or-nothing — would need persisted, cluster-synced state (a new command + event + State field + apply() logic) plus dashboard UI. That's a categorically bigger change than this one and is left as its own follow-up PR; this filter is forward-compatible with it (later: advertised = downloaded AND checked).

Test plan

• uv run basedpyright on touched files — 0 errors; full-repo baseline unchanged (319 errors, 110 warnings)
• uv run ruff check / ruff format — pass
• New tests: /v1/models defaults to downloaded-only, ?status=all returns everything, /models (no status) still returns everything — 3 passed
• uv run pytest src/exo/api — 50 passed
• Confirmed the dashboard's own model browser and chat model selector don't read /v1/models (only dashboard/src/routes/integrations/+page.svelte does, and it already tolerates a shorter list via || []/|| 0 fallbacks)

🤖 Generated with Claude Code

https://claude.ai/code/session_011rjSfwDBTkmySmfU6NgHKF