Extend Ollama preflight to orchestrator and YAML persona models

[rosspeili]

Problem Statement

Ollama preflight (#43) validates only settings.defaults.litellm_model. Users can configure defaults.orchestrator_model or per-persona model fields in rooms.settings.yaml with different ollama/… tags. Preflight exits success, then the session fails later on the first orchestrator or persona turn.

This undermines the purpose of preflight for anyone using non-default orchestrator or custom persona models.

Proposed Solution

1. Collect all ollama/… model strings from settings before the wizard:

   • defaults.litellm_model
   • defaults.orchestrator_model (if set)
   • Each entry in personas[].model when personas is non-empty

2. Deduplicate tags; one /api/tags call; fail with the same Rich panel style if any tag is missing or Ollama is unreachable.

3. Document in docs/ARCHITECTURE.md that wizard-only custom agents (added interactively) are validated at inference time (feat: surface inference failures in CLI with recovery hints (not as agent speech) #37), not at startup.

4. Tests with mocked HTTP covering multi-model settings.

Acceptance criteria

• Preflight fails when orchestrator tag missing but default tag present
• Preflight fails when YAML persona tag missing
• Single /api/tags fetch per run
• ARCHITECTURE documents startup vs wizard-time limits

Alternatives Considered

• Validate only at inference (feat: surface inference failures in CLI with recovery hints (not as agent speech) #37 alone) — rejected; misses cheap early failure for YAML-configured models.
• Re-run full preflight after wizard — see issue feat: add timeout to Agent.generate_response() #2 below as complement.

Priority

High 🔴

Additional Context

No response