[FEATURE] Choose model and effort per task when suggesting a chip (spawn_task) or starting a session

### Problem Statement

An orchestrator session in the Claude desktop app (Code tab) suggests many background tasks with `spawn_task`, which accepts only `prompt`, `title`, `tldr` and `cwd`. The chip card has no model or effort picker. Every chip starts on the model and effort of the session that suggested it, so routine work runs on an expensive tier.

The only fix today is manual:

1. Start the chip.
2. Open its session.
3. Wait for it to start.
4. Change model and effort.
5. Send a message so the change applies.

The parent session can call `set_session_model` / `set_session_effort` on the child, but that triggers a permission pop-up and only applies from the child's **next** turn. A chip that finishes in its first turn never runs on the intended model. Our workaround is a handshake: the chip does setup and stops, the parent switches the model, then sends "continue".

### Proposed Solution

Optional `model` and `effort` parameters on `spawn_task`, and on `start_session` if that ships. Show them on the chip card, editable before launch, and apply them from the first turn. A permission prompt at launch is fine.

### Why

Choosing the cheapest adequate model for each task is a cost decision the dispatching session is best placed to make. Today it can only recommend a model in the task description, and the user has to apply it by hand after launch.

### Related

- #82683: chips inherit the parent's model instead of the user's default. Same mechanism, different ask.
- #89783: auto-starting child sessions, where `start_session` is mentioned.

Start by tracing the spawn_task entry point and the related start_session, set_session_model, and set_session_effort flows; the payload does not identify files or tests. Check how the chip card represents launch parameters and permissions. Done means optional model and effort can be selected or edited before launch and apply on the first turn, with existing behavior unchanged when omitted.

Written by the indexing model from the issue text.