[Feat] Interactive Agent Builder: role templates + explicit per-role provider/model choice

## The problem

CAO's economic premise is **"a frontier model plans; cheaper models do the work."** Today that silently fails:

- The built-in `developer`/`reviewer` profiles pin **no provider and no model**.
- A worker with no provider **inherits the supervisor's provider** (e.g. `claude_code`); with no model it uses **that CLI's configured default**.
- So on a machine whose Claude default is a frontier model, every "cheap worker" actually runs the most expensive model. Observed live: planner + developer both on the frontier tier — zero savings.

The fix isn't "remember to add a `model:` line to a markdown file" — most users won't know that field exists, or that omitting it routes to the frontier (see also the closely-related #174, where the `model` frontmatter was ignored at terminal-creation time). The fix is to make **role + provider + model an explicit, guided choice in the UI**, with the cost of each choice visible.

## What I'd like to build

An **Agent Builder**: a screen where a user creates a reusable agent by picking a **role template**, choosing its **provider + model**, and giving it an **avatar + name** — which writes a normal profile to the agent store, so it plugs into everything that already exists. *Keep the markdown-profile system; add a UI on top of it.*

**Two connected surfaces:**

1. **Agent Builder** (create/edit a profile)
- Pick a **role template** (Developer, Code Reviewer, Planner/Supervisor, Tester, Researcher, Doc Writer, … + Custom) that pre-fills the system prompt, recommended tool restrictions (role→tools defaults already exist), and a *recommended* model tier.
- Choose **provider + model** — the core fix. Model dropdown populated per provider (claude: opus/sonnet/haiku/fable; kimi: kimi-for-coding; …) with a **cost/speed hint** per option (frontier · balanced · cheapest). Default to the template's recommended tier, **not** "inherit".
- **Avatar + friendly name** (dependency-free: curated emoji picker or auto initials-on-color).
- **Live preview** of the resulting agent card, then **Save** → writes `.md` to the agent store (frontmatter: name/description/role/provider/model/avatar/mcpServers + the template's prompt body).

2. **Team composition in the run wizard**
- A "Who's on the team?" step mapping each role the planner uses (developer, reviewer) to a built agent — i.e. *which model fills which role for this run* — showing the projected mix ("Planner: · Developer: Sonnet · Reviewer: Kimi") with a plain-language cost note. This is what makes the planner actually delegate to a `dev-sonnet` instead of the unpinned `developer`.

## Backend

- **Role-template library** ({role, description, recommended_model_tier, system_prompt, recommended_tools}); seed from the existing `code_supervisor`/`developer`/`reviewer` prompt bodies.
- `GET /agents/templates`, `GET /agents/models?provider=…` (selectable models + cost/speed hint), `POST /agents/profiles/create` (build + install a profile `.md` from structured fields, validating the model against the provider — extends the existing install/profile machinery, doesn't reinvent it).
- Carry `avatar` + `model` through `GET /agents/profiles` and session detail (model already flows via `effective_model`).

## Frontend

- `AgentBuilder` modal (role-template grid → provider+model with cost hints → avatar+name → live preview → save), an `AgentAvatar` component reused on the run board / member rows / Agents tab, a new "Agents" library view, the run-wizard team step, and avatars on flow-graph nodes instead of bare dots.

## Phasing

1. **P0 (stops the bleeding):** make per-role model choosable — pin sensible models on the built-in worker profiles (developer→sonnet, reviewer→sonnet/opus) and add a model dropdown to the existing Spawn/Add-Agent forms. Immediately reduces frontier spend.
2. **P1:** the Agent Builder modal (templates + provider/model + name) writing profiles; the Agents library view.
3. **P2:** avatars everywhere; the run-wizard team-composition step with the cost estimate.
4. **P3:** custom role templates, duplicate/edit, export/sharing.

Happy to implement and upstream this as scoped PRs if there's interest.