Allow per-agent-type model overrides in settings

Problem

Built-in subagent types (Explore, Plan, etc.) have hardcoded default models. When a user has many MCP servers configured, the combined tool definitions overflow smaller models' context windows, causing subagents to fail silently with "Prompt is too long" and 0 tokens used.

Current behavior

• The Explore agent defaults to Haiku for speed
• Each MCP server injects tool definitions into the subagent's context
• With ~13+ MCP servers, the system prompt alone exceeds Haiku's context limit
• The agent returns immediately with 0 tokens / 0 tool uses — no error surfaced to the user
• The only workaround is passing model: sonnet per-invocation, which requires knowing about this issue

Proposed solution

Add a subagentModelOverrides setting (or similar) in settings.json / settings.local.json:

{
  "subagentModelOverrides": {
    "Explore": "sonnet",
    "Plan": "sonnet"
  }
}

This would let users with heavy MCP setups configure larger models for subagent types that can't fit their context.

Additional considerations

• Subagents probably shouldn't inherit all MCP tool definitions — Explore only needs Read/Glob/Grep/Bash, but currently gets 300+ MCP tools
• The "Prompt is too long" failure is silent — surfacing it as a visible error/warning would help users diagnose the issue
• An alternative approach: filter tool definitions passed to subagents based on which tools the agent type actually has access to

Environment

• macOS, Claude Code CLI
• Model: opus[1m]
• ~13 MCP servers configured (PostHog, Linear, Notion, Todoist, Slack, Granola, Langfuse, shadcn, AG Grid, BrowserMCP, Chrome DevTools, Figma)