Support denyMainOnly permission scope: restrict tools to subagents only

Problem

When a tool is denied (via deniedTools in settings or hook deny decision), the restriction applies equally to the main conversation and all subagents. There is no way to deny a tool to the main conversation while allowing subagents to use it.

Use Case

I want to restrict write-access MCP tools (e.g., patch_object, delete_object) so the main conversation LLM cannot call them directly, but specialized subagents — which have focused prompts and limited tool sets — can. This provides a safety layer: the main conversation must delegate writes to purpose-built agents rather than calling write tools directly.

Example settings (desired behavior):

{
  "permissions": {
    "denyMainOnly": ["mcp__server__patch_object", "mcp__server__delete_object"]
  }
}

Current Behavior

• deniedTools / deny blocks the tool for both main and subagents
• PreToolUse hooks cannot distinguish main from subagent — all fields (session_id, transcript_path, permission_mode, env vars) are identical
• No workaround exists short of process tree inspection (fragile)

Proposed Solution

Add a permission scope that denies tools to the main conversation but allows subagents to use them. Possible approaches:

1. New permission list: denyMainOnly — denied for main conversation, allowed for subagents
2. Subagent override: allowForSubagents list that overrides deny when called from a subagent
3. Hook context field: Add is_subagent / caller / agent_depth to PreToolUse hook input so hooks can make the decision themselves

Option 1 or 2 would be the cleanest — no hook scripting needed.

Why This Matters

Subagents are the natural delegation mechanism in Claude Code. Being able to restrict the main conversation to read-only while letting specialized agents write is a practical safety pattern, especially for MCP tools that modify external systems.