## Summary

    When a subagent is configured with a custom `base_url` but no explicit `provider`, it should use `'custom'` as the provider rather than inheriting the parent's provider. This fixes credential resolution for subagents using custom endpoints like NVIDIA NIM.

    ## Problem

    The current code in `_build_child_agent()` always inherits the parent's provider when `override_provider` is not set:

    ```python
    effective_provider = override_provider or getattr(parent_agent, "provider", None)
    ```

    This causes issues when:
    1. Parent uses provider `alibaba` with model `kimi-k2.5`
    2. Subagent is configured with `base_url=https://integrate.api.nvidia.com/v1` and `model=nvidia/minimaxai/minimax-m2.7`
    3. Subagent inherits `provider=alibaba` from parent
    4. The wrong API endpoint and auth headers are used for the subagent

    ## Solution

    When `override_base_url` is explicitly set (indicating a custom endpoint), use `"custom"` as the default provider instead of inheriting from parent:

    ```python
    if override_base_url:
        # Custom endpoint configured - use override values or defaults
        effective_provider = override_provider or "custom"
        ...
    else:
        # No custom endpoint - inherit from parent
        effective_provider = override_provider or getattr(parent_agent, "provider", None)
    ```

    ## Additional Fix

    Also updates `_load_config()` to always read fresh from `config.yaml` instead of using the potentially stale `CLI_CONFIG` cache. This ensures delegation config changes are picked up without restarting the CLI (fixes issue #11999).

    The previous implementation short-circuited through `cli.CLI_CONFIG` which is populated once at module-import time and never re-read, causing config changes to be silently ignored in long-lived CLI sessions.

    ## Testing

    Tested with:
    - Parent: `kimi-k2.5` via DashScope (provider: `alibaba`)
    - Subagent: `nvidia/minimaxai/minimax-m2.7` via NVIDIA NIM

    Subagent now correctly uses `provider="custom"` with the NVIDIA endpoint instead of inheriting `provider="alibaba"` from the parent.

    ## Related Issues/PRs

    - Complements PR #12053 (config loading improvements)
    - Fixes credential resolution for custom endpoints in subagents
    - Addresses layer of issue #11999 (stale config in long-lived sessions)

    ## Checklist

    - [x] Code follows project style guidelines
    - [x] Changes are backward compatible
    - [x] Tested with real provider configurations
    - [x] Documentation/comments updated
    ```