MCP prompt responses are submitted as JSON-encoded text

## Summary

MCP prompt responses are JSON-encoded before they are returned as `submit_prompt` content. This changes valid prompt text rather than preserving the text returned by the MCP server.

## Reproduction

1. Configure an MCP prompt that returns a text result such as `He said "hello".
Next line`.
2. Invoke the corresponding slash command.

## Actual behavior

`McpPromptLoader` calls `JSON.stringify()` on the returned text. The submitted prompt therefore contains literal outer quotes and escape sequences, for example `"He said \"hello\".\nNext line"`.

## Expected behavior

The `submit_prompt` action should receive the original MCP text unchanged, including embedded quotes and newlines.

## Impact

Every successful text response from an MCP prompt is altered before it reaches the conversation. Prompts with quotes, code, or multiple lines are especially visibly corrupted.

## Proposed fix

Pass the text response through directly and add a regression test that includes quotes and a newline.

I'd like to take this issue and submit a PR to fix it.

Start at McpPromptLoader and trace how the MCP text reaches the submit_prompt action, focusing on the JSON.stringify() call described in the issue. Add a regression test covering embedded quotes and a newline, then verify that the submitted content matches the original MCP response unchanged.

Written by the indexing model from the issue text.