feat: inject rejected-tool feedback as user message (#96)

[VforVitorio]

Summary

Finishes the last missing piece of #96 (extended interactive prompt).

The arrow-key approval menu in ask mode already offered three options
(Yes, No / Tell lmcode what to do instead, Always) thanks to
PR #95, but when the user typed a redirect instruction after choosing
No, the text was returned inline as a tool error string. That made
the model treat the user's words as tool output and the context felt
off.

This PR routes the feedback through Agent._pending_user_feedback and
replays it as a real user turn on the next iteration of the chat loop.

Changes

• Agent.__init__: new _pending_user_feedback: str | None slot.
• _wrap_tool: when the user types redirect text, stash it and return
  a generic Tool execution cancelled by user error so the model
  stops retrying and waits.
• run(): new auto_input variable. When set, the next loop iteration
  skips session.prompt_async and replays the feedback as if the user
  had typed it at the prompt (complete with a dim › <text> row so the
  injection is visible).
• The interrupted path clears _pending_user_feedback so Ctrl+C cannot
  leak a stale redirect into the next turn.
• CLAUDE.md: documents ui/_interactive_prompt.py in the module map
  and key files table.

Resulting history flow

```
user: write to /x/y.py
assistant: [tool_call write_file]
tool: error: Tool execution cancelled by user. Do not retry...
assistant: ok, cancelled. Waiting for your new instructions.
user: write it to /x/z.py instead ← injected as a real user turn
assistant: [tool_call write_file with new path]
```

Test plan

• `uv run ruff check .`
• `uv run ruff format --check .`
• `uv run mypy src/`
• `uv run pytest` (199 passed)
• Manual: in ask mode, ask for a write_file, pick No / Tell lmcode what to do instead, type a redirect, verify it appears as a new user turn and the model acts on it.

Closes #96.