feat(ui): add interactive REPL with cancellation and OAuth support

[yamaceay]

• Add interactive REPL with prompt toolkit for rich terminal UI
• Implement agent cancellation via ESC key during execution
• Add OAuth token acquisition and caching for MCP HTTP transport
• Integrate MCP servers during bootstrap phase
• Add context compaction progress spinner feedback
• Implement shared state coordination between input and agent threads
• Add .env file to .gitignore for local configuration

This introduces a new interactive mode that allows users to cancel long-running agent operations and provides better visual feedback during execution. The OAuth flow enables authenticated access to MCP servers over HTTP/SSE transports.

[chauncygu]

Thanks for the substantial work here, the threaded REPL, OAuth, and llmwiki plugin are all useful directions. But this PR bundles 6 unrelated features (~1300 LoC) into one commit and CI is red on 3.11. I'd like to split it before merging. A few items below are blockers regardless.

Blockers

1. tools/__init__.py rename is wrong. cc_mcp.tools → mcp.tools doesn't exist (the official mcp pip package has no tools submodule). The import silently fails; MCP tools only register because bootstrap.py adds an explicit import cc_mcp.tools. Please revert the rename or delete the dead entry.

2. Background-thread REPL has correctness gaps.

   • run_query is recursively re-entered (line 946). The new outer wrapper set/clear _agent_running clears the flag while the outer call is still running — main thread then spawns a second thread.
   • The old try/except KeyboardInterrupt around run_query is gone. Daemon-thread exceptions now die silently with no user feedback.
   • Escape calls event.current_buffer.reset(), dropping whatever the user was typing. Should only set _cancel_event.

3. AskUserQuestion routing races with the type-ahead queue. If the user has already pressed Enter to queue a message before the question fires, the Enter binding's _pending_question check eats that text as the answer. Also, wait() timeout returns the string "(timeout)" which the agent will treat as a real answer.

4. Non-verbose tool output regression. print_tool_start/end now overwrites with \r and erases on completion. In default mode the user can no longer see which tools the agent ran (only Edit/Write diffs and errors survive). This is a visible UX downgrade — please keep tool traces visible by default or gate behind a config flag.

5. OAuth security:

   • ~/.cheetahclaws/mcp_oauth_tokens.json is written without chmod 0600 — readable by other users on shared machines.
   • state is generated but never validated in _CallbackHandler.do_GET (CSRF; PKCE is not a substitute).
   • Hardcoded REDIRECT_PORT = 54321 with no fallback if the port is taken.

Smaller things

• bootstrap.py comment numbering jumps 3 → 5.
• ui/input.py imports prompt_toolkit.output.vt100._get_size (private API) — will break on PT upgrades.
• stream_text no longer prints chunks in the non-Live path; output is buffered until flush_response(). SSH / macOS Terminal users lose token-by-token streaming.
• MCPManager.call_tool replaced an O(1) dict lookup with an O(N) sanitize loop — please cache a sanitized → client map.
• commands/advanced.py removed the [:60] truncation on tool descriptions; long MCP tool descriptions will wrap awkwardly in /mcp.
• _load_env doesn't strip quotes (KEY="value" will include the quotes), but the docs claim quotes are fine.

Suggested split

1. Low-risk batch — can land first: .env loader, ANTHROPIC_BASE_URL, MCP env-var expansion, Path.expanduser() in plugin store, llmwiki plugin + docs.
2. OAuth — separate PR, after fixing token file mode, state validation, port fallback.
3. Threaded REPL + /btw + ESC cancel + AskUserQuestion rework — separate PR with tests, with the recursion / KeyboardInterrupt / type-ahead-race issues fixed and CI green.

Happy to re-review once it's split. Thanks again!