#963 feat(server): stop unchanged completed tool-call loops

Description

## Summary

- Add opt-in `tool_loop_max_repeats`; the default remains disabled.
- Stop unchanged single-tool cycles before loading or running the model.
- Return an explicit safety notice, not a claim of task success.
- Preserve changed results, exit codes, payload timestamps, and asynchronous waits.
- Keep the original loading order for normal and vision requests.

The configured ceiling also stops intentional identical terminal calls. Set the value to `0` to disable it.

Related: #728. This guard does not claim to fix model reasoning or quantization quality.

## Verification

- `python tools/test_tool_loop_guard.py`: 13 tests passed.
- Parser and client-shape regressions: 16 tests passed.
- Run configuration regressions: 9 tests passed.
- Python compilation, Ruff, and whitespace checks passed.
- Live Chat, Responses, and Anthropic replays passed with streaming enabled and disabled.
- All six live replay modes returned the notice without executable calls or new engine log output.
- A normal live tool request still executed and produced a final blocker report.

The full API suite ran 167 tests with one error in `OverHttp.test_json_schema_text_format`.
The same error occurred on unchanged commit `6f32ec0`; this patch does not remove or weaken that test.