test: cover repeated LLM span follow-ups

[kocaemre]

Summary

• add the repeated-span ordering guidance that was left out of the merged Fix repeated LLM span localization #86 prompt instructions
• add focused regression coverage for two additive edge cases from fix: localize repeated LLM spans distinctly #87: low-confidence reservation and the single-item first-match fallback
• add prompt-format tests for the simple and reasoning response schemas

Related issue

Follow-up to #85 after #87 was closed as superseded by #86.

Type of change

• Bug fix
• Feature
• Documentation
• Tests
• Refactor or maintenance

Testing

• ruff format --check lettucedetect/ lettucedetect_api/ tests/
• ruff check lettucedetect/ lettucedetect_api/ tests/ --extend-exclude lettucedetect/integrations/
• python -m pytest
• Other: python -m pytest tests/test_llm_detector_pytest.py -q

Checklist

• I kept the PR focused on one change.
• I added or updated tests/docs when needed.
• I checked that no secrets, API keys, or credentials are included.

Rights & sign-off (required)

• I certify that I have the right to submit this code and that it may be
  distributed under the repository's MIT license
  (see CONTRIBUTING).

[adaamko]

Thanks for coming back with the follow-up — this is exactly the additive part from #87 that didn't land with #86, and it's green on top of current main (182 passed). The two new _to_spans tests pin the remaining #85 acceptance items (low-confidence reservation, single-item first-match fallback), and the response-format tests give _response_format_block its first coverage.

Two cosmetic notes, neither blocking: test_simple_response_format_requests_answer_order passes cache_file="unused" where the sibling test uses the cache_file fixture (harmless — nothing is written unless set() is called — but worth aligning in a follow-up); and since this changes the prompt text in llm.py (and so the generic judge's cache key), I'm squashing it as fix: rather than test:. Merging.