fix: localize repeated LLM spans distinctly

[kocaemre]

Summary

• Map repeated LLM-returned span strings to distinct unused answer occurrences in left-to-right order.
• Reserve occurrences before applying is_hallucination and confidence filters so later accepted items keep stable offsets.
• Update generic LLM response instructions to ask for answer-order, one item per distinct occurrence.
• Add network-free regression tests for repeated strings, metadata preservation, filtering, duplicate overflow, first-match fallback, and token projection.

Why

The generic LLM judge path used a fresh first-match search for every returned substring. When the model returned the same text more than once, all items collapsed onto the first occurrence, so later repeated hallucinations were never represented in spans or token output.

Fixes #85.

Test plan

• python3 -m pytest tests/test_llm_detector_pytest.py::TestRepeatedSpanLocalization -q

• python3 tests/run_pytest.py -q — 149 passed

• ruff check lettucedetect/detectors/llm.py tests/test_llm_detector_pytest.py

• git diff --check

• I certify that I have the right to submit this code.

[adaamko]

Thanks — this is careful work, and the test coverage here is broader than what just merged. It crossed with #86 though, which was opened four days earlier for the same issue (#85) with the same reservation design, so first-come wins and #86 is now in main. Two parts of yours are genuinely additive and didn't land with #86: the response-format notes in _RESPONSE_FORMAT/_response_format_block (86 only updated the detection prompt), and several of your edge-case tests (metadata pairing, min_confidence reservation, the documented first-match fallback). A rebased follow-up PR with those would be very welcome — just include the rights-attestation line from the PR template, which this PR is also missing. Closing as superseded.