fix: shared batch length check in BaseDetector (follow-up to #113)

[YassinNouh21]

Summary

This is a follow-up to #113. @adaamko asked for it when closing #114:

1. A shared length check in BaseDetector that every detector calls; RAGFactCheckerDetector.predict_prompt_batch (rag_fact_checker.py:119) still zips silently.

This PR:

• Adds BaseDetector._validate_batch_lengths(prompts, answers). It raises ValueError when the two lengths differ. The message keeps the old text and adds both lengths: Number of prompts must match number of answers, got 2 prompts and 1 answers.
• Replaces the inline length checks in TransformerDetector, LLMDetector and RAGFactCheckerDetector with one call to the helper. The error type and the old message stay the same, so the perf: true batched inference for TransformerDetector (fixes #23) #113 tests pass without edits.
• Makes the zip at rag_fact_checker.py:119 strict.

About rag_fact_checker.py:119

RAGFactCheckerDetector.predict_prompt_batch already rejected a prompts/answers mismatch before this PR. The silent zip at line 119 pairs answers with the results from self.rag.detect_hallucinations_batch(...). If the backend returns fewer results than answers, zip drops the extra answers with no error, and the caller gets a shorter list. With strict=True, this case now raises ValueError. The project requires Python 3.11 or later, so strict=True is available.

Not in this PR:

• RAGFactChecker.detect_hallucinations_batch (lettucedetect/ragfactchecker.py:358) also zips contexts and answers with no length check. The detector path is safe now, because the detector checks the lengths first. A direct caller of RAGFactChecker can still get a shorter list. I can send a separate small PR for this if you want it.
• Idea 2 from the same comment (sort-by-length micro-batching).

Related issue

Follow-up to #113. Refs #114 and #23.

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: 218 passed, with HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1.
• Other: I ran the new RAG tests against the old rag_fact_checker.py. test_short_backend_result_raises_instead_of_truncating fails there with DID NOT RAISE ValueError. It passes with this change.

New tests:

• tests/test_inference_pytest.py::TestBatchLengthValidation: the helper raises with both lengths in the message. It accepts equal lengths, also two empty lists.
• tests/test_rag_fact_checker_pytest.py: this file is new, because the RAG detector had no tests. The backend is a MagicMock, so the tests need no API key and no network. The tests cover three cases:
  1. A length mismatch raises before the backend gets a request.
  2. Equal lengths give one result for each pair, in input order, for spans and for tokens.
  3. A short backend result raises.

The transformer and LLM detectors keep their #113 tests (TestPredictPromptBatchLengthValidation and TestPredictPromptBatchLengthValidationLLM). These tests pass without changes.

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).