[Clannad47]

Closes #117

Summary

The Claude Code Stop hook reported a hallucinated answer as clean whenever
--min-confidence was above 0.0 in --llm-model mode.

check_answer.py filtered spans with
(s.get("confidence", s.get("hallucination_score")) or 0.0) >= args.min_confidence,
so a span without a score counted as 0.0 and was dropped. No span carries a score in
--llm-model mode: the judge path builds LLMDetector without include_reasoning, so no
confidence key is produced, and the native v2 path (KRLabsOrg/lettucedect-v2-qwen-2b)
builds spans through gen.spans_to_offsets, which copies only category, subcategory
and explanation. main() then returned 0 — the exit code that means "nothing to
report" — so the agent was never asked to revise. --api-url and --model-path modes are
unaffected.

This change keeps unscored spans, the same rule as
BaseDetector._filter_spans_by_confidence (lettucedetect/detectors/base.py:91-96).

• check_answer.py: keep a span whose score is missing. The post-filter stays local on
  purpose: LettuceClient.detect_spans takes no threshold, and API spans use
  hallucination_score rather than confidence, so forwarding min_confidence into the
  detectors is not a drop-in for the API mode.
• --min-confidence help text and the integration README flag table now say the threshold
  applies to scored spans only.
• CHANGELOG [Unreleased] gains a Fixed entry.
• Three regression tests.

tests/test_claude_code_hook_pytest.py:

• test_min_confidence_keeps_unscored_spans — two unscored spans, --min-confidence 0.5,
  expects exit 2 and both spans in the report (the scenario from Stop-hook --min-confidence drops every --llm-model span (missing confidence treated as 0.0), reporting hallucinated answers as clean #117).
• test_min_confidence_mixes_scored_and_unscored — scored low / unscored / scored high in
  one result: the low-confidence span is dropped, the other two are kept.
• test_llm_mode_unscored_spans_reported_above_threshold — end-to-end --llm-model with a
  native-path span shape (no confidence), --min-confidence 0.5, expects exit 2.

The existing test_min_confidence_filters (scored span below threshold → exit 0) still
passes unchanged.

Related issue

Closes #117

Type of change

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

Testing

• ruff format --check lettucedetect/ lettucedetect_api/ tests/ — 69 files already formatted
• ruff check lettucedetect/ lettucedetect_api/ tests/ --extend-exclude lettucedetect/integrations/ — All checks passed!
• python -m pytest — 210 passed on Linux (Ubuntu 24.04, Python 3.12, the python-tests.yml platform). On
  Windows the 8 tests in tests/test_benchmark_detectors_pytest.py cannot import scripts/benchmark_detectors.py,
  which does import resource (Unix-only); those 8 pass on Linux, and the remaining 202 pass on Windows.
• Other: ran the hook in a real subprocess with a stubbed detector, --llm-model KRLabsOrg/lettucedect-v2-qwen-2b --llm-base-url http://localhost:8001/v1 --min-confidence 0.5 → exit 2 with the span report on stderr.

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).