feat: Claude Code Stop-hook integration

[adaamko]

What

Implements the Claude Code hook from #50: lettucedetect/integrations/claude_code/ with a Stop-hook script runnable as python -m lettucedetect.integrations.claude_code.check_answer.

• Reads the hook event from stdin, extracts the last user/assistant messages from the session transcript (pure function, junk-tolerant), checks the answer against user-supplied grounding context (--context-file, default context.md; missing context exits 0 silently).
• Two modes per the issue: --api-url via LettuceClient.detect_spans (recommended, model stays loaded) and --model-path in-process, with optional --taxonomy-head so typed spans surface unsupported_addition ("the request did not ask for this") in the report.
• Exit contract: flagged spans print a report to stderr and exit 2 (fed back to the agent); clean answers exit 0. stop_hook_active short-circuits to exit 0 to prevent feedback loops.
• README with a ready-to-paste settings.json block and a 5-line walkthrough, following the langchain integration pattern.
• 15 network-free tests (tests/test_claude_code_hook_pytest.py) with stubbed detectors + a transcript fixture; transcript parsing additionally verified against a real Claude Code session transcript (Claude Code 2.x JSONL format, August 2026).

No changes outside lettucedetect/integrations/claude_code/ and tests/.

Closes #50

Checklist

• python -m pytest tests/test_claude_code_hook_pytest.py -v passes (15/15); python -m pytest full suite green locally; ruff format + check clean.

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

[adaamko]

Added a third mode: --llm-model KRLabsOrg/lettucedect-v2-qwen-2b --llm-base-url http://localhost:8001/v1 routes through the LLM detector, so the generative model (via vLLM or any OpenAI-compatible endpoint) and plain LLM judges both work; the detector's _is_generative heuristic picks the frozen prompt automatically. 17 tests now.

End-to-end verification with the real models (lettucedect-v2-mmbert-base + lettucedect-v2-taxonomy-head, run locally through the actual hook entry point):

Scenario: request says "Fix the bug in normalize_whitespace: consecutive tabs are not collapsed", context is the source file, and the agent's answer describes the correct regex fix plus a sentence nobody asked for: "I also added an LRU cache on normalize_whitespace to speed up repeated calls, and bumped the default max_length from 256 to 512."

Result: exit 2 with

LettuceDetect flagged 1 unsupported span(s) in the answer:
- "I also added an LRU cache on normalize_whitespace to speed up repeated calls, and bumped the default max_length from 256 to 512 for longer documents." (confidence 0.96, unsupported_addition/behavior — the request did not ask for this)

The grounded fix description is not flagged. Negative control (same request and context, honest answer without the extra sentence): exit 0, no report. This is the capability the integration is for: catching side effects the request never asked for, typed as such.