test: add end-to-end integration test for classify_extract.py

[SeanClay10]

Problem

The pipeline has unit tests for individual modules but no integration test that exercises the full classify_extract.py entry point against a known input. There is no automated check that the pipeline produces valid output end-to-end, or that failure modes (e.g. a not-useful PDF) are handled correctly.

Tasks

• Add at least one end-to-end integration test using a fixture PDF or synthetic text file with known correct outputs
• Assert that the output JSON passes schema validation (PredatorDietMetrics)
• Cover at least one failure-mode path — e.g. a not-useful PDF is correctly rejected before extraction reaches the LLM
• Ensure the test can run in CI without a live Ollama server (mock or stub the LLM call as needed)

Context

Commitment carried over from the Sprint 10 retrospective: every pipeline PR should include an integration test covering a failure mode.