Fix content recommender SQL parameter binding

[cursor]

Summary

• Fix pgvector recommendation queries so SQL placeholders receive parameters in the correct order.
• Add regression tests for content recommendations with and without excluded recipe IDs.

Validation

• python3 -m pytest tests/test_recommend.py (apps/ml-service)
• python3 -m pytest (apps/ml-service)

Bug and impact

Trained users hitting the ML /recommend endpoint could trigger a database error because content-based recommendation queries sent the LIMIT value where the second %s::vector placeholder expected the preference vector, breaking personalized recommendations.

Root cause

The parameter list was initialized with both the preference vector and limit, then appended the order-by vector and limit again, creating too many parameters and misaligning the placeholders.