v0.4.1

• Apply extraction-method quality gates to each required evidence type independently.
• Prevent unrelated manual/parser evidence from admitting an LLM-only required result.
• Add an attributed, frozen VBO canine name-mapping case and source rebuild script.
• Evaluate 72 real-source names, 160 controlled faults, and 16 explicit trust-boundary cases separately.
• Preserve generic main and the legacy canine-breed branch.
• Publish to PyPI from version tags; add package metadata and citation file.
• Reject CLI input nested deeper than 100 levels (exit 3) on every platform and Python version.