feat(wiki): persist configurable multimodal knowledge bundles

[MarinaMackay]

Summary

Part 5 of the multimodal repository knowledge stack.

Upstream tracking issue: sysevol-ai#655

This slice hardens the stack by adding configurable VLM visual fact extraction and a persisted, versioned multimodal knowledge bundle format.

Changes

• Add QAConfig and environment-variable wiring for optional visual fact VLM extraction.
• Add CLI flags for OpenAI-compatible visual fact extraction.
• Add codenib.multimodal-knowledge-bundle.v1 persisted bundle schema.
• Add component hashes, bundle hash, atomic save, bounded load, and validation.
• Update documentation.

Testing

python -m pytest test/scripts/test_build_multimodal_knowledge.py test/wiki test/web/test_config.py
git diff --check

Latest local result:

413 passed

[MarinaMackay]

Closing this older draft because config/storage is now part of the smaller replacement stack after PR4a-PR4d.