Replace VectorDB pickle caches with JSON

[armorer-labs]

Summary

• replace the evaluation VectorDB pickle caches with JSON caches
• convert the checked-in classification and text-to-SQL vector DB cache files to JSON
• update the matching guide notebook snippets so the cookbook examples use the same cache format

Why

These examples only use the cache files as local embedding stores, so JSON is enough for the stored data and avoids teaching pickle.load as the default copy-paste pattern. Existing generated .pkl caches are not loaded automatically; users can regenerate them by rerunning the relevant example setup. The checked-in caches used by the cookbook are converted in this PR.

Related

• Related to Security hygiene: copy-paste-ready vulnerable patterns + Dockerfile running as root #531, which called out pickle use in these evaluation VectorDB examples as security hygiene.
• Related to Fix: Inconsistent serialization of query_cache in VectorDB #168, which touches VectorDB serialization consistency but keeps the pickle format.

Validation

• python -m py_compile capabilities/text_to_sql/evaluation/vectordb.py capabilities/classification/evaluation/vectordb.py capabilities/retrieval_augmented_generation/evaluation/vectordb.py
• git diff --cached --check
• confirmed no remaining pickle references in the targeted classification, text-to-SQL, and RAG evaluation/example files with rg
• ran offline behavior checks with a fake Voyage client covering missing-cache generation, JSON save/reload, shipped cache loading, search results, cached-query reuse, and notebook code parsing