Add Privacy Guard NER engine

[johnnygreco]

Summary

• add a resource-backed built-in ner entity-processing engine with policy-owned labels, threshold, overlap behavior, and deterministic replacement
• add a provider-neutral NERModel facade with explicit /v1/extract and serialized local-model implementations
• register NER only when deployment resources are supplied, while preserving the default Regex-only registry
• add endpoint and local registry examples, mixed-engine coverage, safety limits, and architecture documentation
• keep the Anonymizer chat-completions endpoint as a future dedicated adapter without protocol autodetection

Why

Privacy Guard needs general named-entity recognition without coupling policy to a model checkpoint, endpoint, or execution location. This preserves the existing engine boundary while allowing operators to deploy the same policy against a remote GLiNER-compatible endpoint or an already-loaded local model.

Validation

• make check — 297 passed, 2 explicitly opt-in live-model smoke tests skipped
• make check-py311
• python3 tests/test_render_dev_notes.py
• scripts/build-docs.sh

Notes

The live DGX endpoint and local downloaded-model smoke tests remain opt-in through explicit environment variables; normal validation never contacts or downloads a model.

[github-actions]

Documentation preview

View the deployed preview

Built from af25bc4.