Enable pinned multilingual Prompt Guard

[0xsl1m]

Summary

• add use_transformer="multilingual" for the true mDeBERTa-based Meta Prompt Guard 2 86M model
• pin known gated Meta models to reviewed immutable Hugging Face revisions
• resolve authentication only through HF_TOKEN/the HF cache; never retain or log raw tokens
• add a credentialed same-repository CI smoke plus an honest deepset benchmark; fork PRs receive no secret
• correct documentation that previously described the faster 22M DeBERTa-xsmall model as multilingual mDeBERTa

Local validation

• 323 passed, 3 opt-in model skips
• strict mypy clean across 54 source files
• Ruff lint/format clean across 80 files
• actionlint clean

Release gate

The trusted-branch ML job intentionally fails if HF_TOKEN is missing or the associated Hugging Face account has not been approved for the manually gated Meta model. The credentialed live-model result and measured benchmark must be green before merge.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
shadowshield	Ready	Preview, Comment	Jul 25, 2026 11:38pm