privacy-linterlisted

# Privacy Linter Catches leakage at the one cheap intervention point — the moment of creation — rather than after content is already committed or posted. Advisory by default: a low false-positive rate matters more than catching everything, because an ignored linter is useless. ## Why this is deterministic-only, not model-based The design this implements (`Projects/Privacy OS - Pre-Disclosure Privacy Linter` in the source vault) requires local-only analysis: **sending content to an external LLM API for leak detection is itself a potential disclosure.** Of the four leak classes in that design — Direct PII, Metadata, Inference cues, Stylometric fingerprint — only the first two are mechanically detectable without a model. This version implements those two, end to end, with real tests, the same order `deid-reid-harness` used for its own tracks (ship the model-independent slice first) — plus **Secrets/credentials**, a class added afterward that isn't part of the source design's original four but is mechanically detectable the exact same way (see `references/leak-taxonomy.md` for why it's called out as an addition, not silently folded into "Direct PII"). Inference cues and stylometric fingerprinting are documented, not built — see `references/leak-taxonomy.md` — because building them today would mean either standing up a local-model pipeline (real setup work, not yet done) or using an external LLM for exactly the analysis this project exists to keep local. Don't silently reach fo