Brief Sanitize

Sanitizer for untrusted text (email bodies, calendar invites, scraped pages, third-party payloads) before it is fed to an LLM prompt. Strips HTML to plain text via a real parser (sanitize-html), decodes entities, strips Unicode obfuscation characters (NFKC-normalized, \p{Cf} + variation selectors), defangs http(s)/javascript/data/protocol-relative URLs, truncates, and wraps each item behind a random per-call nonce delimiter (Microsoft "spotlighting", arXiv:2403.14720) so a downstream prompt can safely quote it as data rather than instructions. Defense-in-depth against prompt injection: no LLM, no network, no secrets.