Add safe-outputs.url-policy with audit and reputation URL handling - #34868

safe-outputs only supported strict domain allowlisting, which made open-web research outputs unreadable by redacting nearly all citations. This PR introduces policy-driven URL sanitization so workflows can opt into broader URL handling without removing safety controls entirely.

• Policy model for URL sanitization

  • Added safe-outputs.url-policy with:
    • allowlist (current/default behavior)
    • audit (preserve URL, log out-of-allowlist domains)
    • reputation (preserve unless flagged malicious)
  • Added safe-outputs.reputation config:
    • provider (currently google-safe-browsing)
    • api-key-secret

• Compiler + schema integration

  • Extended workflow schema to include url-policy and reputation.
  • Added config parsing and validation for new fields.
  • Enforced reputation.provider and reputation.api-key-secret when url-policy: reputation.
  • Propagated policy/reputation settings into safe-output runtime env vars.
  • Included new meta fields in safe-outputs import merge behavior.

• Sanitizer behavior updates

  • allowlist: unchanged redaction behavior.
  • audit: keeps external URLs intact and emits audit logs for non-allowlisted domains.
  • reputation: checks non-allowlisted URLs against Google Safe Browsing and redacts only malicious matches.
  • Added bounded in-memory caching for reputation lookups and fail-open behavior on provider/API errors.

• Docs and contract clarity

  • Updated safe-outputs reference docs to describe policy modes and reputation configuration.

safe-outputs:
  url-policy: reputation
  reputation:
    provider: google-safe-browsing
    api-key-secret: SB_API_KEY
  allowed-domains: [default-safe-outputs]