fix(llm): strip inline HTML + entities from rule text (refs #595)

src/llm/source.py

@@ -15,6 +15,8 @@ so parsing the comment id is the reliable path to a comment's docket.
 
 from __future__ import annotations
 
+import html
+import re
 from pathlib import Path
 from typing import Iterator
 
@@ -123,23 +125,38 @@ def _attachment_text(store: Store, item_key: str) -> str:
     return "\n\n".join(parts)
 
 
+_TAG_RE = re.compile(r"<[^>]+>")
+
+
 def _strip_html_wrapper(text: str) -> str:
-    """Drop the FR ``.txt`` attachment's HTML wrapper, if present.
+    """Drop the FR ``.txt`` attachment's HTML wrapper and inline markup.
 
     FR TXT attachments are HTML-wrapped: ``<html><head>...<body><pre>``
     then the plain text, closed with ``</pre></body></html>``. Strip
     everything through the first ``<pre>`` and any trailing close tags;
     text without that wrapper passes through unchanged.
+
+    The FR text itself also carries inline HTML — e.g. Cloudflare's
+    obfuscated ``<a href="/cdn-cgi/l/email-protection#...">`` mailto
+    anchors and page-break anchors. Those are stripped with a plain
+    ``<[^>]+>`` regex (this is a controlled, known corpus — not
+    attacker-supplied HTML — so no HTML-parsing dependency is needed),
+    keeping the anchor's inner text. HTML entities (``&amp;`` etc.) are
+    then unescaped. Plain-text markers like ``[[Page NNNNN]]`` contain no
+    angle brackets and pass through untouched.
     """
     idx = text.find("<pre>")
     if idx == -1:
-        return text
-    body = text[idx + len("<pre>") :]
-    stripped = body.rstrip()
-    for suffix in ("</pre></body></html>", "</pre>"):
-        if stripped.endswith(suffix):
-            return stripped[: -len(suffix)]
-    return body
+        body = text
+    else:
+        body = text[idx + len("<pre>") :]
+        stripped = body.rstrip()
+        for suffix in ("</pre></body></html>", "</pre>"):
+            if stripped.endswith(suffix):
+                body = stripped[: -len(suffix)]
+                break
+    body = _TAG_RE.sub("", body)
+    return html.unescape(body)
 
 
 def _rule_text(store: Store, item_key: str) -> str:

tests/llm/test_source.py

@@ -145,6 +145,27 @@ class TestRuleDocs:
             "item_key": rule_key,
         }
 
+    def test_inline_html_and_entities_stripped(self, store, rule_key, tmp_path):
+        txt = tmp_path / "2025-13271.txt"
+        txt.write_text(
+            "<html><head><title>x</title></head><body><pre>\n"
+            "[[Page 12345]]\n"
+            'Contact us at <a href="/cdn-cgi/l/email-protection#x">'
+            "someone</a> regarding CMS &amp; PFS.\n"
+            "</pre></body></html>"
+        )
+        store.attach_file(rule_key, txt)
+
+        docs = list(iter_rule_docs(store))
+
+        assert len(docs) == 1
+        text = docs[0].text
+        assert "<a href=" not in text
+        assert "</a>" not in text
+        assert "someone" in text
+        assert "CMS & PFS" in text
+        assert "[[Page 12345]]" in text
+
     def test_pdf_only_falls_back_to_extract_attachment(self, store, rule_key, tmp_path):
         pdf = tmp_path / "rule.pdf"
         pdf.write_bytes(b"%PDF-1.4 fake")