fix(llm): strip NUL/control chars from chunk text before pgvector insert

Extracted PDF comment text carries NUL (0x00) and form-feed bytes that
Postgres text columns reject (DataError: cannot contain NUL bytes) — the
2017 pilot was clean by luck, the 2018+ dockets crash the indexer. Strip
C0 control chars (except tab/newline/return) in chunk_doc.

src/llm/chunk.py

@@ -35,6 +35,10 @@ def content_hash(text: str) -> str:
 
 _FRONTMATTER = re.compile(r"\A---\n.*?\n---\n", re.DOTALL)
 _HEADING = re.compile(r"^#{1,6} ", re.MULTILINE)
+# C0 control chars except tab/newline/carriage-return. Extracted PDF text
+# carries NUL (0x00) and form-feeds that Postgres text columns reject
+# ("cannot contain NUL bytes") and that add no semantic value to embeddings.
+_CONTROL = re.compile(r"[\x00-\x08\x0b\x0c\x0e-\x1f]")
 
 
 def _sections(text: str) -> list[str]:
@@ -90,7 +94,7 @@ def chunk_doc(
     """
     if overlap_chars >= target_chars:
         raise ValueError("overlap_chars must be smaller than target_chars")
-    body = _FRONTMATTER.sub("", doc.text).strip()
+    body = _CONTROL.sub("", _FRONTMATTER.sub("", doc.text)).strip()
     if not body:
         return []
     prefix = f"{doc.key}:{content_hash(doc.text)[:12]}"

tests/llm/test_chunk.py

@@ -61,3 +61,19 @@ class TestChunkDoc:
     def test_overlap_gte_target_raises(self):
         with pytest.raises(ValueError, match="overlap_chars must be smaller"):
             chunk_doc(_doc("some text"), target_chars=100, overlap_chars=100)
+
+
+class TestControlCharStripping:
+    def test_strips_nul_and_control_chars(self):
+        doc = _doc("Clean text\x00 with\x0c a NUL\x07 and formfeed.")
+        chunks = chunk_doc(doc)
+        joined = "".join(c.text for c in chunks)
+        assert "\x00" not in joined
+        assert "\x0c" not in joined
+        assert "\x07" not in joined
+        assert "Clean text with a NUL and formfeed." in joined
+
+    def test_keeps_tab_newline_return(self):
+        doc = _doc("line1\n\nline2\twith tab")
+        joined = "".join(c.text for c in chunk_doc(doc))
+        assert "\t" in joined and "line2" in joined