fix(kb/ingest): block-based PDF extraction preserves paragraph structure

pymupdf's `get_text("text")` reads glyph runs in raw order, which on
multi-column Hebrew PDFs (ספר הליקויים, ספר המבחנים) interleaves column
headers, page numbers, form labels, and body paragraphs into jumbled
lines — words split to single characters on separate lines, paragraphs
lost. Switch to `get_text("blocks")` and sort each page's blocks by
(y-axis ascending, x-axis descending) so each paragraph stays intact
and the reading order is top-to-bottom, right-to-left as expected.

Side-effect: chunk content becomes readable in the browse view.

Refs Task Master #2

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

api/services/kb/ingest.py

@@ -38,11 +38,33 @@ def _coerce_date(value) -> _dt.date | None:
 
 
 def _parse_pdf(data: bytes) -> str:
+    """Extract text from a PDF, preferring block-based parsing.
+
+    Plain `get_text("text")` reads runs in raw glyph order which interleaves
+    column labels, page numbers, and body copy for multi-column Hebrew PDFs
+    (e.g. ספר הליקויים). Using `get_text("blocks")` and sorting by vertical
+    position then horizontal position keeps each paragraph intact and
+    preserves reading order top-to-bottom.
+    """
     parts: list[str] = []
     with fitz.open(stream=data, filetype="pdf") as doc:
         for page in doc:
-            parts.append(page.get_text("text"))
-    return "\n\n".join(parts).strip()
+            blocks = page.get_text("blocks") or []
+            # Filter out empty / image-only blocks and sort by (y, x). For RTL
+            # languages, sorting by descending x within the same y keeps the
+            # reading order correct (right-to-left); pymupdf's default block
+            # order already respects this but only per-column — sorting
+            # explicitly makes it robust across oddly-placed floaters.
+            textual = [
+                (b[1], -b[0], b[4])
+                for b in blocks
+                if isinstance(b, tuple) and len(b) >= 5 and isinstance(b[4], str) and b[4].strip()
+            ]
+            textual.sort()  # by y ascending, then x descending (via negation)
+            # Join paragraphs with a blank line so the chunker's header regexes
+            # still see line boundaries.
+            parts.append("\n\n".join(b[2].rstrip() for b in textual))
+    return "\n\n".join(p for p in parts if p).strip()
 
 
 def _parse_docx(data: bytes) -> str: