fix(chunker): drop empty H2/H3 sections instead of chunking them

_split_long("") returned [""], so an H2/H3 heading immediately followed by
another heading (or by the end of the body) produced a real chunk with an
empty body: a wasted FTS row, trigram row and embedding call. ord stays
contiguous since the loop only increments when a chunk is actually appended.

Test: tests/test_chunker.py::test_empty_h2_sections_do_not_produce_empty_chunks

src/mdmem/chunker.py

@@ -49,7 +49,11 @@ def _sections(body: str) -> list[tuple[str, str]]:
 
 def _split_long(text: str, limit: int = MAX_CHUNK_CHARS) -> list[str]:
     if len(text) <= limit:
-        return [text]
+        # An empty/whitespace-only section (an H2/H3 immediately followed by
+        # another heading, or by the end of the body) must not become a real
+        # chunk: an empty body still gets an FTS row, a trigram row and a
+        # wasted embedding call.
+        return [text] if text.strip() else []
     parts: list[str] = []
     buf: list[str] = []
     size = 0

tests/test_chunker.py

@@ -49,6 +49,21 @@ def test_embed_text_merges_keys_at_index_time():
     assert t.endswith("Körper")
 
 
+def test_empty_h2_sections_do_not_produce_empty_chunks():
+    """An empty H2/H3 section (heading immediately followed by another heading,
+    or by the end of the body) used to become a real chunk with an empty body
+    -- a wasted FTS row, trigram row and embedding call for nothing."""
+    body = "## A\n\n## B\n\n" + "wort " * 1400 + "\n\n## C\n\n"
+    n = Note.new(title="T", body=body, type="note")
+    chunks = chunk_note(n)
+    assert all(c.body.strip() for c in chunks), [c.body for c in chunks]
+    assert [c.ord for c in chunks] == list(range(len(chunks)))
+    headings = {c.heading for c in chunks}
+    assert "A" not in headings
+    assert "C" not in headings
+    assert "B" in headings
+
+
 def test_empty_body_yields_one_chunk_with_title_text():
     n = Note.new(title="Nur Titel", body="", type="fact")
     chunks = chunk_note(n)