fix(embeddings): no chunk is a heading alone (#4784)

Judging live auto-inject menus (#4772) found "## Work log — 2026-08-21",
"## Findings worth carrying forward" and "Two dev→main PRs this session."
each embedded as a whole chunk. Text about nothing embeds close to
everything, so they took ranks 1–3 on vague queries and pushed real records
down.

Three ways the chunker made them, each closed:
- _split_paragraphs flushed a heading on its own when the paragraph under it
  was long. A thin lead is now carried into the paragraph that follows, and
  a hard cut never lands in the first quarter of the budget, where the
  newline it finds closes a heading.
- A heading with no body (above its ### parts) became a section of its own
  when the section before it was full. A thin section now takes the next.
- A short lead-in or a short last section stood alone. The lead-in joins
  what follows; a thin tail joins what precedes it.

A continuation piece now repeats the LAST heading before it, since one
section can now hold several. CHUNKER_VERSION 2 → 3, so the startup backfill
re-embeds; tuned dials will report their shape as changed, which is true.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/test_chunking.py

@@ -8,8 +8,11 @@ untouched), and long records lose NOTHING — every line of the body lands in
 some chunk, each chunk inside the window budget, each carrying the title as
 its topical anchor.
 """
+import re
+
 from scribe.services.embeddings import (
     _CHUNK_CHAR_BUDGET,
+    _MIN_CHUNK_CONTENT,
     chunk_document,
     embedding_text,
 )
@@ -136,6 +139,87 @@ def test_a_monster_single_paragraph_is_hard_split_not_dropped():
     assert total_words == 2000
 
 
+# --- no chunk is a heading alone (#4784) -------------------------------------
+#
+# Found by judging live auto-inject menus: "## Work log — 2026-08-21",
+# "## Findings worth carrying forward" and "Two dev→main PRs this session."
+# were each a whole chunk, and took ranks 1–3 on vague queries — text about
+# nothing embeds close to everything. Each case below is one of those shapes.
+
+_HEADING = re.compile(r"^#{1,6}\s")
+
+
+def _said(chunk: str, title: str) -> str:
+    """A chunk's text once its title prefix and heading lines are set aside."""
+    body = chunk[len(title) + 1:] if chunk.startswith(title + "\n") else chunk
+    return "\n".join(ln for ln in body.splitlines() if not _HEADING.match(ln)).strip()
+
+
+def _assert_nothing_thin_and_nothing_lost(title: str, body: str) -> list[str]:
+    chunks = chunk_document(title, body)
+    assert len(chunks) > 1, "the case must be long enough to split"
+    for chunk in chunks:
+        assert len(_said(chunk, title)) >= _MIN_CHUNK_CONTENT, (
+            f"a chunk says almost nothing: {chunk[:120]!r}")
+        assert len(chunk) <= _CHUNK_CHAR_BUDGET + len(title) + 1 + 80
+    joined = "\n".join(chunks)
+    for line in (ln.strip() for ln in body.splitlines()):
+        # A line longer than a chunk is cut across two, so only its ends can
+        # be looked for whole.
+        for part in ([line] if len(line) < 400 else [line[:60], line[-60:]]):
+            assert part in joined, f"content dropped: {part[:60]!r}"
+    return chunks
+
+
+def test_a_heading_over_one_long_paragraph_is_not_a_chunk_of_its_own():
+    """A work log whose entry is one long paragraph: the split used to flush
+    the heading as its own piece before cutting the paragraph."""
+    entry = " ".join(["The deploy was verified against the live instance."] * 40)
+    _assert_nothing_thin_and_nothing_lost(
+        "Step 6 — webhook drift-flagging",
+        "The step body.\n\n## Work log — 2026-08-16\n\n" + entry,
+    )
+
+
+def test_a_heading_with_nothing_under_it_joins_what_follows():
+    """'## Findings' directly above its '### …' parts is a section with no
+    body, and stood alone whenever the section before it was full."""
+    body = (
+        # Five paragraphs fill the section before it, so the heading cannot
+        # merge backwards and has to be carried forward.
+        "## Context\n\n" + _long_section("context", paragraphs=5)
+        + "\n\n## Findings worth carrying forward\n\n"
+        + "### One\n\n" + _long_section("one", paragraphs=5)
+        + "\n\n### Two\n\n" + _long_section("two", paragraphs=5)
+    )
+    _assert_nothing_thin_and_nothing_lost("A milestone dev-log", body)
+
+
+def test_a_lone_lead_in_line_joins_the_section_after_it():
+    _assert_nothing_thin_and_nothing_lost(
+        "2026-06-10 — UI batch",
+        "Two dev→main PRs this session.\n\n## PR 89\n\n"
+        + _long_section("pr89", paragraphs=8),
+    )
+
+
+def test_a_thin_last_section_joins_the_one_before_it():
+    _assert_nothing_thin_and_nothing_lost(
+        "T",
+        "## Body\n\n" + _long_section("body", paragraphs=8) + "\n\n## Next\n\nTBD.",
+    )
+
+
+def test_a_continuation_piece_repeats_the_heading_it_sits_under():
+    """A thin heading merged forward puts two headings in one section; the
+    pieces of the second must carry the second, not the first."""
+    body = "## Parent\n\n## Child\n\n" + _long_section("child", paragraphs=40)
+    chunks = chunk_document("T", body)
+    assert len(chunks) > 2
+    for chunk in chunks[1:]:
+        assert "## Child" in chunk
+
+
 # --- the read path: best chunk wins (#280 step 4) ----------------------------