fix: surface result titles in QA context assembly

LoCoMo v2 q300 exposed a structural asymmetry: the session-date fix
lifted multi_hop to 0.57 (mem0) and 0.52 (grep) but bm-local stayed at
0.06. BM returns bullet-level matched chunks as hit text — precise,
but they strip the document context where the session date lives.
grep returns doc heads (date included) and mem0 stores whole documents,
so only BM lost temporal anchoring.

BM's search results DO return the title (which carries the date); the
provider now passes it through hit metadata and context assembly adds
it to the section header when it isn't already in the snippet. This
uses only what the provider returns — no special-casing.

The underlying product gap (BM chunks lack document-level context at
retrieval time) is queued as the first basic-memory Phase 2 change.

2 new tests; suite green, lint clean.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: Drew Cain <groksrc@gmail.com>

Author: groksrc

src/basic_memory_benchmarks/scoring/qa.py

@@ -140,7 +140,11 @@ def assemble_context(hits: list[SearchHit], max_chars: int = CONTEXT_MAX_CHARS)
             if not snippet:
                 break
         source = hit.source_doc_id or hit.source_path or "unknown"
-        sections.append(f"[Memory {rank} | source: {source}]\n{snippet}")
+        title = (hit.metadata or {}).get("title")
+        header = f"[Memory {rank} | source: {source}]"
+        if title and str(title) not in snippet:
+            header = f"[Memory {rank} | source: {source} | {title}]"
+        sections.append(f"{header}\n{snippet}")
         used += len(snippet)
         if used >= max_chars:
             break

tests/test_qa_scoring.py

@@ -350,3 +350,31 @@ def test_budget_override_flows_through_run_qa(self):
             max_context_chars=30_000,
         )
         assert answerer.prompts[0].count("z") == 30_000
+
+
+class TestHitTitleInContext:
+    def test_title_metadata_lands_in_header(self):
+        from basic_memory_benchmarks.models import SearchHit
+        from basic_memory_benchmarks.scoring.qa import assemble_context
+
+        hit = SearchHit(
+            source_doc_id="doc-a",
+            text="- **Melanie:** I hiked yesterday!",
+            score=1.0,
+            metadata={"title": "locomo-c00-s18 (3:01 pm on 20 October, 2023)"},
+        )
+        ctx = assemble_context([hit])
+        assert "| locomo-c00-s18 (3:01 pm on 20 October, 2023)]" in ctx
+
+    def test_title_skipped_when_already_in_text(self):
+        from basic_memory_benchmarks.models import SearchHit
+        from basic_memory_benchmarks.scoring.qa import assemble_context
+
+        hit = SearchHit(
+            source_doc_id="doc-a",
+            text="# Chat session at 8 May 2023\nfull body",
+            score=1.0,
+            metadata={"title": "Chat session at 8 May 2023"},
+        )
+        ctx = assemble_context([hit])
+        assert ctx.count("Chat session at 8 May 2023") == 1