refactor(memory): extract F2 recall path (#23)

harness/tht/memory.py

@@ -137,6 +137,48 @@ REUSABLE_TYPES = frozenset({"concept_clarified"})
 MAX_PROMOTION_CANDIDATES = 5
 
 
+def recall_memories(
+    question: str,
+    *,
+    records: list[MemoryRecord],
+    decisions: list[DecisionRecord],
+    searcher,
+    embedder,
+    top: int = 5,
+) -> list[dict]:
+    """Return reviewer-visible F2 candidates in semantic-search rank order.
+
+    Memory owns filtering against the canonical registry, reusable decision types,
+    and decisions already persisted in the current (including resumed) session.
+    The CLI remains a thin adapter that supplies the vector ports and renders output.
+    """
+    excluded = decided_memory_ids(decisions)
+    hits = searcher.search(
+        embedder.embed_query(question),
+        top_n=top,
+        kinds=["memory"],
+    )
+    by_id = {record.id: record for record in records}
+    results = []
+    for hit in hits:
+        record = by_id.get(hit.ref)
+        if record is None or record.type not in REUSABLE_TYPES or record.id in excluded:
+            continue
+        results.append({
+            "id": record.id,
+            "type": record.type,
+            "subject": record.subject,
+            "detail": record.detail,
+            "rationale": record.rationale,
+            "question_context": record.question_context,
+            "tables": record.tables,
+            "concepts": record.concepts,
+            "session_id": record.session_id,
+            "score": round(hit.similarity, 4),
+        })
+    return results
+
+
 def _compute_promotions(
     session_dir: Path, manifest: SessionManifest, *,
     seqs: list[int] | None, existing: list[MemoryRecord],