Merge pull request 'fix(metadata): לא להתיישב 'completed' בכשל-חילוץ-Gemini חולף (#138)' (#261) from worktree-metadata-no-settle-on-fail into main

mcp-server/src/legal_mcp/services/precedent_library.py

@@ -291,10 +291,16 @@ async def process_pending_extractions(kind: str = "metadata", limit: int = 20) -
                 if result.get("status") == "extraction_failed":
                     await db.set_case_law_halacha_status(cid, "failed")
                 await db.clear_extraction_request(cid, kind=kind)
+            elif result.get("status") == "extraction_failed":
+                # metadata transient failure (Gemini hiccup despite full text) —
+                # do NOT settle 'completed' or the row is silently stranded with
+                # empty metadata and the drain never revisits it (#138). Revert
+                # to 'pending' (the queue timestamp is preserved) so it re-drains.
+                await db.set_case_law_metadata_status(cid, "pending")
             else:
-                # metadata — set terminal 'completed' status (also clears the
-                # request timestamp) so the UI badge settles instead of
-                # lingering on 'processing'.
+                # metadata success / no_changes / no_metadata(no text) — set
+                # terminal 'completed' (also clears the request timestamp) so the
+                # UI badge settles instead of lingering on 'processing'.
                 await db.set_case_law_metadata_status(cid, "completed")
             processed += 1
             results.append({

mcp-server/src/legal_mcp/services/precedent_metadata_extractor.py

@@ -428,7 +428,20 @@ async def extract_and_apply(
     """Convenience wrapper: extract → merge into row → return summary."""
     suggested = await extract_metadata(case_law_id)
     if not suggested:
-        return {"status": "no_metadata", "fields": []}
+        # Empty result has two very different meanings (#138): the precedent has
+        # NO text to extract from (permanent — nothing the queue can ever do), vs
+        # the Gemini call FAILED despite the row having full text (transient — a
+        # key/network/rate-limit hiccup that a retry can recover). Conflating
+        # them as 'no_metadata' let the drain settle the row to 'completed' on a
+        # transient failure, silently stranding it with empty metadata. Branch on
+        # whether text was actually present so the caller can retry the transient
+        # case and only settle the genuinely-empty one.
+        record = await db.get_case_law(case_law_id)
+        has_text = bool(((record or {}).get("full_text") or "").strip())
+        return {
+            "status": "extraction_failed" if has_text else "no_metadata",
+            "fields": [],
+        }
     result = await apply_to_record(case_law_id, suggested, overwrite_case_number=overwrite_case_number)
     if result["updated"]:
         await db.recompute_searchable(case_law_id)