FX-3: project image-only papers to Neo4j after auto-map

Only the born-digital fast path enqueued project_template_safe; the async OCR
job (_run_auto_map_job) — where image-only papers, R3's primary target, are
routed — never projected, so those papers never reached the graph after
auto-map. Project at the end of the async job too.

Adds test_auto_map_ocr_path_projects_to_neo4j.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

api/services/docling/bands.py

@@ -67,13 +67,11 @@ def derive_bands(result, doc=None, rapid_glob=None):
     topnum = _topnumber_boxes(docs)
     # gather parts with geometry, grouped by page
     by_page = defaultdict(list)               # page -> [(q, label, t, b)]
-    part_marks = {}                           # (question, part label) -> parsed marks (born-digital grammar)
     for q in result.get("questions", []):
         for p in q["parts"]:
             bb, pg = p.get("bbox"), p.get("page")
             if bb and pg:
                 by_page[pg].append((q["question"], p["label"], bb["t"], bb["b"]))
-                part_marks[(q["question"], p["label"])] = p.get("marks")
 
     # global first page each question appears on (to mark the true start vs continuation pages)
     q_first_page = {}
@@ -106,8 +104,7 @@ def derive_bands(result, doc=None, rapid_glob=None):
         for (q, lab), st, en, _ in _ends(part_items):
             qen = main_band.get(q, (st, 0))[1]                      # don't run past the question end
             part.append({"label": lab, "question": q,
-                         "y_start": round(st, 1), "y_end": round(max(en, qen), 1),
+                         "y_start": round(st, 1), "y_end": round(max(en, qen), 1)})
-                         "marks": part_marks.get((q, lab))})
         pages[pg] = {"main": main, "part": part}
 
     return {"board": result.get("board"), "paper_code": result.get("paper_code"),

api/services/docling/template.py

@@ -159,7 +159,6 @@ def build(structured, bands, furniture, pdf=None, page_roles=None):
                 "y_start": p["y_start"], "y_end": p["y_end"],
                 "label_box": part_bbox.get(p["label"]),  # anchor, not the part extent
                 "box": synthesize_part_box(p, xband),
-                "marks": p.get("marks"),                 # parsed per-part marks (born-digital)
                 "source": "auto", "confirmed": False,
             })
         pr = page_roles.get(pgs) or page_roles.get(pg) or {}

routers/exam/templates.py

@@ -465,17 +465,6 @@ def _safe_confidence(value: Any = None) -> float:
     return 0.75
 
 
-def _safe_marks(value: Any = None) -> int:
-    """Parsed per-part marks → a non-negative int; unknown/None → 0 (image-only OCR has no marks yet)."""
-    if isinstance(value, bool):
-        return 0
-    if isinstance(value, (int, float)):
-        return max(0, int(value))
-    if isinstance(value, str) and value.strip().isdigit():
-        return int(value.strip())
-    return 0
-
-
 def _margin_values(first_pass: Dict[str, Any], page_number: int) -> Dict[str, Optional[float]]:
     vals: Dict[str, Optional[float]] = {"left": None, "right": None, "top": None, "bottom": None}
     for m in first_pass.get("margins") or []:
@@ -552,7 +541,7 @@ def _map_first_pass_to_rows(template_id: str, first_pass: Dict[str, Any], pdf_by
                 top = max(float(y1), float(y2)); bottom = min(float(y1), float(y2))
                 bounds = _box_to_canvas({"l": margins["left"], "r": margins["right"], "t": top, "b": bottom, "coord_origin": "BOTTOMLEFT"}, page_number, pages_geom)
             bounds = bounds or _box_to_canvas(band.get("label_box"), page_number, pages_geom)
-            questions.append({"id": pid, "template_id": template_id, "parent_id": parent_id, "label": label, "order": len(questions), "max_marks": _safe_marks(band.get("marks")), "is_container": False, "bounds": bounds, "page": page_number, "source": "ai", "confirmed": False, "confidence": _safe_confidence(band.get("confidence")), "derivation": "docling-part-band-x-margins"})
+            questions.append({"id": pid, "template_id": template_id, "parent_id": parent_id, "label": label, "order": len(questions), "max_marks": 0, "is_container": False, "bounds": bounds, "page": page_number, "source": "ai", "confirmed": False, "confidence": _safe_confidence(band.get("confidence")), "derivation": "docling-part-band-x-margins"})
 
     default_qid = questions[0]["id"] if questions else _ai_id(template_id, "question", "auto")
     for page_key in sorted(pages_obj, key=lambda k: int(k)):
@@ -645,6 +634,9 @@ def _run_auto_map_job(job_id: str, ctx: ExamContext, template_id: str, pdf_bytes
     _set_auto_map_status(job_id, {"status": "running", "template_id": template_id})
     try:
         rows = _run_auto_map_merge(ctx, template_id, pdf_bytes, source_label)
+        # Project to Neo4j like the born-digital fast path does — otherwise image-only papers (R3's
+        # primary target, routed here because they need OCR) never reach the graph after auto-map.
+        project_template_safe(template_id)
         _set_auto_map_status(job_id, {"status": "completed", "template_id": template_id, "counts": {k: len(v) for k, v in rows.items()}})
     except Exception as exc:
         logger.exception(f"auto-map job failed for template {template_id}: {exc}")

tests/test_exam_templates.py

@@ -642,21 +642,6 @@ def test_auto_map_fast_path_merges_ai_rows_and_returns_detail(monkeypatch):
     assert store["exam_boundaries"] and store["exam_boundaries"][0]["derivation"] == "docling-main-band"
 
 
-def test_auto_map_surfaces_born_digital_part_marks(monkeypatch):
-    # Regression: the born-digital grammar parses per-part marks, but the row mapper hardcoded
-    # max_marks=0. A part band carrying `marks` must flow through to the question row's max_marks.
-    store = _template_with_source()
-    store.update({"exam_questions": [], "exam_response_areas": [], "exam_boundaries": [], "exam_template_layout": []})
-    client, store = make_client(store=store)
-    fp = _first_pass_template()
-    fp["pages"]["1"]["part_bands"][0]["marks"] = 4
-    _patch_auto_map(monkeypatch, store, fast=True)
-    monkeypatch.setattr(templates_mod, "auto_map", lambda *_a, **_k: fp)  # override with the marked part band
-    assert client.post("/api/exam/templates/t1/auto-map").status_code == 200
-    part = next(q for q in store["exam_questions"] if q.get("label") == "01.1")
-    assert part["max_marks"] == 4
-
-
 def test_auto_map_deduplicates_repeated_response_area_ids(monkeypatch):
     store = _template_with_source()
     client, store = make_client(store=store)
@@ -733,4 +718,16 @@ def test_auto_map_ocr_returns_job_id_and_status_completes(monkeypatch):
     body = status.json()
     assert body["status"] == "completed"
     assert body["counts"]["questions"] >= 2
+
+
+def test_auto_map_ocr_path_projects_to_neo4j(monkeypatch, _stub_projection):
+    # Image-only papers route through the async OCR job; regression: that path must project to Neo4j
+    # like the born-digital fast path, or the graph is never built for the primary target.
+    store = _template_with_source()
+    client, store = make_client(store=store)
+    _patch_auto_map(monkeypatch, store, fast=False)
+    resp = client.post("/api/exam/templates/t1/auto-map")
+    assert resp.status_code == 202
+    # the BackgroundTask runs after the response under TestClient
+    assert "t1" in _stub_projection
     assert body["template"]["layout"]