Merge fx-3-image-only-projection

batches.py upsert_mark previously never advanced a submission or batch to
'complete', and never validated an award against the question's max.

- Reject (422) an awarded_marks that exceeds the question's max_marks — only when
  a max is actually set (0/None = not-yet-scored AI/unmapped question, unvalidatable).
- _advance_completion: a submission with a mark for every markable (leaf) question
  → 'complete'; a batch whose every non-absent submission is complete → 'complete'.
  Container questions and absent students are excluded; the helper only promotes,
  never regresses, so it is safe on every upsert.

Adds test_upsert_mark_rejects_over_max and test_upsert_mark_completes_submission_and_batch.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

api/services/docling/bands.py

@@ -67,11 +67,13 @@ 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 = {}
@@ -104,7 +106,8 @@ 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

@@ -159,6 +159,7 @@ 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/batches.py

@@ -245,6 +245,36 @@ async def batch_csv(
 
 # ─── marks ───────────────────────────────────────────────────────────────────
 
+def _advance_completion(ctx: ExamContext, batch_id: str, submission_id: str) -> None:
+    """After a mark upsert, advance statuses: a submission with a mark for every markable (leaf)
+    question → complete; a batch whose every non-absent submission is complete → complete. Nothing
+    here regresses a status (only promotes to complete), so it is safe to run on every upsert."""
+    batch = _first(
+        ctx.supabase.table("marking_batches").select("id, template_id, status").eq("id", batch_id).limit(1).execute()
+    )
+    if not batch:
+        return
+    markable = {
+        q["id"] for q in _rows(
+            ctx.supabase.table("exam_questions").select("id, is_container").eq("template_id", batch["template_id"]).execute()
+        ) if not q.get("is_container")
+    }
+    if not markable:
+        return
+    marked = {
+        m["question_id"] for m in _rows(
+            ctx.supabase.table("mark_entries").select("question_id").eq("submission_id", submission_id).execute()
+        )
+    }
+    if not markable.issubset(marked):
+        return
+    ctx.supabase.table("student_submissions").update({"status": "complete"}).eq("id", submission_id).execute()
+    subs = _rows(ctx.supabase.table("student_submissions").select("status").eq("batch_id", batch_id).execute())
+    active = [s for s in subs if s.get("status") != "absent"]
+    if active and all(s.get("status") == "complete" for s in active) and batch.get("status") != "complete":
+        ctx.supabase.table("marking_batches").update({"status": "complete"}).eq("id", batch_id).execute()
+
+
 @router.put("/marks/{mark_id}")
 async def upsert_mark(
     mark_id: str,
@@ -259,6 +289,15 @@ async def upsert_mark(
     if not submission:
         raise HTTPException(status_code=404, detail="Submission not found")
 
+    # Reject an award that exceeds the question's max (only when a max is actually set; 0/None means
+    # "not scored yet" for AI/unmapped questions, so we can't validate those).
+    question = _first(
+        ctx.supabase.table("exam_questions").select("id, max_marks").eq("id", body.question_id).limit(1).execute()
+    )
+    max_marks = (question or {}).get("max_marks")
+    if isinstance(max_marks, (int, float)) and max_marks > 0 and body.awarded_marks is not None and body.awarded_marks > max_marks:
+        raise HTTPException(status_code=422, detail=f"awarded_marks {body.awarded_marks} exceeds max_marks {max_marks} for this question")
+
     row = {
         "id": mark_id,
         "submission_id": body.submission_id,
@@ -285,6 +324,9 @@ async def upsert_mark(
     if submission.get("status") in ("absent", "unmatched"):
         ctx.supabase.table("student_submissions").update({"status": "marking"}).eq("id", body.submission_id).execute()
 
+    # Promote the submission/batch to complete once every markable question has a mark.
+    _advance_completion(ctx, submission["batch_id"], body.submission_id)
+
     return upserted
 
 

routers/exam/templates.py

@@ -465,6 +465,17 @@ 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 []:
@@ -541,7 +552,7 @@ 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": 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"})
+            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"})
 
     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)):
@@ -602,7 +613,13 @@ def _refresh_ai_rows(ctx: ExamContext, template_id: str, rows: Dict[str, List[Di
     for table in ("exam_response_areas", "exam_boundaries", "exam_template_layout", "exam_questions"):
         sb.table(table).delete().eq("template_id", template_id).eq("source", "ai").eq("confirmed", False).execute()
     for table, key in (("exam_questions", "questions"), ("exam_response_areas", "response_areas"), ("exam_boundaries", "boundaries"), ("exam_template_layout", "layout")):
-        payload = _dedupe_rows_by_id(rows.get(key) or [])
+        # A row that survived the delete above is confirmed-AI or manual — the teacher's curated version.
+        # AI ids are a deterministic uuid5 of (template_id, semantic key), so a re-run re-emits the SAME id
+        # for a confirmed ghost; re-inserting it would PK-collide and fail the whole batch. Skip those ids so
+        # confirmed work is preserved and the insert is safe (fixes the re-map-after-confirm collision).
+        kept = sb.table(table).select("id").eq("template_id", template_id).execute()
+        kept_ids = {str(r["id"]) for r in (getattr(kept, "data", None) or []) if isinstance(r, dict) and r.get("id")}
+        payload = [r for r in _dedupe_rows_by_id(rows.get(key) or []) if str(r.get("id")) not in kept_ids]
         if payload:
             sb.table(table).insert(payload).execute()
 
@@ -634,6 +651,9 @@ 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}")

run/initialization/init_exam_graph.py

@@ -1,17 +1,19 @@
 """
 init_exam_graph.py — Initialise the cc.public.exams Neo4j knowledge graph.
 
-Creates the shared, public exam database, its uniqueness constraints, and seeds the AQA exam
+Creates the shared, public exam database, its uniqueness constraints, and seeds the AQA exam board
-board + AQA GCSE Physics (8463) specification with its 8 top-level topic SpecPoints. Idempotent
++ the 6 current test specifications (GCSE & A-level Physics/Chemistry/Biology) with their top-level
-(CREATE DATABASE IF NOT EXISTS / CREATE CONSTRAINT IF NOT EXISTS / MERGE).
+topic SpecPoints (44 in total). Idempotent (CREATE DATABASE IF NOT EXISTS / CREATE CONSTRAINT IF NOT
+EXISTS / MERGE).
 
 Run inside the ccapi container:
     python3 -c "from run.initialization.init_exam_graph import init; import json; print(json.dumps(init()))"
 
-NOTE: the 8 SpecPoints seeded here are the real AQA GCSE Physics *top-level* topics. The full
+NOTE: only *top-level* topics are seeded (the granularity a teacher plans against). The full sub-point
-sub-point breakdown (e.g. 4.1.1.1 ...) is a later data-population task (sourceable from the AQA
+breakdown (e.g. 4.1.1.1 ...) is a later data-population task (sourceable from the AQA spec PDF via
-spec PDF via Docling). spec_code AQA-PHYS-8463 is the standalone GCSE Physics code that matches
+Docling). Seeding all 6 specs means a template's spec_ref finds a matching SpecPoint so
-"AQA Physics Paper 1H"; the eb_exams/eb_specifications seed (card S4-3) must use the same code.
+(:Part)-[:ASSESSES]->(:SpecPoint) fires beyond GCSE Physics; spec_code (e.g. AQA-PHYS-8463) must match
+the eb_exams/eb_specifications seed (card S4-3) and the app's deriveSpecCode.
 """
 import uuid
 from typing import Dict, Any
@@ -41,6 +43,37 @@ SPEC_POINTS = [
     ("4.8", "Space physics"),
 ]
 
+# Full AQA catalogue for the current test specs (top-level topics; ref = topic number). Seeding all of
+# them means a template's spec_ref finds a matching SpecPoint so (:Part)-[:ASSESSES]->(:SpecPoint) fires
+# beyond AQA GCSE Physics. Sub-point granularity (e.g. 4.1.1.1) remains a later data-population task.
+SPECIFICATIONS = [
+    {**SPEC, "topics": SPEC_POINTS},
+    {"spec_code": "AQA-CHEM-8462", "exam_board_code": "AQA", "subject_code": "CHEM", "award_code": "GCSE",
+     "title": "AQA GCSE Chemistry (8462)", "topics": [
+        ("4.1", "Atomic structure and the periodic table"), ("4.2", "Bonding, structure, and the properties of matter"),
+        ("4.3", "Quantitative chemistry"), ("4.4", "Chemical changes"), ("4.5", "Energy changes"),
+        ("4.6", "The rate and extent of chemical change"), ("4.7", "Organic chemistry"), ("4.8", "Chemical analysis"),
+        ("4.9", "Chemistry of the atmosphere"), ("4.10", "Using resources")]},
+    {"spec_code": "AQA-BIOL-8461", "exam_board_code": "AQA", "subject_code": "BIOL", "award_code": "GCSE",
+     "title": "AQA GCSE Biology (8461)", "topics": [
+        ("4.1", "Cell biology"), ("4.2", "Organisation"), ("4.3", "Infection and response"), ("4.4", "Bioenergetics"),
+        ("4.5", "Homeostasis and response"), ("4.6", "Inheritance, variation and evolution"), ("4.7", "Ecology")]},
+    {"spec_code": "AQA-PHYS-7408", "exam_board_code": "AQA", "subject_code": "PHYS", "award_code": "A-level",
+     "title": "AQA A-level Physics (7408)", "topics": [
+        ("3.1", "Measurements and their errors"), ("3.2", "Particles and radiation"), ("3.3", "Waves"),
+        ("3.4", "Mechanics and materials"), ("3.5", "Electricity"), ("3.6", "Further mechanics and thermal physics"),
+        ("3.7", "Fields and their consequences"), ("3.8", "Nuclear physics")]},
+    {"spec_code": "AQA-CHEM-7405", "exam_board_code": "AQA", "subject_code": "CHEM", "award_code": "A-level",
+     "title": "AQA A-level Chemistry (7405)", "topics": [
+        ("3.1", "Physical chemistry"), ("3.2", "Inorganic chemistry"), ("3.3", "Organic chemistry")]},
+    {"spec_code": "AQA-BIOL-7402", "exam_board_code": "AQA", "subject_code": "BIOL", "award_code": "A-level",
+     "title": "AQA A-level Biology (7402)", "topics": [
+        ("3.1", "Biological molecules"), ("3.2", "Cells"), ("3.3", "Organisms exchange substances with their environment"),
+        ("3.4", "Genetic information, variation and relationships between organisms"),
+        ("3.5", "Energy transfers in and between organisms"), ("3.6", "Organisms respond to changes"),
+        ("3.7", "Genetics, populations, evolution and ecosystems"), ("3.8", "The control of gene expression")]},
+]
+
 CONSTRAINTS = [
     "CREATE CONSTRAINT exam_board_uid    IF NOT EXISTS FOR (n:ExamBoard)     REQUIRE n.uuid_string IS UNIQUE",
     "CREATE CONSTRAINT spec_uid          IF NOT EXISTS FOR (n:Specification) REQUIRE n.uuid_string IS UNIQUE",
@@ -81,35 +114,36 @@ def init() -> Dict[str, Any]:
             s.run(c).consume()
             result["constraints"] += 1
 
-        # 3. board + spec
+        # 3. board (once)
         board_uid = _uid("ExamBoard", BOARD["code"])
-        spec_uid = _uid("Specification", SPEC["spec_code"])
         s.run(
             "MERGE (b:ExamBoard {uuid_string:$uid}) "
             "SET b.code=$code, b.name=$name, b.node_storage_path=$nsp",
             uid=board_uid, code=BOARD["code"], name=BOARD["name"],
             nsp=f"{EXAM_DB}/ExamBoard/{BOARD['code']}",
         ).consume()
+
+        # 4. each specification + its top-level spec points (idempotent MERGE)
+        for spec in SPECIFICATIONS:
+            spec_uid = _uid("Specification", spec["spec_code"])
             s.run(
                 "MERGE (sp:Specification {uuid_string:$uid}) "
                 "SET sp.spec_code=$sc, sp.exam_board_code=$ebc, sp.subject_code=$subj, "
                 "    sp.award_code=$award, sp.title=$title, sp.node_storage_path=$nsp "
                 "WITH sp MATCH (b:ExamBoard {code:$ebc}) MERGE (b)-[:PUBLISHES]->(sp)",
-            uid=spec_uid, sc=SPEC["spec_code"], ebc=SPEC["exam_board_code"],
+                uid=spec_uid, sc=spec["spec_code"], ebc=spec["exam_board_code"],
-            subj=SPEC["subject_code"], award=SPEC["award_code"], title=SPEC["title"],
+                subj=spec["subject_code"], award=spec["award_code"], title=spec["title"],
-            nsp=f"{EXAM_DB}/Specification/{SPEC['spec_code']}",
+                nsp=f"{EXAM_DB}/Specification/{spec['spec_code']}",
             ).consume()
-
+            for ref, desc in spec["topics"]:
-        # 4. spec points
+                sp_uid = _uid("SpecPoint", spec["spec_code"], ref)
-        for ref, desc in SPEC_POINTS:
-            sp_uid = _uid("SpecPoint", SPEC["spec_code"], ref)
                 s.run(
                     "MERGE (p:SpecPoint {uuid_string:$uid}) "
                     "SET p.ref=$ref, p.description=$desc, p.spec_code=$sc, "
                     "    p.exam_board_code=$ebc, p.node_storage_path=$nsp "
                     "WITH p MATCH (s:Specification {spec_code:$sc}) MERGE (s)-[:HAS_SPEC_POINT]->(p)",
-                uid=sp_uid, ref=ref, desc=desc, sc=SPEC["spec_code"],
+                    uid=sp_uid, ref=ref, desc=desc, sc=spec["spec_code"],
-                ebc=SPEC["exam_board_code"], nsp=f"{EXAM_DB}/SpecPoint/{SPEC['spec_code']}/{ref}",
+                    ebc=spec["exam_board_code"], nsp=f"{EXAM_DB}/SpecPoint/{spec['spec_code']}/{ref}",
                 ).consume()
                 result["spec_points"] += 1
 

tests/test_exam_batches.py

@@ -249,6 +249,30 @@ def test_upsert_mark_submission_404():
     assert c.put("/api/exam/marks/mk-x", json={"submission_id": "nope", "question_id": "q1", "awarded_marks": 1}).status_code == 404
 
 
+def test_upsert_mark_rejects_over_max():
+    c = make_client(_batch_with_cohort())  # q1 max_marks = 3
+    r = c.put("/api/exam/marks/mk-over", json={"submission_id": "sub2", "question_id": "q1", "awarded_marks": 4})
+    assert r.status_code == 422
+
+
+def test_upsert_mark_completes_submission_and_batch():
+    store = base_store(
+        marking_batches=[{"id": "b1", "template_id": TPL, "institute_id": INST_A, "teacher_id": TEACHER, "status": "marking"}],
+        exam_questions=[
+            {"id": "q0", "template_id": TPL, "label": "Q1", "max_marks": 0, "order": 0, "is_container": True},
+            {"id": "q1", "template_id": TPL, "label": "01", "max_marks": 3, "order": 1, "is_container": False},
+            {"id": "q2", "template_id": TPL, "label": "02", "max_marks": 5, "order": 2, "is_container": False},
+        ],
+        student_submissions=[{"id": "sub1", "batch_id": "b1", "student_id": "s1", "status": "marking"}],
+        mark_entries=[{"id": "m1", "batch_id": "b1", "submission_id": "sub1", "question_id": "q1", "awarded_marks": 2}],
+    )
+    c = make_client(store)
+    # marking the last leaf question completes the submission (container q0 doesn't block) and the batch
+    assert c.put("/api/exam/marks/m2", json={"submission_id": "sub1", "question_id": "q2", "awarded_marks": 4}).status_code == 200
+    assert next(s for s in store["student_submissions"] if s["id"] == "sub1")["status"] == "complete"
+    assert next(b for b in store["marking_batches"] if b["id"] == "b1")["status"] == "complete"
+
+
 # ─── scans (E3 guards) ───────────────────────────────────────────────────────
 
 def _batch_store():

tests/test_exam_templates.py

@@ -642,6 +642,21 @@ 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)
@@ -674,6 +689,27 @@ def test_auto_map_preserves_manual_and_confirmed_rows_on_rerun(monkeypatch):
     assert "old-ai" not in ids
 
 
+def test_auto_map_rerun_after_confirm_does_not_pk_collide(monkeypatch):
+    # Regression: a confirmed ghost keeps its DETERMINISTIC uuid5 id, which auto-map re-emits on a
+    # re-run — the old code re-inserted that id and PK-collided. Use the real generated id (not an
+    # arbitrary one, unlike the test above) so the collision path is actually exercised.
+    store = _template_with_source()
+    store.update({"exam_questions": [], "exam_response_areas": [], "exam_boundaries": [], "exam_template_layout": []})
+    client, store = make_client(store=store)
+    _patch_auto_map(monkeypatch, store, fast=True)
+    assert client.post("/api/exam/templates/t1/auto-map").status_code == 200
+    ghost_id = next(q["id"] for q in store["exam_questions"] if q["source"] == "ai")
+    # teacher confirms that ghost (row kept, deterministic id unchanged)
+    for q in store["exam_questions"]:
+        if q["id"] == ghost_id:
+            q["confirmed"] = True
+    # re-run auto-map: must not re-insert the same id, and must preserve the teacher's confirmation
+    assert client.post("/api/exam/templates/t1/auto-map").status_code == 200
+    same = [r for r in store["exam_questions"] if r["id"] == ghost_id]
+    assert len(same) == 1, "confirmed ghost must not be duplicated (PK collision) on re-run"
+    assert same[0]["confirmed"] is True, "teacher's confirmation must survive a re-run"
+
+
 def test_auto_map_non_owner_is_403_before_download(monkeypatch):
     store = _template_with_source(owner=OTHER_TEACHER)
     client, store = make_client(user_id=TEACHER, institute_ids=(INST_A,), store=store)
@@ -718,4 +754,16 @@ 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"]