G1 add homework field to lessons API

Persist the paper's question-layout signals (A-level Section A/B, EITHER/OR choice
groups, marks confidence) onto exam_templates.extraction_meta so the setup UI can
show them. Data was carried by the contract but dropped at persistence.

Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01GruxHXxfdp4kZCgAMVFgvV

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 {}

database/migrations/20260724_g1_homework_field.sql

@@ -0,0 +1,14 @@
+-- G1: Homework as a first-class lesson/planner field.
+-- Apply to DEV Supabase only before promoting app/api changes.
+
+alter table if exists public.taught_lessons
+  add column if not exists homework text not null default '';
+
+alter table if exists public.planned_lessons
+  add column if not exists homework text not null default '';
+
+comment on column public.taught_lessons.homework is
+  'Teacher-authored homework for this taught lesson; surfaced in lesson/timetable views.';
+
+comment on column public.planned_lessons.homework is
+  'Default homework carried by a lesson plan and copied to a taught lesson when delivered.';

docker-compose.dev.yml

@@ -55,6 +55,8 @@ services:
       - CC_COMPOSE_SERVICE=backend-dev
       - RUN_INIT=false
       - INIT_MODE=infra
+      # P2: route exam auto-map through the spike's full recognition pipeline (extraction service)
+      - EXAM_EXTRACT_URL=${EXAM_EXTRACT_URL:-http://192.168.0.203:8899}
     ports:
       - "18000:8000"
     depends_on:

modules/database/schemas/nodes/workers/timetable.py

@@ -53,6 +53,7 @@ class PlannedLessonNode(CCBaseNode):
     subject: str
     teacher_code: str
     planning_status: str
+    homework: Optional[str] = None
     topic_code: Optional[str] = None
     topic_name: Optional[str] = None
     lesson_code: Optional[str] = None

modules/services/exam_extract.py

@@ -0,0 +1,86 @@
+"""Client for the exam extraction SERVICE (docling-exam-spike, P1).
+
+The service runs the spike's FULL recognition pipeline (textlayer → question tree → OMR/figure/table
+sidecars → structure fusion → analyse) and returns the ghost-region contract the app already consumes.
+This module is a thin HTTP client: POST the paper, poll, return the `analyse` suggestions. The app's
+auto-map merges them (coordinate-adapted, id-remapped) in routers/exam/templates.py.
+
+Config: EXAM_EXTRACT_URL (e.g. http://192.168.0.203:8899). If unset, the app keeps its thin first-pass.
+"""
+from __future__ import annotations
+
+import base64
+import os
+import time
+from typing import Any, Dict, Optional
+
+import requests
+
+from modules.logger_tool import initialise_logger
+
+logger = initialise_logger(__name__, os.getenv("LOG_LEVEL"), os.getenv("LOG_PATH"), "default", True)
+
+
+class ExtractError(RuntimeError):
+    pass
+
+
+def service_url() -> Optional[str]:
+    url = os.getenv("EXAM_EXTRACT_URL")
+    return url.rstrip("/") if url else None
+
+
+def is_enabled() -> bool:
+    return bool(service_url())
+
+
+def _get(base: str, slug: str, timeout: int = 30) -> Dict[str, Any]:
+    r = requests.get(f"{base}/api/extract/{slug}", timeout=timeout)
+    r.raise_for_status()
+    return r.json()
+
+
+def get_replica(slug: str, timeout: int = 30) -> Dict[str, Any]:
+    """The digital-replica markdown for a paper (P4): {slug, title, n_questions, total_marks, markdown,
+    questions:[{label, marks, markdown}]}. Raises ExtractError if the paper has no replica yet."""
+    base = service_url()
+    if not base:
+        raise ExtractError("EXAM_EXTRACT_URL not configured")
+    r = requests.get(f"{base}/api/replica/{slug}", timeout=timeout)
+    if r.status_code == 404:
+        raise ExtractError(f"no digital replica for {slug}")
+    r.raise_for_status()
+    return r.json()
+
+
+def extract_suggestions(slug: str, pdf_bytes: bytes, *, force: bool = False,
+                        poll_timeout: int = 1500, poll_interval: int = 5) -> Dict[str, Any]:
+    """POST the paper to the service and poll until the analyse contract is ready.
+
+    Returns the full analyse payload: {status, coordinate_space:'page_fraction', margins,
+    suggestions:{questions[], response_areas[], boundaries[]}, meta}. Raises ExtractError on
+    failure/timeout. A cold paper is ~15 min (Docling per masked page); cached papers return instantly.
+    """
+    base = service_url()
+    if not base:
+        raise ExtractError("EXAM_EXTRACT_URL not configured")
+    payload = {"slug": slug, "pdf_b64": base64.b64encode(pdf_bytes).decode(), "force": force}
+    r = requests.post(f"{base}/api/extract", json=payload, timeout=120)
+    r.raise_for_status()
+    started = r.json()
+    if not started.get("ok", True):
+        raise ExtractError(started.get("error") or "service rejected the request")
+    # cached → fetch the contract straight away; otherwise poll the running job
+    deadline = time.time() + poll_timeout
+    while True:
+        d = _get(base, slug)
+        status = d.get("status")
+        if status == "complete":
+            if not (d.get("suggestions") or {}):
+                raise ExtractError("service returned complete with no suggestions")
+            return d
+        if status == "error":
+            raise ExtractError(f"extraction failed: {d.get('error')}")
+        if time.time() >= deadline:
+            raise ExtractError(f"extraction timed out after {poll_timeout}s (slug={slug})")
+        time.sleep(poll_interval)

routers/database/tools/lesson_plans_router.py

@@ -83,6 +83,7 @@ class CreatePlanRequest(BaseModel):
     estimated_duration_minutes: Optional[int] = None
     objectives: Optional[List[Dict[str, Any]]] = None
     activities: Optional[List[Dict[str, Any]]] = None
+    homework: Optional[str] = None
     status: Optional[str] = "draft"
     tags: Optional[List[str]] = None
     topic_code: Optional[str] = None
@@ -99,6 +100,7 @@ class UpdatePlanRequest(BaseModel):
     estimated_duration_minutes: Optional[int] = None
     objectives: Optional[List[Dict[str, Any]]] = None
     activities: Optional[List[Dict[str, Any]]] = None
+    homework: Optional[str] = None
     status: Optional[str] = None
     tags: Optional[List[str]] = None
     topic_code: Optional[str] = None
@@ -249,6 +251,7 @@ async def create_plan(
         "status": body.status or "draft",
         "objectives": body.objectives or [],
         "activities": body.activities or [],
+        "homework": body.homework or "",
         "tags": body.tags or [],
     }
     for field in (
@@ -362,7 +365,7 @@ async def update_plan(
     updates: Dict[str, Any] = {}
     for field in (
         "title", "class_id", "subject", "year_group", "estimated_duration_minutes",
-        "objectives", "activities", "status", "tags", "topic_code",
+        "objectives", "activities", "homework", "status", "tags", "topic_code",
         "whiteboard_room_id", "course_id", "sequence_number",
     ):
         val = getattr(body, field)
@@ -425,7 +428,7 @@ async def deliver_plan(
 
     plan_res = (
         sb.supabase.table("planned_lessons")
-        .select("id, whiteboard_room_id")
+        .select("id, whiteboard_room_id, homework")
         .eq("id", plan_id)
         .eq("institute_id", institute_id)
         .single()
@@ -456,6 +459,19 @@ async def deliver_plan(
 
     res = sb.supabase.table("lesson_deliveries").insert(row).execute()
     delivery = (res.data or [{}])[0]
+
+    # If the plan has default homework and was delivered into a taught lesson,
+    # copy it onto the lesson so homework appears in daily/timetable views.
+    plan_homework = plan_res.data.get("homework")
+    if body.taught_lesson_id and plan_homework is not None:
+        try:
+            sb.supabase.table("taught_lessons").update({
+                "homework": plan_homework,
+                "updated_at": datetime.utcnow().isoformat(),
+            }).eq("id", body.taught_lesson_id).eq("teacher_id", user_id).execute()
+        except Exception as e:
+            logger.warning(f"Could not copy homework to taught_lesson {body.taught_lesson_id}: {e}")
+
     logger.info(f"Lesson delivery created: {delivery.get('id')} for plan {plan_id} by {user_id}")
     return delivery
 

routers/database/tools/taught_lessons_router.py

@@ -4,7 +4,7 @@ Taught Lessons Router — materialization and lesson CRUD.
 POST /materialize   — slot template × academic_periods → taught_lessons rows
 GET  /lessons       — teacher's lessons for a date range
 GET  /lessons/{id}  — single lesson detail
-PATCH /lessons/{id} — update lesson_plan, notes, status (teacher-owned)
+PATCH /lessons/{id} — update lesson_plan, homework, notes, status (teacher-owned)
 """
 import os
 from collections import defaultdict
@@ -44,6 +44,7 @@ def _require_institute(user_id: str) -> str:
 
 class UpdateLessonRequest(BaseModel):
     lesson_plan: Optional[Dict[str, Any]] = None
+    homework: Optional[str] = None
     notes: Optional[str] = None
     status: Optional[str] = None      # planned | in_progress | completed | cancelled | substituted
 
@@ -310,7 +311,7 @@ async def get_lessons(
     lessons = (
         sb.supabase.table("taught_lessons")
         .select(
-            "id,date,period_code,week_cycle,day_of_week,status,lesson_plan,notes,whiteboard_room_id,"
+            "id,date,period_code,week_cycle,day_of_week,status,lesson_plan,homework,notes,whiteboard_room_id,"
             "class_id,academic_period_id"
         )
         .eq("teacher_id", user_id)
@@ -418,13 +419,15 @@ async def update_lesson(
     body: UpdateLessonRequest,
     credentials: dict = Depends(SupabaseBearer()),
 ) -> Dict[str, Any]:
-    """Teacher updates their own lesson content: plan, notes, status."""
+    """Teacher updates their own lesson content: plan, homework, notes, status."""
     user_id = credentials.get("sub", "")
     sb = _sb()
 
     updates: Dict[str, Any] = {}
     if body.lesson_plan is not None:
         updates["lesson_plan"] = body.lesson_plan
+    if body.homework is not None:
+        updates["homework"] = body.homework
     if body.notes is not None:
         updates["notes"] = body.notes
     if body.status is not None:
@@ -508,7 +511,7 @@ async def get_student_lessons(
     lessons = (
         sb.supabase.table("taught_lessons")
         .select(
-            "id,date,period_code,week_cycle,day_of_week,status,lesson_plan,notes,whiteboard_room_id,"
+            "id,date,period_code,week_cycle,day_of_week,status,lesson_plan,homework,notes,whiteboard_room_id,"
             "class_id,academic_period_id,teacher_id"
         )
         .in_("class_id", class_ids)

routers/database/tools/timetable_builder_router.py

@@ -675,7 +675,7 @@ def _sync_taught_lessons_to_neo4j(
     """
     lessons = (
         sb.supabase.table("taught_lessons")
-        .select("id,neo4j_node_id,academic_period_id,teacher_id,date,period_code,week_cycle,day_of_week,status")
+        .select("id,neo4j_node_id,academic_period_id,teacher_id,date,period_code,week_cycle,day_of_week,status,homework")
         .eq("institute_id", institute_id)
         .execute()
         .data or []
@@ -728,6 +728,7 @@ def _sync_taught_lessons_to_neo4j(
                 SET tl.date = date($date), tl.period_code = $pcode,
                     tl.week_cycle = $wc, tl.day_of_week = $dow,
                     tl.status = $status,
+                    tl.homework = $homework,
                     tl.node_storage_path = $path
             """,
                 id=tl_id,
@@ -736,6 +737,7 @@ def _sync_taught_lessons_to_neo4j(
                 wc=lesson.get("week_cycle", ""),
                 dow=lesson.get("day_of_week", ""),
                 status=lesson.get("status", "planned"),
+                homework=lesson.get("homework") or "",
                 path=f"taught_lessons/{tl_id}",
             )
             count += 1

routers/exam/__init__.py

@@ -8,9 +8,13 @@ from fastapi import APIRouter
 
 from routers.exam.templates import router as templates_router
 from routers.exam.batches import router as batches_router
+from routers.exam.bank import router as bank_router
+from routers.exam.corpus import router as corpus_router
 
 router = APIRouter()
 router.include_router(templates_router)
 router.include_router(batches_router)
+router.include_router(bank_router)
+router.include_router(corpus_router)
 
 __all__ = ["router"]

routers/exam/bank.py

@@ -0,0 +1,152 @@
+"""Question bank + custom-paper assembly (/api/exam/bank, /api/exam/custom-papers) — mode 3.
+
+Build-your-own-from-the-spec. The BANK is a browsable index over the leaf questions across the templates
+the caller can see (RLS-scoped, institute-wide); a CUSTOM PAPER is a normal exam_template whose questions
+are COPIES of the selected bank questions, so it flows through the existing setup / marking / projection
+pipeline unchanged. Copy-on-assemble (v1) avoids decoupling question identity from a template; a shared-
+question model is Phase-2. Design: ~/cc/ideas/2026-07-02-mode3-question-bank-design.md.
+
+All access is as-the-user (E1): the bank query returns only questions in templates RLS lets the caller see,
+and a custom paper is created owned by the caller + institute-scoped (writes pass the same RLS with-check).
+"""
+from __future__ import annotations
+
+import os
+import uuid
+from typing import Any, Dict, List, Optional
+
+from fastapi import APIRouter, Depends, HTTPException
+from pydantic import BaseModel
+
+from modules.database.services.exam_projection import project_template_safe
+from modules.logger_tool import initialise_logger
+from routers.exam.dependencies import ExamContext, get_exam_context
+
+logger = initialise_logger(__name__, os.getenv("LOG_LEVEL"), os.getenv("LOG_PATH"), "default", True)
+
+router = APIRouter()
+
+
+def _rows(result: Any) -> List[Dict[str, Any]]:
+    data = getattr(result, "data", None)
+    if not data:
+        return []
+    return data if isinstance(data, list) else [data]
+
+
+def _first(result: Any) -> Optional[Dict[str, Any]]:
+    rows = _rows(result)
+    return rows[0] if rows else None
+
+
+@router.get("/bank")
+async def list_bank(
+    spec_ref: Optional[str] = None,
+    subject: Optional[str] = None,
+    ctx: ExamContext = Depends(get_exam_context),
+) -> Dict[str, Any]:
+    """Leaf questions across the caller's institute templates, with their source-paper context.
+
+    Filter by `spec_ref` (spec point) and/or `subject`. Facets over the full visible set drive the UI
+    filters (so filtering by one axis doesn't hide the others' options). This is the spec-planning surface:
+    a teacher picks a spec point → gets every question tagged it across their papers → assembles a paper.
+    """
+    sel = (
+        "id, template_id, label, max_marks, answer_type, spec_ref, bounds, page, "
+        "exam_templates(id, title, subject, exam_code)"
+    )
+    query = ctx.supabase.table("exam_questions").select(sel).eq("is_container", False)
+    if spec_ref:
+        query = query.eq("spec_ref", spec_ref)
+    rows = _rows(query.execute())
+
+    facets_spec: Dict[str, int] = {}
+    facets_subject: Dict[str, int] = {}
+    items: List[Dict[str, Any]] = []
+    for r in rows:
+        tmpl = r.get("exam_templates") or {}
+        subj = tmpl.get("subject")
+        if r.get("spec_ref"):
+            facets_spec[r["spec_ref"]] = facets_spec.get(r["spec_ref"], 0) + 1
+        if subj:
+            facets_subject[subj] = facets_subject.get(subj, 0) + 1
+        if subject and subj != subject:
+            continue
+        items.append({
+            "id": r["id"], "template_id": r["template_id"], "label": r.get("label"),
+            "max_marks": r.get("max_marks"), "answer_type": r.get("answer_type"),
+            "spec_ref": r.get("spec_ref"), "bounds": r.get("bounds"), "page": r.get("page"),
+            "paper": {"id": tmpl.get("id"), "title": tmpl.get("title"),
+                      "subject": subj, "exam_code": tmpl.get("exam_code")},
+        })
+    return {"questions": items, "n": len(items),
+            "facets": {"spec_ref": facets_spec, "subject": facets_subject}}
+
+
+class CustomPaperRequest(BaseModel):
+    title: str
+    subject: Optional[str] = None
+    institute_id: Optional[str] = None
+    question_ids: List[str]
+
+
+@router.post("/custom-papers")
+async def create_custom_paper(
+    body: CustomPaperRequest,
+    ctx: ExamContext = Depends(get_exam_context),
+) -> Dict[str, Any]:
+    """Assemble the selected bank questions into a new template (copy-on-assemble), then project it."""
+    if not body.question_ids:
+        raise HTTPException(status_code=400, detail="question_ids is required")
+    institute_id = ctx.resolve_institute(body.institute_id)
+
+    # RLS makes this return only leaf questions the caller may see; unknown/forbidden ids are silently dropped.
+    srcs = _rows(
+        ctx.supabase.table("exam_questions").select("*")
+        .in_("id", body.question_ids).eq("is_container", False).execute()
+    )
+    by_id = {s["id"]: s for s in srcs}
+    ordered = [by_id[qid] for qid in body.question_ids if qid in by_id]  # preserve the caller's order
+    if not ordered:
+        raise HTTPException(status_code=404, detail="No accessible questions for the given ids")
+
+    template_id = str(uuid.uuid4())
+    ctx.supabase.table("exam_templates").insert({
+        "id": template_id, "title": body.title, "subject": body.subject,
+        "institute_id": institute_id, "teacher_id": ctx.user_id, "status": "draft",
+    }).execute()
+
+    id_map: Dict[str, str] = {}
+    q_rows: List[Dict[str, Any]] = []
+    for order, s in enumerate(ordered):
+        new_id = str(uuid.uuid4())
+        id_map[s["id"]] = new_id
+        q_rows.append({
+            "id": new_id, "template_id": template_id, "parent_id": None,
+            "label": s.get("label"), "order": order, "max_marks": s.get("max_marks") or 0,
+            "answer_type": s.get("answer_type"), "mcq_options": s.get("mcq_options"),
+            "mark_scheme": s.get("mark_scheme") or {}, "is_container": False,
+            "spec_ref": s.get("spec_ref"), "bounds": s.get("bounds"), "page": s.get("page"),
+            "source": "manual", "confirmed": True,
+        })
+    ctx.supabase.table("exam_questions").insert(q_rows).execute()
+
+    # Copy each selected question's response areas onto its new copy (geometry travels with the question).
+    ra_rows: List[Dict[str, Any]] = []
+    if id_map:
+        for ra in _rows(
+            ctx.supabase.table("exam_response_areas").select("*")
+            .in_("question_id", list(id_map.keys())).execute()
+        ):
+            ra_rows.append({
+                "id": str(uuid.uuid4()), "question_id": id_map[ra["question_id"]], "template_id": template_id,
+                "page": ra.get("page"), "bounds": ra.get("bounds"), "kind": ra.get("kind"),
+                "response_form": ra.get("response_form"), "context_type": ra.get("context_type"),
+                "source": "manual", "confirmed": True,
+            })
+        if ra_rows:
+            ctx.supabase.table("exam_response_areas").insert(ra_rows).execute()
+
+    project_template_safe(template_id)  # a custom paper is a normal paper in the graph
+    return {"id": template_id, "title": body.title, "subject": body.subject,
+            "n_questions": len(q_rows), "n_response_areas": len(ra_rows)}

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/corpus.py

@@ -0,0 +1,98 @@
+"""Exam-bank corpus coverage (/api/exam/corpus) — the state of the collected exam bank.
+
+Read-only view over the seeded exam-board catalogue (eb_specifications + eb_exams): board → subject →
+specification → papers, with per-session QP/MS/ER coverage and rollup counts. Shows what the app has
+COLLECTED (question papers, mark schemes, examiner reports) so a teacher/admin can see the bank's state
+and where science coverage is complete vs thin. Catalogue data is public reference — read as-the-user.
+"""
+from __future__ import annotations
+
+import os
+import re
+from typing import Any, Dict, List, Optional
+
+from fastapi import APIRouter, Depends
+
+from modules.logger_tool import initialise_logger
+from routers.exam.dependencies import ExamContext, get_exam_context
+
+logger = initialise_logger(__name__, os.getenv("LOG_LEVEL"), os.getenv("LOG_PATH"), "default", True)
+
+router = APIRouter()
+
+DOC_TYPES = ("QP", "MS", "ER")
+
+
+def _rows(result: Any) -> List[Dict[str, Any]]:
+    data = getattr(result, "data", None)
+    if not data:
+        return []
+    return data if isinstance(data, list) else [data]
+
+
+def _award_level(spec: Dict[str, Any]) -> str:
+    """Best-effort GCSE / AS / A-level from the spec code (AQA GCSE = 8xxx, A-level/AS = 7xxx)."""
+    code = re.sub(r"\D", "", spec.get("award_code") or spec.get("spec_code") or "")
+    if code.startswith("8"):
+        return "GCSE"
+    if code.startswith("7"):
+        return "A-level"
+    return spec.get("award_level") or "Other"
+
+
+@router.get("/corpus")
+async def corpus_coverage(ctx: ExamContext = Depends(get_exam_context)) -> Dict[str, Any]:
+    specs = _rows(
+        ctx.supabase.table("eb_specifications")
+        .select("spec_code, exam_board_code, subject_code, award_code, first_teach").execute()
+    )
+    exams = _rows(
+        ctx.supabase.table("eb_exams")
+        .select("exam_code, spec_code, paper_code, tier, session, type_code, storage_loc").execute()
+    )
+
+    # group exam docs → per spec → per paper (paper_code + session) → which doc types are present
+    by_spec: Dict[str, Dict[str, Dict[str, Any]]] = {}
+    for e in exams:
+        sc = e.get("spec_code")
+        if not sc:
+            continue
+        key = f"{e.get('paper_code') or '?'}|{e.get('session') or '?'}"
+        paper = by_spec.setdefault(sc, {}).setdefault(key, {
+            "paper_code": e.get("paper_code"), "session": e.get("session"),
+            "tier": e.get("tier"), "docs": {}, "exam_codes": {},
+        })
+        dt = (e.get("type_code") or "").upper()
+        if dt in DOC_TYPES:
+            paper["docs"][dt] = bool(e.get("storage_loc"))
+            paper["exam_codes"][dt] = e.get("exam_code")
+
+    totals = {"specs": 0, "papers": 0, "sessions": 0, **{d: 0 for d in DOC_TYPES}}
+    boards: Dict[str, Dict[str, Any]] = {}
+    for s in specs:
+        sc = s["spec_code"]
+        papers_map = by_spec.get(sc, {})
+        if not papers_map:
+            continue
+        totals["specs"] += 1
+        board = s.get("exam_board_code") or "?"
+        level = _award_level(s)
+        papers = sorted(papers_map.values(), key=lambda p: (str(p["session"]), str(p["paper_code"])))
+        counts = {d: sum(1 for p in papers if p["docs"].get(d)) for d in DOC_TYPES}
+        for d in DOC_TYPES:
+            totals[d] += counts[d]
+        totals["papers"] += len(papers)
+        totals["sessions"] += len({p["session"] for p in papers})
+        spec_entry = {
+            "spec_code": sc, "subject": (s.get("subject_code") or "").title(), "level": level,
+            "board": board, "first_teach": s.get("first_teach"),
+            "n_papers": len(papers), "counts": counts, "papers": papers,
+        }
+        boards.setdefault(board, {"board": board, "specs": []})["specs"].append(spec_entry)
+
+    board_list = []
+    for board in sorted(boards):
+        specs_sorted = sorted(boards[board]["specs"], key=lambda x: (x["level"], x["subject"], x["spec_code"]))
+        board_list.append({"board": board, "n_specs": len(specs_sorted), "specs": specs_sorted})
+
+    return {"totals": totals, "boards": board_list}

routers/exam/schemas.py

@@ -80,6 +80,9 @@ class ResponseAreaPayload(BaseModel):
     ] = None
     # Optional Context differentiation (v1 generic; future graph/chart/data_table/diagram/code_block/passage).
     context_type: Optional[str] = None
+    # Rich recognition payload (75-exam-marker-region-meta.sql): figure name/description, OMR geometry, unit…
+    # Carried on canvas save so a named context figure survives a round-trip.
+    meta: Optional[Dict[str, Any]] = None
     source: Literal["manual", "ai"] = "manual"
     confirmed: bool = True
     confidence: Optional[float] = Field(default=None, ge=0, le=1)

routers/exam/templates.py

@@ -13,7 +13,9 @@ join keys (spec §2).
 from __future__ import annotations
 
 import json
+import math
 import os
+import re
 import tempfile
 import time
 import uuid
@@ -29,6 +31,7 @@ from modules.database.services.exam_projection import project_template, project_
 from modules.database.supabase.utils.client import SupabaseServiceRoleClient
 from modules.database.supabase.utils.storage import StorageAdmin
 from modules.upload_validation import read_pdf_upload_bytes
+from modules.services import exam_extract
 from modules.logger_tool import initialise_logger
 from routers.exam.dependencies import ExamContext, get_exam_context, lookup_exam_code
 from routers.exam.schemas import (
@@ -342,6 +345,13 @@ def _pdf_has_text_layer(pdf_bytes: bytes) -> bool:
             pass
 
 
+# Canvas page width the frontend renders each PDF page at (app src/utils/exam-canvas/model.ts
+# PAGE_WIDTH). All auto-map canvas coords are emitted in this 780-wide, proportional-height space.
+CANVAS_PAGE_WIDTH = 780.0
+# Response/answer-region detector (api/services/docling/regions.py) renders at 144 DPI = 2 px / PDF point.
+REGIONS_PX_PER_PT = 2.0
+
+
 def _pdf_page_geometry(pdf_bytes: bytes) -> List[Dict[str, float]]:
     with tempfile.NamedTemporaryFile(prefix="cc-auto-map-geom-", suffix=".pdf", delete=False) as fh:
         fh.write(pdf_bytes)
@@ -355,14 +365,23 @@ def _pdf_page_geometry(pdf_bytes: bytes) -> List[Dict[str, float]]:
             for page in doc:
                 media = page.mediabox
                 crop = page.cropbox
-                rendered_w = float(crop.width or page.rect.width or 595.0)
-                rendered_h = float(crop.height or page.rect.height or 842.0)
+                page_pt_w = float(crop.width or page.rect.width or 1.0)
+                page_pt_h = float(crop.height or page.rect.height or 1.0)
+                # Emit canvas coords in the FRONTEND render space: the app draws each page at
+                # CANVAS_PAGE_WIDTH (app model.ts PAGE_WIDTH=780) with proportional height and stacks
+                # pages by those heights. Previously rendered_w/h were left in PDF points (~595x842),
+                # so every shape landed shrunk (~0.76x) and shifted up-left on the 780-wide canvas.
+                rendered_w = CANVAS_PAGE_WIDTH
+                # Mirror the app's canvas.height = Math.ceil(viewport.height) EXACTLY (pdfLoader.ts),
+                # so page_top accumulates identically. Using the raw float drifts ~1px/page, compounding
+                # to a visible upward shift on later pages of long papers (~36px over 40 pages).
+                rendered_h = float(math.ceil(CANVAS_PAGE_WIDTH * page_pt_h / page_pt_w))
                 pages.append({
                     "media_x0": float(media.x0),
                     "crop_x0": float(crop.x0),
                     "crop_y0": float(crop.y0),
-                    "page_pt_w": float(crop.width or page.rect.width or 1),
-                    "page_pt_h": float(crop.height or page.rect.height or 1),
+                    "page_pt_w": page_pt_w,
+                    "page_pt_h": page_pt_h,
                     "rendered_w": rendered_w,
                     "rendered_h": rendered_h,
                     "page_top": page_top,
@@ -384,11 +403,12 @@ def _pdf_page_geometry(pdf_bytes: bytes) -> List[Dict[str, float]]:
 def _page_geom(pages: List[Dict[str, float]], page_number: int) -> Dict[str, float]:
     if 1 <= page_number <= len(pages):
         return pages[page_number - 1]
+    _fallback_h = float(math.ceil(CANVAS_PAGE_WIDTH * 842.0 / 595.0))
     return {
         "media_x0": 0.0, "crop_x0": 0.0, "crop_y0": 0.0,
         "page_pt_w": 595.0, "page_pt_h": 842.0,
-        "rendered_w": 595.0, "rendered_h": 842.0,
-        "page_top": (page_number - 1) * 842.0,
+        "rendered_w": CANVAS_PAGE_WIDTH, "rendered_h": _fallback_h,
+        "page_top": (page_number - 1) * _fallback_h,
     }
 
 
@@ -397,12 +417,16 @@ def _box_to_canvas(box: Optional[Dict[str, Any]], page_number: int, pages: List[
         return None
     g = _page_geom(pages, page_number)
     if box.get("coord_origin") == "TOPLEFT" and {"x", "y", "w", "h"}.issubset(box):
-        scale = 0.5 if box.get("unit") == "px" else 1.0
+        # Scale the box into the 780-wide canvas space. px boxes (opencv/gemma regions) are in
+        # rendered-image px at REGIONS_PX_PER_PT px/point; TOPLEFT point boxes are 1 px/point.
+        px_per_pt = REGIONS_PX_PER_PT if box.get("unit") == "px" else 1.0
+        sx = g["rendered_w"] / (g["page_pt_w"] * px_per_pt)
+        sy = g["rendered_h"] / (g["page_pt_h"] * px_per_pt)
         return {
-            "x": round(float(box["x"]) * scale, 2),
-            "y": round(g["page_top"] + float(box["y"]) * scale, 2),
-            "w": round(float(box["w"]) * scale, 2),
-            "h": round(float(box["h"]) * scale, 2),
+            "x": round(float(box["x"]) * sx, 2),
+            "y": round(g["page_top"] + float(box["y"]) * sy, 2),
+            "w": round(float(box["w"]) * sx, 2),
+            "h": round(float(box["h"]) * sy, 2),
         }
     if not {"l", "t", "r", "b"}.issubset(box):
         return None
@@ -433,6 +457,32 @@ def _y_to_canvas(y_value: float, page_number: int, pages: List[Dict[str, float]]
     return round(g["page_top"] + (g["page_pt_h"] - (float(y_value) - g["crop_y0"])) / g["page_pt_h"] * g["rendered_h"], 2)
 
 
+def _frac_box_to_canvas(bounds: Optional[Dict[str, Any]], page_number: int,
+                        pages: List[Dict[str, float]]) -> Optional[Dict[str, float]]:
+    """Page-fraction {x,y,w,h} (0..1 per page, from the extraction service) → 780-wide stacked canvas."""
+    if not bounds:
+        return None
+    g = _page_geom(pages, page_number)
+    try:
+        x, y, w, h = (float(bounds["x"]), float(bounds["y"]), float(bounds["w"]), float(bounds["h"]))
+    except (KeyError, TypeError, ValueError):
+        return None
+    return {
+        "x": round(x * g["rendered_w"], 2),
+        "y": round(g["page_top"] + y * g["rendered_h"], 2),
+        "w": round(w * g["rendered_w"], 2),
+        "h": round(h * g["rendered_h"], 2),
+    }
+
+
+def _frac_y_to_canvas(y_frac: Any, page_number: int, pages: List[Dict[str, float]]) -> Optional[float]:
+    g = _page_geom(pages, page_number)
+    try:
+        return round(g["page_top"] + float(y_frac) * g["rendered_h"], 2)
+    except (TypeError, ValueError):
+        return None
+
+
 def _ai_id(template_id: str, *parts: Any) -> str:
     return str(uuid.uuid5(uuid.NAMESPACE_URL, "/".join(["cc-auto-map", template_id, *[str(p) for p in parts]])))
 
@@ -443,6 +493,17 @@ 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 []:
@@ -519,7 +580,7 @@ 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)):
@@ -540,15 +601,53 @@ def _map_first_pass_to_rows(template_id: str, first_pass: Dict[str, Any], pdf_by
             response_form = _response_form_from_region_type(region.get("region_type"))
             if response_form:
                 response_areas.append({"id": _ai_id(template_id, "region", page_index, idx), "template_id": template_id, "question_id": first_part_by_page.get(page_index, default_qid), "page": page_index + 1, "bounds": bounds, "kind": "response", "response_form": response_form, "source": "ai", "confirmed": False, "confidence": _safe_confidence(region.get("confidence")), "derivation": region.get("detection_method") or "opencv-response-region"})
+    # Integrity guard: every response_area/boundary question_id must reference an inserted question
+    # (FK exam_response_areas/exam_boundaries -> exam_questions). On papers where band detection yields
+    # few/no questions but opencv/gemma still emit regions, those regions point at the synthetic
+    # default_qid which was never inserted. Ensure that fallback container question exists and reattach
+    # any orphan child rows to it, so persistence can't violate the FK.
+    qid_set = {q["id"] for q in questions}
+    orphans = [r for r in (response_areas + boundaries) if r.get("question_id") not in qid_set]
+    if orphans:
+        if default_qid not in qid_set:
+            questions.insert(0, {"id": default_qid, "template_id": template_id, "label": "Unassigned",
+                                 "order": 0, "max_marks": 0, "is_container": True, "source": "ai",
+                                 "confirmed": False, "confidence": 0.5,
+                                 "derivation": "auto-map-fallback-container"})
+            qid_set.add(default_qid)
+        for r in orphans:
+            r["question_id"] = default_qid
+
     return {"questions": questions, "response_areas": response_areas, "boundaries": boundaries, "layout": layout}
 
 
+def _dedupe_rows_by_id(rows: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    """Preserve first occurrence of stable AI row ids emitted by noisy OCR detectors."""
+    out: List[Dict[str, Any]] = []
+    seen: set[str] = set()
+    for row in rows:
+        row_id = row.get("id")
+        if row_id:
+            key = str(row_id)
+            if key in seen:
+                continue
+            seen.add(key)
+        out.append(row)
+    return out
+
+
 def _refresh_ai_rows(ctx: ExamContext, template_id: str, rows: Dict[str, List[Dict[str, Any]]]) -> None:
     sb = ctx.supabase
     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 = 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()
 
@@ -580,12 +679,153 @@ 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}")
         _set_auto_map_status(job_id, {"status": "failed", "template_id": template_id, "error": str(exc)})
 
 
+_ALLOWED_RESPONSE_FORMS = {"lines", "answer-box", "working", "diagram", "tick-boxes", "table", "blanks"}
+_ALLOWED_ANSWER_TYPES = {"written", "mcq", "short", "diagram"}
+_ALLOWED_KINDS = {"response", "context", "question_number", "mark_area", "reference", "furniture"}
+_BOARD_RE = re.compile(r"^(aqa|edexcel|ocr|wjec|eduqas|ccea)-")
+
+
+def _extract_slug(ctx: ExamContext, template: Dict[str, Any]) -> str:
+    """A stable, board-prefixed slug for the extraction-service cache. Prefer the catalogue exam_code
+    (e.g. 'AQA-8463-1H-2022JUN-QP' → 'aqa-8463-1h-2022jun-qp' → board 'aqa' for the right margins);
+    fall back to a template-id slug (structure.py then defaults to AQA content-box margins)."""
+    code = None
+    exam_id = template.get("exam_id")
+    if exam_id:
+        try:
+            row = _first(ctx.supabase.table("eb_exams").select("exam_code").eq("id", exam_id).limit(1).execute())
+            code = (row or {}).get("exam_code")
+        except Exception as exc:
+            logger.info(f"extract slug: eb_exams lookup failed for {exam_id}: {exc}")
+    slug = re.sub(r"[^a-z0-9._-]+", "-", (code or "").lower()).strip("-")
+    if _BOARD_RE.match(slug):
+        return slug
+    return f"aqa-tmpl-{str(template.get('id') or '')[:12]}"
+
+
+def _map_service_contract_to_rows(template_id: str, contract: Dict[str, Any],
+                                  pdf_bytes: bytes) -> Dict[str, List[Dict[str, Any]]]:
+    """Map the extraction service's page-fraction analyse contract onto the app's canvas-space ghost rows.
+
+    Coordinates: page-fraction → the 780-wide stacked canvas. IDs: the service's deterministic uuid5s are
+    re-namespaced per template via _ai_id so two templates of the same paper don't collide and a re-map of
+    the same template re-emits stable ids (so _refresh_ai_rows preserves confirmed ghosts). FK-safe: parts
+    whose parent/owner question is absent are de-parented / dropped rather than crashing the insert.
+    """
+    pages = _pdf_page_geometry(pdf_bytes)
+    sug = contract.get("suggestions") or {}
+
+    def qid(uid: Any) -> str:
+        return _ai_id(template_id, "svc-q", uid)
+
+    questions: List[Dict[str, Any]] = []
+    q_ids: set = set()
+    for q in sug.get("questions") or []:
+        uid = q.get("uid")
+        if not uid:
+            continue
+        rid = qid(uid)
+        q_ids.add(rid)
+        at = q.get("answer_type") if q.get("answer_type") in _ALLOWED_ANSWER_TYPES else None
+        questions.append({
+            "id": rid, "template_id": template_id,
+            "parent_id": qid(q["parent_uid"]) if q.get("parent_uid") else None,
+            "label": q.get("label") or "?", "order": q.get("order", len(questions)),
+            "max_marks": _safe_marks(q.get("max_marks")), "answer_type": at,
+            "is_container": bool(q.get("is_container")),
+            "bounds": _frac_box_to_canvas(q.get("bounds"), q.get("page") or 1, pages),
+            "page": q.get("page"), "source": "ai", "confirmed": False,
+            "confidence": _safe_confidence(q.get("confidence")), "derivation": "extract-service",
+            # analyse contract v2 (migration 78): command verb + stem prose per part
+            "command_word": (q.get("command_word") or None) if not q.get("is_container") else None,
+            "preamble": q.get("preamble") or None,
+        })
+    for q in questions:                                 # FK safety: de-parent a dangling parent_id
+        if q["parent_id"] and q["parent_id"] not in q_ids:
+            q["parent_id"] = None
+
+    response_areas: List[Dict[str, Any]] = []
+    for ra in sug.get("response_areas") or []:
+        uid = ra.get("uid")
+        quid = qid(ra.get("question_uid") or "")
+        if not uid or quid not in q_ids:               # orphan region → drop (FK safety)
+            continue
+        bounds = _frac_box_to_canvas(ra.get("bounds"), ra.get("page") or 1, pages)
+        if not bounds:
+            continue
+        kind = ra.get("kind") if ra.get("kind") in _ALLOWED_KINDS else "response"
+        form = ra.get("response_form") if ra.get("response_form") in _ALLOWED_RESPONSE_FORMS else None
+        response_areas.append({
+            "id": _ai_id(template_id, "svc-ra", uid), "template_id": template_id,
+            "question_id": quid, "page": ra.get("page"), "bounds": bounds,
+            "kind": kind, "response_form": form if kind == "response" else None,
+            "context_type": ra.get("context_type"), "meta": ra.get("meta") or {},
+            "source": "ai", "confirmed": False,
+            "confidence": _safe_confidence(ra.get("confidence")), "derivation": "extract-service",
+        })
+
+    boundaries: List[Dict[str, Any]] = []
+    for i, b in enumerate(sug.get("boundaries") or []):
+        quid = qid(b.get("question_uid") or "")
+        if quid not in q_ids:
+            continue
+        page_index = b.get("page_index")
+        y = _frac_y_to_canvas(b.get("y"), (page_index or 0) + 1, pages)
+        if y is None:
+            continue
+        boundaries.append({
+            "id": _ai_id(template_id, "svc-b", b.get("question_uid") or i), "template_id": template_id,
+            "question_id": quid, "label": b.get("label") or "", "page_index": page_index,
+            "y": y, "bounds": None, "source": "ai", "confirmed": False,
+            "confidence": _safe_confidence(b.get("confidence")), "derivation": "extract-service",
+        })
+
+    return {"questions": questions, "response_areas": response_areas, "boundaries": boundaries}
+
+
+def _run_service_extract_merge(ctx: ExamContext, template_id: str, pdf_bytes: bytes, slug: str) -> Dict[str, List[Dict[str, Any]]]:
+    contract = exam_extract.extract_suggestions(slug, pdf_bytes)
+    rows = _map_service_contract_to_rows(template_id, contract, pdf_bytes)
+    _refresh_ai_rows(ctx, template_id, rows)
+    meta = contract.get("meta") or {}
+    # P3: record provenance + the audit cover-reconciliation gate so the setup UI can flag under-reads,
+    # and store the slug so the digital-text view can resolve this paper's replica.
+    updates: Dict[str, Any] = {"extraction_meta": {
+        "engine": "extract-service", "slug": slug, "audit": meta.get("audit") or {},
+        "counts": {k: len(v) for k, v in rows.items()},
+        # question-layout signals for the setup UI: paper sections (e.g. A-level Section A/B) + EITHER/OR choices
+        "sections": meta.get("sections") or [],
+        "choice_groups": meta.get("choice_groups") or [],
+        "marks_confidence": meta.get("marks_confidence"),
+    }}
+    n_pages = meta.get("n_pages") or meta.get("pages")
+    if n_pages:
+        updates["page_count"] = n_pages
+    ctx.supabase.table("exam_templates").update(updates).eq("id", template_id).execute()
+    return rows
+
+
+def _run_service_extract_job(job_id: str, ctx: ExamContext, template_id: str, pdf_bytes: bytes, slug: str) -> None:
+    _set_auto_map_status(job_id, {"status": "running", "template_id": template_id, "engine": "extract-service", "slug": slug})
+    try:
+        rows = _run_service_extract_merge(ctx, template_id, pdf_bytes, slug)
+        project_template_safe(template_id)
+        _set_auto_map_status(job_id, {"status": "completed", "template_id": template_id,
+                                      "engine": "extract-service", "counts": {k: len(v) for k, v in rows.items()}})
+    except Exception as exc:
+        logger.exception(f"extract-service job failed for template {template_id}: {exc}")
+        _set_auto_map_status(job_id, {"status": "failed", "template_id": template_id, "engine": "extract-service", "error": str(exc)})
+
+
 # ─── templates ───────────────────────────────────────────────────────────────
 
 
@@ -763,6 +1003,14 @@ async def auto_map_template(
         raise HTTPException(status_code=409, detail="Template has recorded marks; auto-map structural refresh is blocked.")
     bucket, path, pdf_bytes = _resolve_template_source(ctx, template)
     source_label = f"{bucket}/{path}"
+    # Extraction-service path (P2): when EXAM_EXTRACT_URL is set, route auto-map through the spike's full
+    # recognition pipeline instead of the thin first-pass. Always async — a cold paper is ~15 min.
+    if exam_extract.is_enabled():
+        slug = _extract_slug(ctx, template)
+        job_id = str(uuid.uuid4())
+        _set_auto_map_status(job_id, {"status": "queued", "template_id": template_id, "engine": "extract-service", "slug": slug})
+        background_tasks.add_task(_run_service_extract_job, job_id, ctx, template_id, pdf_bytes, slug)
+        return JSONResponse(status_code=202, content={"status": "accepted", "job_id": job_id, "engine": "extract-service"})
     try:
         fast_path = _pdf_has_text_layer(pdf_bytes)
     except Exception as exc:
@@ -801,6 +1049,24 @@ async def auto_map_status(
     return body
 
 
+@router.get("/templates/{template_id}/digital-text")
+async def template_digital_text(
+    template_id: str,
+    ctx: ExamContext = Depends(get_exam_context),
+) -> Dict[str, Any]:
+    """P4: the digital-replica markdown for this template's paper (stem text, parts, marks, answer-space
+    placeholders, spec refs). Resolves the paper via the slug the extraction used."""
+    template = _fetch_template_or_404(ctx, template_id)
+    _require_source_visibility_or_404(ctx, template)
+    if not exam_extract.is_enabled():
+        raise HTTPException(status_code=503, detail="Extraction service not configured")
+    slug = (template.get("extraction_meta") or {}).get("slug") or _extract_slug(ctx, template)
+    try:
+        return exam_extract.get_replica(slug)
+    except exam_extract.ExtractError as exc:
+        raise HTTPException(status_code=404, detail=f"No digital text yet — run auto-map first ({exc})")
+
+
 @router.put("/templates/{template_id}")
 async def replace_template(
     template_id: str,
@@ -881,6 +1147,7 @@ async def replace_template(
                 "kind": ra.kind,
                 "response_form": ra.response_form,
                 "context_type": ra.context_type,   # 73: optional Context differentiation
+                "meta": ra.meta,                   # 75: rich recognition payload (name/description/OMR/…)
                 "source": ra.source,
                 "confirmed": ra.confirmed,
                 "confidence": ra.confidence,

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
-board + AQA GCSE Physics (8463) specification with its 8 top-level topic SpecPoints. Idempotent
-(CREATE DATABASE IF NOT EXISTS / CREATE CONSTRAINT IF NOT EXISTS / MERGE).
+Creates the shared, public exam database, its uniqueness constraints, and seeds the AQA exam board
++ the 6 current test specifications (GCSE & A-level Physics/Chemistry/Biology) with their top-level
+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
-sub-point breakdown (e.g. 4.1.1.1 ...) is a later data-population task (sourceable from the AQA
-spec PDF via Docling). spec_code AQA-PHYS-8463 is the standalone GCSE Physics code that matches
-"AQA Physics Paper 1H"; the eb_exams/eb_specifications seed (card S4-3) must use the same code.
+NOTE: only *top-level* topics are seeded (the granularity a teacher plans against). The full sub-point
+breakdown (e.g. 4.1.1.1 ...) is a later data-population task (sourceable from the AQA spec PDF via
+Docling). Seeding all 6 specs means a template's spec_ref finds a matching SpecPoint so
+(: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",
@@ -81,37 +114,38 @@ 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()
-        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"],
-            subj=SPEC["subject_code"], award=SPEC["award_code"], title=SPEC["title"],
-            nsp=f"{EXAM_DB}/Specification/{SPEC['spec_code']}",
-        ).consume()
 
-        # 4. spec points
-        for ref, desc in SPEC_POINTS:
-            sp_uid = _uid("SpecPoint", SPEC["spec_code"], ref)
+        # 4. each specification + its top-level spec points (idempotent MERGE)
+        for spec in SPECIFICATIONS:
+            spec_uid = _uid("Specification", spec["spec_code"])
             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:$sc}) MERGE (s)-[:HAS_SPEC_POINT]->(p)",
-                uid=sp_uid, ref=ref, desc=desc, sc=SPEC["spec_code"],
-                ebc=SPEC["exam_board_code"], nsp=f"{EXAM_DB}/SpecPoint/{SPEC['spec_code']}/{ref}",
+                "MERGE (sp:Specification {uuid_string:$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:$ebc}) MERGE (b)-[:PUBLISHES]->(sp)",
+                uid=spec_uid, sc=spec["spec_code"], ebc=spec["exam_board_code"],
+                subj=spec["subject_code"], award=spec["award_code"], title=spec["title"],
+                nsp=f"{EXAM_DB}/Specification/{spec['spec_code']}",
             ).consume()
-            result["spec_points"] += 1
+            for ref, desc in spec["topics"]:
+                sp_uid = _uid("SpecPoint", spec["spec_code"], ref)
+                s.run(
+                    "MERGE (p:SpecPoint {uuid_string:$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"],
+                    ebc=spec["exam_board_code"], nsp=f"{EXAM_DB}/SpecPoint/{spec['spec_code']}/{ref}",
+                ).consume()
+                result["spec_points"] += 1
 
         counts = s.run(
             "MATCH (b:ExamBoard) WITH count(b) AS boards "

tests/test_exam_bank.py

@@ -0,0 +1,134 @@
+"""Tests for the question-bank + custom-paper assembly router (/api/exam/bank, /custom-papers) — mode 3.
+
+Same FakeSupabase + dependency-override pattern as test_exam_templates.py; the as-user RLS itself is
+verified against the live dev DB (.94), not here.
+"""
+import pytest
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+
+import routers.exam.bank as bank_mod
+from routers.exam.bank import router
+from routers.exam.dependencies import ExamContext, get_exam_context
+
+
+TEACHER = "00000000-0000-0000-0000-000000000001"
+INST_A = "10000000-0000-0000-0000-000000000001"
+
+
+@pytest.fixture(autouse=True)
+def _stub_projection(monkeypatch):
+    monkeypatch.setattr(bank_mod, "project_template_safe", lambda tid: None)
+
+
+class FakeResult:
+    def __init__(self, data):
+        self.data = data
+
+
+class FakeQuery:
+    def __init__(self, store, table):
+        self.store = store
+        self.table = table
+        self.rows = list(store.get(table, []))
+        self._op = None
+        self._payload = None
+
+    def select(self, *_a, **_k):
+        self._op = "select"
+        return self
+
+    def insert(self, payload):
+        self._op = "insert"
+        self._payload = payload
+        return self
+
+    def eq(self, key, value):
+        self.rows = [r for r in self.rows if r.get(key) == value]
+        return self
+
+    def in_(self, key, values):
+        values = set(values)
+        self.rows = [r for r in self.rows if r.get(key) in values]
+        return self
+
+    def execute(self):
+        backing = self.store.setdefault(self.table, [])
+        if self._op == "insert":
+            payloads = self._payload if isinstance(self._payload, list) else [self._payload]
+            inserted = []
+            for p in payloads:
+                row = dict(p)
+                row.setdefault("id", f"gen-{self.table}-{len(backing)}")
+                backing.append(row)
+                inserted.append(row)
+            return FakeResult(inserted)
+        return FakeResult(self.rows)
+
+
+class FakeSupabase:
+    def __init__(self, store):
+        self.store = store
+
+    def table(self, name):
+        return FakeQuery(self.store, name)
+
+
+def make_client(store, user_id=TEACHER, institute_ids=(INST_A,)):
+    app = FastAPI()
+    app.include_router(router, prefix="/api/exam")
+    app.dependency_overrides[get_exam_context] = lambda: ExamContext(user_id, "tok", FakeSupabase(store), list(institute_ids))
+    return TestClient(app), store
+
+
+def test_bank_lists_leaf_questions_with_facets_excluding_containers():
+    store = {"exam_questions": [
+        {"id": "c1", "template_id": "t1", "label": "Q1", "is_container": True, "spec_ref": None,
+         "exam_templates": {"id": "t1", "title": "Physics 1H", "subject": "physics", "exam_code": "AQA-PHYS-8463"}},
+        {"id": "q1", "template_id": "t1", "label": "01.1", "is_container": False, "max_marks": 3, "spec_ref": "4.2",
+         "exam_templates": {"id": "t1", "title": "Physics 1H", "subject": "physics", "exam_code": "AQA-PHYS-8463"}},
+        {"id": "q2", "template_id": "t1", "label": "01.2", "is_container": False, "max_marks": 2, "spec_ref": "4.2",
+         "exam_templates": {"id": "t1", "title": "Physics 1H", "subject": "physics", "exam_code": "AQA-PHYS-8463"}},
+    ]}
+    c, _ = make_client(store)
+    body = c.get("/api/exam/bank").json()
+    labels = {q["label"] for q in body["questions"]}
+    assert labels == {"01.1", "01.2"}  # container excluded
+    assert body["facets"]["spec_ref"] == {"4.2": 2}
+    assert body["questions"][0]["paper"]["exam_code"] == "AQA-PHYS-8463"
+
+
+def test_bank_filters_by_spec_ref():
+    store = {"exam_questions": [
+        {"id": "q1", "template_id": "t1", "label": "a", "is_container": False, "spec_ref": "4.2", "exam_templates": {"subject": "physics"}},
+        {"id": "q2", "template_id": "t1", "label": "b", "is_container": False, "spec_ref": "4.5", "exam_templates": {"subject": "physics"}},
+    ]}
+    c, _ = make_client(store)
+    body = c.get("/api/exam/bank?spec_ref=4.2").json()
+    assert [q["id"] for q in body["questions"]] == ["q1"]
+
+
+def test_custom_paper_copies_selected_questions_in_order():
+    store = {"exam_questions": [
+        {"id": "q1", "template_id": "t1", "label": "01.1", "is_container": False, "max_marks": 3, "spec_ref": "4.2", "answer_type": "short"},
+        {"id": "q2", "template_id": "t2", "label": "05.1", "is_container": False, "max_marks": 6, "spec_ref": "4.5", "answer_type": "written"},
+    ]}
+    c, store = make_client(store)
+    r = c.post("/api/exam/custom-papers", json={"title": "Electricity mini-test", "subject": "physics", "question_ids": ["q2", "q1"]})
+    assert r.status_code == 200
+    body = r.json()
+    assert body["n_questions"] == 2
+    # a new template row was created, owned + institute-scoped
+    new_tpls = [t for t in store["exam_templates"] if t["id"] == body["id"]]
+    assert new_tpls and new_tpls[0]["teacher_id"] == TEACHER and new_tpls[0]["institute_id"] == INST_A
+    # two COPIES exist under the new template, order preserving the requested [q2, q1], with fresh ids
+    copies = sorted([q for q in store["exam_questions"] if q["template_id"] == body["id"]], key=lambda q: q["order"])
+    assert [q["label"] for q in copies] == ["05.1", "01.1"]
+    assert all(q["id"] not in ("q1", "q2") for q in copies)  # not the originals
+    assert copies[0]["max_marks"] == 6 and copies[0]["spec_ref"] == "4.5"
+
+
+def test_custom_paper_requires_accessible_questions():
+    c, _ = make_client({"exam_questions": []})
+    assert c.post("/api/exam/custom-papers", json={"title": "x", "question_ids": ["nope"]}).status_code == 404
+    assert c.post("/api/exam/custom-papers", json={"title": "x", "question_ids": []}).status_code == 400

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_corpus.py

@@ -0,0 +1,48 @@
+"""Test the exam-bank corpus coverage grouping (/api/exam/corpus)."""
+from fastapi import FastAPI
+from fastapi.testclient import TestClient
+from routers.exam.corpus import router
+from routers.exam.dependencies import ExamContext, get_exam_context
+
+TEACHER = "00000000-0000-0000-0000-000000000001"
+INST_A = "10000000-0000-0000-0000-000000000001"
+
+
+class FakeResult:
+    def __init__(self, data): self.data = data
+
+
+class FakeQuery:
+    def __init__(self, store, table): self.store, self.table = store, table
+    def select(self, *_a, **_k): return self
+    def execute(self): return FakeResult(list(self.store.get(self.table, [])))
+
+
+class FakeSupabase:
+    def __init__(self, store): self.store = store
+    def table(self, name): return FakeQuery(self.store, name)
+
+
+def _client(store):
+    app = FastAPI(); app.include_router(router, prefix="/api/exam")
+    app.dependency_overrides[get_exam_context] = lambda: ExamContext(TEACHER, "tok", FakeSupabase(store), [INST_A])
+    return TestClient(app)
+
+
+def test_corpus_groups_papers_by_session_with_qp_ms_er_coverage():
+    store = {
+        "eb_specifications": [{"spec_code": "AQA-PHYS-8463", "exam_board_code": "AQA", "subject_code": "PHYSICS", "award_code": "8463", "first_teach": "2016"}],
+        "eb_exams": [
+            {"exam_code": "a", "spec_code": "AQA-PHYS-8463", "paper_code": "8463/1H", "session": "2022-Jun", "type_code": "QP", "storage_loc": "cc.examboards/a.pdf"},
+            {"exam_code": "b", "spec_code": "AQA-PHYS-8463", "paper_code": "8463/1H", "session": "2022-Jun", "type_code": "MS", "storage_loc": "cc.examboards/b.pdf"},
+            {"exam_code": "c", "spec_code": "AQA-PHYS-8463", "paper_code": "8463/1H", "session": "2022-Jun", "type_code": "ER", "storage_loc": None},
+            {"exam_code": "d", "spec_code": "AQA-PHYS-8463", "paper_code": "8463/2H", "session": "2023-Jun", "type_code": "QP", "storage_loc": "cc.examboards/d.pdf"},
+        ],
+    }
+    body = _client(store).get("/api/exam/corpus").json()
+    assert body["totals"]["specs"] == 1 and body["totals"]["papers"] == 2
+    spec = body["boards"][0]["specs"][0]
+    assert spec["level"] == "GCSE" and spec["subject"] == "Physics"
+    assert spec["counts"] == {"QP": 2, "MS": 1, "ER": 0}  # ER present but not stored → not counted
+    p = next(p for p in spec["papers"] if p["paper_code"] == "8463/1H")
+    assert p["docs"] == {"QP": True, "MS": True, "ER": False}

tests/test_exam_extract_mapping.py

@@ -0,0 +1,75 @@
+"""Test the extraction-service contract → app ghost-row mapping (P2).
+
+Mocks the PDF geometry so the mapper is tested in isolation: page-fraction → 780-canvas conversion,
+per-template uuid remap, FK-safety (orphan regions / dangling parents), and rich meta passthrough.
+"""
+import routers.exam.templates as T
+
+TID = "11111111-1111-1111-1111-111111111111"
+
+# two pages, each rendered 780 wide × 1000 tall; page 2 stacked below page 1
+_GEOM = [
+    {"rendered_w": 780.0, "rendered_h": 1000.0, "page_top": 0.0, "page_pt_w": 595.0, "page_pt_h": 842.0, "crop_x0": 0.0, "crop_y0": 0.0},
+    {"rendered_w": 780.0, "rendered_h": 1000.0, "page_top": 1000.0, "page_pt_w": 595.0, "page_pt_h": 842.0, "crop_x0": 0.0, "crop_y0": 0.0},
+]
+
+_CONTRACT = {
+    "coordinate_space": "page_fraction",
+    "suggestions": {
+        "questions": [
+            {"uid": "Q1", "label": "1", "order": 0, "max_marks": 5, "is_container": True, "page": 1},
+            {"uid": "Q1a", "parent_uid": "Q1", "label": "1(a)", "order": 1, "max_marks": 3,
+             "answer_type": "short", "is_container": False, "page": 1},
+            {"uid": "Q1b", "parent_uid": "Q1", "label": "1(b)", "order": 2, "max_marks": 2,
+             "answer_type": "mcq", "is_container": False, "page": 2},
+        ],
+        "response_areas": [
+            {"uid": "RA1", "question_uid": "Q1a", "page": 1, "kind": "response", "response_form": "answer-box",
+             "confidence": 0.9, "bounds": {"x": 0.1, "y": 0.2, "w": 0.5, "h": 0.05},
+             "meta": {"unit": "m/s", "quantity": "v", "n_lines": 1}},
+            {"uid": "RA2", "question_uid": "Q1b", "page": 2, "kind": "response", "response_form": "tick-boxes",
+             "bounds": {"x": 0.2, "y": 0.3, "w": 0.4, "h": 0.2},
+             "meta": {"select_n": 2, "n_options": 5, "boxes": [{"x0": 0.2, "y0": 0.3, "fill": 0.8}]}},
+            {"uid": "CTX1", "question_uid": "Q1b", "page": 2, "kind": "context", "response_form": None,
+             "bounds": {"x": 0.1, "y": 0.1, "w": 0.6, "h": 0.15}},
+            {"uid": "ORPH", "question_uid": "GHOST", "page": 1, "kind": "response", "response_form": "lines",
+             "bounds": {"x": 0.1, "y": 0.5, "w": 0.5, "h": 0.1}},   # orphan owner → must be dropped
+        ],
+        "boundaries": [
+            {"question_uid": "Q1", "label": "1", "page_index": 0, "y": 0.18, "confidence": 0.85},
+        ],
+    },
+    "meta": {"n_pages": 2},
+}
+
+
+def test_contract_maps_to_canvas_rows(monkeypatch):
+    monkeypatch.setattr(T, "_pdf_page_geometry", lambda _b: _GEOM)
+    rows = T._map_service_contract_to_rows(TID, _CONTRACT, b"%PDF-1.4")
+
+    q = rows["questions"]
+    assert len(q) == 3
+    container = next(x for x in q if x["label"] == "1")
+    a = next(x for x in q if x["label"] == "1(a)")
+    assert container["is_container"] is True and container["parent_id"] is None
+    assert a["parent_id"] == container["id"]          # parent remapped to the container's per-template id
+    assert a["answer_type"] == "short" and a["max_marks"] == 3
+    assert all(x["source"] == "ai" and x["confirmed"] is False for x in q)
+
+    ra = rows["response_areas"]
+    assert len(ra) == 3                                # orphan (owner GHOST) dropped
+    ids = {r["id"] for r in ra}
+    assert len(ids) == 3                               # ids unique
+    box = next(r for r in ra if r["response_form"] == "answer-box")
+    # page-fraction → 780-canvas: x=0.1*780=78, y=0.2*1000=200, w=0.5*780=390, h=0.05*1000=50
+    assert box["bounds"] == {"x": 78.0, "y": 200.0, "w": 390.0, "h": 50.0}
+    assert box["meta"]["unit"] == "m/s"               # rich meta preserved
+    tick = next(r for r in ra if r["response_form"] == "tick-boxes")
+    # page 2 y stacked: 0.3*1000 + page_top(1000) = 1300
+    assert tick["bounds"]["y"] == 1300.0 and tick["meta"]["select_n"] == 2
+    ctx = next(r for r in ra if r["kind"] == "context")
+    assert ctx["response_form"] is None               # context carries no answer form
+    assert all(r["question_id"] in {x["id"] for x in q} for r in ra)   # FK-safe
+
+    b = rows["boundaries"]
+    assert len(b) == 1 and b[0]["y"] == 180.0 and b[0]["question_id"] == container["id"]

tests/test_exam_templates.py

@@ -642,6 +642,35 @@ 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)
+    _patch_auto_map(monkeypatch, store, fast=True)
+    dup = {"page_index": 0, "bbox": {"l": 50, "t": 700, "r": 100, "b": 680, "coord_origin": "BOTTOMLEFT"}, "region_type": "answer_lines", "confidence": 0.9}
+    monkeypatch.setattr(templates_mod, "detect_response_regions_from_pdf", lambda *_a, **_k: [dup, dict(dup)])
+
+    resp = client.post("/api/exam/templates/t1/auto-map")
+
+    assert resp.status_code == 200
+    response_area_ids = [r["id"] for r in store["exam_response_areas"]]
+    assert len(response_area_ids) == len(set(response_area_ids))
+
+
 def test_auto_map_preserves_manual_and_confirmed_rows_on_rerun(monkeypatch):
     store = _template_with_source()
     store.update({
@@ -660,6 +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)
@@ -704,4 +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"]