fix: footnote blocks emit after refs; frontmatter_noise renders on backmatter pages; visual heading grading; figure inner-text crop exclusion

paperforge/worker/ocr_figures.py

@@ -1501,6 +1501,21 @@ def build_figure_inventory(structured_blocks: list[dict], page_width: float = 12
         entry["bridge_block_ids"] = [str(b.get("block_id", "")) for b in local_bridges if b.get("block_id")]
         if len(matched_assets) > 1:
             entry["cluster_bbox"] = _cluster_bbox([a.get("bbox", [0, 0, 0, 0]) for a in matched_assets])
+            # Push cluster_bbox top down past figure_inner_text blocks (panel labels)
+            # to exclude (A), (B) etc. from the cropped figure image.
+            leg_page = int(legend_page or 0)
+            for ib in structured_blocks:
+                if int(ib.get("page", 0) or 0) != leg_page or ib.get("role") != "figure_inner_text":
+                    continue
+                ib_bbox = ib.get("bbox") or ib.get("block_bbox") or [0, 0, 0, 0]
+                cb = entry["cluster_bbox"]
+                if len(ib_bbox) >= 4 and len(cb) >= 4:
+                    # Horizontal overlap: inner text sits in the figure's column
+                    x_overlap = ib_bbox[0] < cb[2] and ib_bbox[2] > cb[0]
+                    # Text starts above cluster top and extends into it
+                    is_above = ib_bbox[1] < cb[1] and ib_bbox[3] > cb[1]
+                    if x_overlap and is_above:
+                        cb[1] = max(cb[1], ib_bbox[3])
         matched_figures.append(entry)
 
     for _i, asset in enumerate(assets):

paperforge/worker/ocr_render.py

@@ -381,9 +381,10 @@ def _reorder_tail_run(
     # Column-sort is maintained; ref items are grouped at the end.
     if skip_section_grouping:
         ref_roles = frozenset({"reference_heading", "reference_item", "reference_body"})
-        non_ref = [b for b in tail_blocks if b.get("role") not in ref_roles]
+        non_ref = [b for b in tail_blocks if b.get("role") not in ref_roles and b.get("role") != "footnote"]
         refs = [b for b in tail_blocks if b.get("role") in ref_roles]
-        return non_ref + refs, carried_ref, carried_backmatter
+        fnotes = [b for b in tail_blocks if b.get("role") == "footnote"]
+        return non_ref + refs + fnotes, carried_ref, carried_backmatter
 
     # Quick check: do blocks have bbox data?  If not, use FIFO fallback.
     has_geo = any(
@@ -400,6 +401,7 @@ def _reorder_tail_run(
     ref_section: dict | None = carried_ref
     ref_items: list[dict] = []
     non_tail_pass: list[dict] = []
+    footnote_blocks: list[dict] = []
     orphan_blocks: list[dict] = []
     carried_bodies: list[dict] = []
 
@@ -430,6 +432,8 @@ def _reorder_tail_run(
                     non_tail_pass.append(block)
             else:
                 non_tail_pass.append(block)
+        elif role == "footnote":
+            footnote_blocks.append(block)
         else:
             non_tail_pass.append(block)
 
@@ -505,9 +509,9 @@ def _reorder_tail_run(
 
     # Phase 5 — emit: non-tail pass (frontmatter), then backmatter
     # sections sorted by y-position (boundary-level and sub-headings
-    # emit in reading order), then reference sections.  The y-sort is
-    # applied only to backmatter_sections to avoid carrying cross-page
-    # relationship artifacts from previous pages' carried blocks.
+    # emit in reading order), then reference sections, then footnotes.  The
+    # y-sort is applied only to backmatter_sections to avoid carrying
+    # cross-page relationship artifacts from previous pages' carried blocks.
     result: list[dict] = []
     result.extend(non_tail_pass)
     result.extend(carried_bodies)
@@ -526,6 +530,7 @@ def _reorder_tail_run(
         if ref_section.get("heading"):
             result.append(ref_section["heading"])
         result.extend(ref_section["bodies"])
+    result.extend(footnote_blocks)
     result.extend(orphan_blocks)
 
     next_backmatter = carried_backmatter
@@ -1237,7 +1242,9 @@ def render_fulltext_markdown(
         if role == "structured_insert":
             pass  # render as callout below
         elif not block.get("render_default", True):
-            continue
+            bm_start = getattr(document_structure, "spread_start", None) if document_structure else None
+            if role != "frontmatter_noise" or block_page is None or not bm_start or block_page < bm_start:
+                continue
         if role in CONSUMED_FRONTMATTER_ROLES and int(block.get("page", 0) or 0) <= 2:
             continue
         _SKIPPED_BODY_ROLES = {
@@ -1251,7 +1258,11 @@ def render_fulltext_markdown(
             "figure_inner_text",
         }
         if role in _SKIPPED_BODY_ROLES:
-            continue
+            bm_start = getattr(document_structure, "spread_start", None) if document_structure else None
+            if role == "frontmatter_noise" and block_page is not None and bm_start and block_page >= bm_start:
+                pass
+            else:
+                continue
 
         block_id = block.get("block_id")
         block_key = _page_block_key(block_page, block_id)
@@ -1271,7 +1282,9 @@ def render_fulltext_markdown(
         if role == "structured_insert":
             pass  # render as callout below
         elif not block.get("render_default", True):
-            continue
+            bm_start2 = getattr(document_structure, "spread_start", None) if document_structure else None
+            if role != "frontmatter_noise" or block_page is None or not bm_start2 or block_page < bm_start2:
+                continue
         if role == "backmatter_boundary_heading" or role == "backmatter_heading":
             last_structured_insert_page = None
             last_structured_insert_bbox = None

paperforge/worker/ocr_roles.py

@@ -442,7 +442,21 @@ def _infer_heading_level(
                     if match["size_compatible"] and match["match_score"] > 0.6:
                         return candidate_role
 
-    # Legacy heuristics (fallback when no profile data)
+    # Visual signal: font_size + bold overrides text heuristics
+    ss = (block or {}).get("span_signature") or {}
+    fs = ss.get("font_size_median", 0) or ss.get("font_size", 0) or font_size
+    is_bold = ss.get("bold", False)
+    if fs > 0:
+        # Larger font (section_heading tier) → highest level
+        if fs >= 11:
+            return "section_heading"
+        # Same-as-body bold → intermediate level
+        if is_bold:
+            return "subsection_heading"
+        # Same-as-body regular → lowest level
+        return "sub_subsection_heading"
+
+    # Legacy text heuristics (fallback when no visual data)
     upper_ratio = sum(1 for c in text if c.isupper()) / max(len(text), 1)
     is_mostly_upper = upper_ratio > 0.7
     sentence_verbs = {" is ", " are ", " was ", " were ", " have ", " has ", " been "}
@@ -1211,7 +1225,7 @@ def assign_block_role(
                 confidence=0.5,
                 evidence=[f"unnumbered paragraph_title on page 1 outside title zone: {text[:60]}"],
             )
-        level = _infer_heading_level(text)
+        level = _infer_heading_level(text, block=block)
         return RoleAssignment(
             role=level,
             confidence=0.6,
@@ -1473,7 +1487,7 @@ def assign_block_role(
             bbox = block.get("block_bbox", [0, 0, 0, 0])
             in_top_80 = page_height == 0 or (len(bbox) >= 4 and bbox[1] < page_height * 0.8)
             if in_top_80:
-                heading_level = _infer_heading_level(text, font_size)
+                heading_level = _infer_heading_level(text, font_size, block=block)
                 return RoleAssignment(
                     role=heading_level,
                     confidence=0.65,