diff --git a/paperforge/worker/ocr_figures.py b/paperforge/worker/ocr_figures.py
index 91157340..24ae187d 100644
--- a/paperforge/worker/ocr_figures.py
+++ b/paperforge/worker/ocr_figures.py
@@ -88,30 +88,31 @@ def build_figure_inventory(structured_blocks: list[dict]) -> dict[str, Any]:
             if raw_label in {"image", "chart", "figure_title", "figure"} or not raw_label:
                 assets.append(block)
 
-    # --- Step 1: Split legends into formal and candidate paths ---
-    formal_legends: list[dict] = []
-    candidate_legends: list[dict] = []
-    for leg in legends:
+    # --- Step 1: Classify each legend with confidence level ---
+    # All legends are included; the classification affects confidence scoring
+    legend_classification: dict[int, str] = {}  # legend index -> "formal" | "candidate" | "fallback"
+    for i, leg in enumerate(legends):
         text = leg.get("text", "")
         bbox = leg.get("bbox", [0, 0, 0, 0])
         if _extract_figure_number(text) is not None:
-            formal_legends.append(leg)
-        elif _compute_text_confidence(text, bbox, formal_legends):
-            candidate_legends.append(leg)
-
-    all_legends = formal_legends + candidate_legends
+            legend_classification[i] = "formal"
+        elif _compute_text_confidence(text, bbox, [legends[j] for j in range(i) if legend_classification.get(j) == "formal"]):
+            legend_classification[i] = "candidate"
+        else:
+            legend_classification[i] = "fallback"
 
     # --- Step 2: Multi-signal matching ---
     used_asset_indices: set[int] = set()
     matched_figures: list[dict] = []
     unmatched_legends: list[dict] = []
 
-    for legend in all_legends:
+    for i_leg, legend in enumerate(legends):
         legend_page = legend.get("page", 0)
         legend_bbox = legend.get("bbox", [0, 0, 0, 0])
         legend_text = legend.get("text", "")
         fig_num = _extract_figure_number(legend_text)
-        is_formal = legend in formal_legends
+        cls = legend_classification.get(i_leg, "fallback")
+        is_formal = cls == "formal"
 
         candidate_pages = [legend_page, legend_page + 1, legend_page - 1]
         matched_assets = []
diff --git a/paperforge/worker/ocr_rebuild.py b/paperforge/worker/ocr_rebuild.py
index c43e1ce7..f5a0b967 100644
--- a/paperforge/worker/ocr_rebuild.py
+++ b/paperforge/worker/ocr_rebuild.py
@@ -236,6 +236,11 @@ def backfill_from_result(vault: Path, key: str) -> dict:
         meta["is_backfilled"] = True
         meta["backfilled_at"] = __import__("datetime").datetime.now().isoformat()
         meta["ocr_status"] = "done"
+        # Fix assets_path from legacy images/ to structured assets/
+        if meta.get("assets_path", "").endswith("/images"):
+            meta["assets_path"] = meta["assets_path"].replace("/images", "/assets")
+        elif meta.get("assets_path", "").endswith("\\images"):
+            meta["assets_path"] = meta["assets_path"].replace("\\images", "\\assets")
         write_json(meta_path, meta)
 
         return {"backfill_status": "done", "paper_key": key}
diff --git a/paperforge/worker/ocr_render.py b/paperforge/worker/ocr_render.py
index 307170f6..c629db8e 100644
--- a/paperforge/worker/ocr_render.py
+++ b/paperforge/worker/ocr_render.py
@@ -113,8 +113,16 @@ def render_fulltext_markdown(
         block_page = block.get("page")
 
         if block_page is not None and block_page != current_page:
-            # Fill in page markers for any pages that had no renderable blocks
-            # (e.g. page 1 with only frontmatter_noise blocks)
+            # Emit objects for the page we just finished rendering
+            if current_page is not None:
+                for fig_id in figures_by_page.get(current_page, []):
+                    lines.append(f"![[render/figures/{fig_id}.md]]")
+                    lines.append("")
+                for tbl_id in tables_by_page.get(current_page, []):
+                    lines.append(f"![[render/tables/{tbl_id}.md]]")
+                    lines.append("")
+                emitted_pages.add(current_page)
+            # Fill in page markers for skipped pages (no renderable blocks)
             first_new_page = (current_page or 0) + 1
             for p in range(first_new_page, block_page):
                 lines.append(f"<!-- page {p} -->")
@@ -129,6 +137,7 @@ def render_fulltext_markdown(
             current_page = block_page
             lines.append(f"<!-- page {block_page} -->")
             lines.append("")
+            emitted_pages.add(block_page)
 
         if role == "section_heading":
             if text.strip().lower() in FRONTMATTER_NOISE: