Update app.py

app.py

@@ -18,11 +18,10 @@ Explicitly omitted vs the heavy Space build:
   - No text-layer row injection, institution-specific splits (UCB / Navy /
     TD / First Horizon / …), or doc-wide dedupe passes.
 
-Included (data-driven, header-agnostic):
-  - HTML tables: infer modal logical column width from each table's own rows
-    (colspan-aware), pad short rows, trim trailing empty cells on over-wide
-    rows. No fixed N, no header keywords, no date/money heuristics. Does not
-    fix same-width rows with content in the wrong cell (needs bbox / PDF text).
+Included (data-driven, no institution names):
+  - HTML tables: modal logical width from rowspan-free rows (colspan-aware);
+    pad short rows; trim trailing empty cells; if last cell is empty and the
+    previous cell is only a currency token, move that token to the last column.
 
 Configure GLMOCR_API_KEY (environment variable). Optional: glmocr + gradio +
 pymupdf + pillow installed.
@@ -315,7 +314,7 @@ def _logical_row_width(entries: List[Tuple[str, int]]) -> int:
 
 
 def _cell_text_empty(full_cell: str) -> bool:
-    m = _CELL.fullmatch(full_cell.strip(), flags=re.IGNORECASE | re.DOTALL)
+    m = _CELL.fullmatch(full_cell.strip())
     if not m:
         inner = re.sub(r"<[^>]+>", " ", full_cell)
     else:
@@ -325,6 +324,47 @@ def _cell_text_empty(full_cell: str) -> bool:
     return inner == ""
 
 
+def _cell_plain_text(full_cell: str) -> str:
+    """Visible text of one td/th, no tags."""
+    m = _CELL.fullmatch(full_cell.strip())
+    if not m:
+        t = re.sub(r"<[^>]+>", " ", full_cell)
+    else:
+        t = m.group(3)
+    t = html.unescape(re.sub(r"\s+", " ", t).strip())
+    return t
+
+
+def _is_whole_cell_currency(text: str) -> bool:
+    """
+    True iff the cell is nothing but a currency-looking amount (optional $, commas, 2 decimals).
+    Excludes dates (slashes) and arbitrary prose — not keyed to column headers.
+    """
+    if not text or "/" in text:
+        return False
+    return bool(
+        re.fullmatch(
+            r"-?(?:\$|€|£)?\s*\d{1,3}(?:,\d{3})*\.\d{2}\s*",
+            text,
+        )
+        or re.fullmatch(r"-?(?:\$|€|£)?\s*\d+\.\d{2}\s*", text)
+    )
+
+
+def _realign_money_if_last_cell_empty(cells: List[str]) -> List[str]:
+    """
+    … | X | empty -> … | empty | X when X is currency-only (fixes amount parked
+    one column left of an empty trailing cell, including after width padding).
+    """
+    if len(cells) < 2:
+        return cells
+    if not _cell_text_empty(cells[-1]):
+        return cells
+    if not _is_whole_cell_currency(_cell_plain_text(cells[-2])):
+        return cells
+    return cells[:-2] + ["<td></td>", cells[-2]]
+
+
 def _infer_modal_logical_width(tr_inners: List[str]) -> int:
     """
     Modal logical column count across rows (colspan sums). On frequency ties,
@@ -355,13 +395,16 @@ def _normalize_one_tr_inner(tr_inner: str, target: int) -> str:
     w = sum(spans)
     if w < target:
         cells.extend(["<td></td>"] * (target - w))
-        return "".join(cells)
+        spans.extend([1] * (target - w))
+        w = target
     while w > target and cells:
         if spans[-1] != 1 or not _cell_text_empty(cells[-1]):
             break
         w -= spans[-1]
         cells.pop()
         spans.pop()
+    if cells and all(s == 1 for s in spans):
+        cells = _realign_money_if_last_cell_empty(cells)
     return "".join(cells)
 
 
@@ -371,8 +414,11 @@ def normalize_html_table_row_widths(md: str) -> str:
     rows (colspan-aware), then pad rows that are too narrow or strip trailing
     empty single-colspan cells from rows that are too wide.
 
-    No column names, dates, currency patterns, or fixed N — only per-table
-    statistics. Tables containing rowspan are left unchanged (unsafe to infer).
+    No column names or fixed N: width comes from per-table row statistics.
+    If the last cell is empty and the previous cell is only a currency token,
+    that amount is moved into the last column (whole-cell shape, not headers).
+    Tables with rowspan are skipped. Same-width wrong text that is not
+    currency-shaped is not altered.
     """
     if not md or "<table" not in md.lower():
         return md
@@ -664,8 +710,8 @@ def _create_gradio_demo():
     with gr.Blocks(title="GLM-OCR (simple)") as demo:
         gr.Markdown(
             "# GLM-OCR (simple)\n"
-            "Upload a PDF or image. Header and footer bands are included; "
-            "body OCR is passed through with only light markdown cleanup."
+            "Upload a PDF or image. Header and footer bands use PDF text when available. "
+            "Body markdown is produced by the vision model and may disagree with the PDF text layer."
         )
         file_in = gr.File(
             label="Upload PDF or image",