Update app.py

app.py

@@ -41,20 +41,24 @@ Primary goals for reconcile rate:
 - Fix 21: Cross-table DA3 repair for MSBILL / MICROSOFT#G… PIN rows — when Amount is junk in one table
             but another table repeats the same authorization id with a strict amount, copy that amount
             (then subset dedupe can drop the stray duplicate block, e.g. under Deposits).
-- Fix 22: Drop a single OCR “flat line” between pages: (continued) + DESCRIPTION + AMOUNT + many MM/DD
+- Fix 22: Drop a single OCR "flat line" between pages: (continued) + DESCRIPTION + AMOUNT + many MM/DD
             tokens (noise before a proper <table>).
 - Fix 23: Doc-wide DA3 table dedupe using multiset of (date, amount) keys — catches duplicate sections when
             OCR varies description text between copies (e.g. withdrawals repeated under Deposits).
 - Fix 24: Run orphan flatline strip on merged PDF output and after text-layer patch so page-boundary
             noise is removed consistently.
-- Fix 25: After a centered “deposits + credits/interest” heading, if two DA3 tables appear back-to-back
+- Fix 25: After a centered "deposits + credits/interest" heading, if two DA3 tables appear back-to-back
             and the first is mostly credit/deposit rows while the second is mostly debit/withdrawal-style
             rows (generic keywords), drop the second — fixes OCR pasting the withdrawals block again.
 - Fix 23b: Loose multiset keys use normalized float amounts so comma/spacing variants still match.
 - Fix 26: Drop small DA3 fragments (3–14 rows) that are a multiset subset of an earlier larger DA3
             table — matches PDF layouts that re-print a header and repeat a few rows.
-- Fix 27: Strip a lone “…(continued)” account banner line (no DATE/DESCRIPTION/AMOUNT) between
+- Fix 27: Strip a lone "…(continued)" account banner line (no DATE/DESCRIPTION/AMOUNT) between
             page separators — duplicate header noise without the long flatline.
+- Fix 28: Strip leading credit/deposit rows that OCR prepends to a withdrawals/debits DA3 table
+            when the preceding DA3 table was under a deposits/credits heading. Only strips the
+            contiguous prefix of the debit table whose rows appear in the credit table; requires
+            ≥1 data row to remain; entire function wrapped in try/except (safe no-op on any error).
 """
 
 # Patch asyncio first (before Gradio imports it) to suppress Python 3.13 cleanup noise
@@ -612,10 +616,6 @@ _HEADER_ROW_KEYWORD_THRESHOLD = 0.5
 # --------------------------
 
 def _split_keyword_remainder(cell_text: str):
-    """
-    If cell_text starts with a known header keyword followed by extra content,
-    return (keyword, remainder).  Otherwise return (cell_text, "").
-    """
     s = cell_text.strip()
     for pattern in _HEADER_KW_RES:
         m = pattern.match(s)
@@ -626,18 +626,6 @@ def _split_keyword_remainder(cell_text: str):
     return s, ""
 
 def _extract_fused_header_artifacts(header_row):
-    """
-    Generic detector for the OCR artifact where the first data row of a table
-    gets fused into the header cells during OCR.
-    The artifact pattern (BOTH signals must fire simultaneously):
-      - Signal 1 — text-fused: a header cell contains KEYWORD + free descriptive text
-        e.g.  "DESCRIPTIONBeginning Balance"  "NARRATIONOpening Balance"
-      - Signal 2 — money-fused: a DIFFERENT header cell contains KEYWORD + money amount
-        e.g.  "BALANCE$3,447.10"  "AMOUNT 5,000.00"
-    Two-signal guard prevents false positives on clean PDFs from any bank.
-    Returns:
-        (cleaned_header : list[str], recovered_row : list[str] | None)
-    """
     if not header_row:
         return list(header_row), None
 
@@ -676,10 +664,6 @@ def _extract_fused_header_artifacts(header_row):
     return list(header_row), None
 
 def _clean_header_artifacts(grid):
-    """
-    Entry point for Fix 1 called from normalize_html_tables.
-    Returns (grid, recovered_row | None).
-    """
     if not grid or not grid[0]:
         return grid, None
 
@@ -692,12 +676,6 @@ def _clean_header_artifacts(grid):
 # --------------------------
 
 def _row_keyword_score(row):
-    """
-    Return the fraction of non-empty cells in this row that are pure header
-    keywords (e.g. DATE, DESCRIPTION, DEBIT, CREDIT, BALANCE).
-    Also handles cells like "DATE DESCRIPTION" where two keywords are
-    space-joined into one cell — these count as a keyword cell too.
-    """
     non_empty = [str(c or "").strip() for c in row if str(c or "").strip()]
     if not non_empty:
         return 0.0
@@ -714,8 +692,6 @@ def _row_keyword_score(row):
 
     return keyword_hits / len(non_empty)
 
-# Pattern: header cell contains balance/account info — do NOT promote away from it.
-# Matches things like "Beginning Balance:", "Ending Balance:", account numbers, "$20.48$3.46"
 _BALANCE_INFO_RE = re.compile(
     r"(beginning|ending|opening|closing)\s+balance"
     r"|account\s*#?\s*\d{5,}"
@@ -725,39 +701,12 @@ _BALANCE_INFO_RE = re.compile(
 )
 
 def _header_contains_balance_info(header_row):
-    """
-    Return True if any cell in the header row contains balance or account
-    information — indicating this IS a legitimate header row even if its
-    keyword score is low (e.g. Citi's "208479667 | Beginning Balance:$20.48 | ...").
-    """
     for cell in header_row:
         if _BALANCE_INFO_RE.search(str(cell or "")):
             return True
     return False
 
 def _promote_misplaced_header_row(grid):
-    """
-    Detects and fixes the OCR artifact where real column headers land in a
-    <td> data row instead of the <th> header row.
-    Two scenarios handled:
-    Scenario A — Simple misplaced header (e.g. TD Bank):
-      <th>ACCOUNT ACTIVITY</th>        ← section title, not a real header
-      <td>DATE DESCRIPTION</td> ...    ← real column headers in a data row
-      <td>01/05 ...</td>               ← data
-    → Promote the keyword row to header, discard the section title rows above.
-    Scenario B — Metadata header + misplaced column headers (e.g. Citi page 1):
-      <th>208479667</th> <th>Beginning Balance:$20.48 Ending Balance:$3.46</th>
-      <td>Date Description</td> <td>Debits</td> <td>Credits</td> <td>Balance</td>
-      <td>04/01 DEBIT CARD... 8.01</td> ...
-    → The existing th row has balance/account metadata (NOT a real column header).
-    → Promote the keyword data row to header.
-    → Preserve metadata row cells as a plain-text prefix OUTSIDE the table
-      by embedding them as a leading data row with colspan (kept for reference).
-    → Split any "Date Description" fused cells in data rows.
-    Guard conditions (no-op if not met — safe for all other PDFs):
-      - Current header keyword score < threshold.
-      - A data row within the first 5 rows has keyword score >= threshold.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -765,7 +714,6 @@ def _promote_misplaced_header_row(grid):
     if current_header_score >= _HEADER_ROW_KEYWORD_THRESHOLD:
         return grid
 
-    # Search the first few data rows for a candidate keyword header row.
     candidate_idx = None
     candidate_score = 0.0
     for i in range(1, min(len(grid), 6)):
@@ -779,7 +727,6 @@ def _promote_misplaced_header_row(grid):
 
     candidate_row = grid[candidate_idx]
 
-    # Expand any merged keyword cells (e.g. "DATE DESCRIPTION" → ["DATE", "DESCRIPTION"])
     expanded_header = []
     for cell in candidate_row:
         cell_s = str(cell or "").strip()
@@ -792,9 +739,6 @@ def _promote_misplaced_header_row(grid):
     new_ncols = len(expanded_header)
     col_expansion = new_ncols - len(candidate_row)
 
-    # If the existing header row contains balance/account metadata (Scenario B),
-    # preserve it as the first data row so the information is not lost.
-    # This row will have its content in col 0 (joined) and blanks elsewhere.
     metadata_row = None
     if _header_contains_balance_info(grid[0]):
         meta_cells = [str(c or "").strip() for c in grid[0]]
@@ -802,7 +746,6 @@ def _promote_misplaced_header_row(grid):
         if meta_text:
             metadata_row = [meta_text] + [""] * (new_ncols - 1)
 
-    # Re-align data rows below the candidate header
     new_data_rows = []
     for row in grid[candidate_idx + 1:]:
         if col_expansion > 0 and row:
@@ -823,7 +766,6 @@ def _promote_misplaced_header_row(grid):
         return r[:n]
 
     new_grid = [pad(expanded_header, new_ncols)]
-    # Insert metadata row first if present (preserves Beginning/Ending Balance)
     if metadata_row is not None:
         new_grid.append(pad(metadata_row, new_ncols))
     for row in new_data_rows:
@@ -835,39 +777,16 @@ def _promote_misplaced_header_row(grid):
 # Fix 3: Fused key-value rows in summary sections
 # --------------------------
 
-# Matches a value suffix fused onto a label with no space.
-# e.g. "Beginning Interest Rate0.00%" → label + "0.00%"
-#      "Number of days in this Period31" → label + "31"
-# Strict: no \s* between groups so label trailing-space detects space-separated values
 _FUSED_KV_RE = re.compile(
     r"^(.+?)(-?\d{1,3}(?:,\d{3})*(?:\.\d+)?%?)$",
     re.DOTALL,
 )
 
 def _fix_fused_keyvalue_rows(grid):
-    """
-    Fix rows where label+value are fused into the first cell with all other
-    cells empty — common in Interest Summary / Fee Summary sections appended
-    to a transaction table by OCR.
-    e.g. "Beginning Interest Rate0.00%"  → label="Beginning Interest Rate"  value="0.00%"
-         "Number of days in this Period31" → label="..."  value="31"
-    Guard conditions (no-op if not met — safe for all other tables):
-      - All cells except first must be empty.
-      - A non-space character must immediately precede the digit run (fused).
-      - Both label and value parts must be non-empty after splitting.
-      - The cell must NOT contain an account number (7+ digit run) — prevents
-        incorrectly splitting "STREAMLINED CHECKING #208479667" or
-        "Charges debited from account #208479667".
-      - The cell must NOT contain a proper dollar amount with decimal — prevents
-        incorrectly splitting metadata like "Beginning Balance:$20.48$3.46".
-    """
     if not grid or len(grid) < 2:
         return grid
 
-    # Matches a proper dollar/currency amount with decimal point — these appear
-    # in legitimate label cells and must not trigger the fused-KV split.
     _DOLLAR_AMOUNT_RE = re.compile(r"\$\s*\d{1,3}(?:,\d{3})*\.\d{2}")
-    # Matches a long account/reference number (7+ consecutive digits)
     _LONG_NUMBER_RE = re.compile(r"\d{7,}")
 
     ncols = len(grid[0])
@@ -886,21 +805,14 @@ def _fix_fused_keyvalue_rows(grid):
             new_grid.append(row)
             continue
 
-        # Guard: skip rows containing a proper dollar amount with decimal
-        # (these are metadata/balance rows, not fused key-value summary rows)
         if _DOLLAR_AMOUNT_RE.search(first):
             new_grid.append(row)
             continue
 
-        # Guard: skip rows containing a long account/reference number (7+ digits)
-        # (e.g. "STREAMLINED CHECKING #208479667", "account #208479667")
         if _LONG_NUMBER_RE.search(first):
             new_grid.append(row)
             continue
 
-        # Guard: skip rows containing a # followed by digits or X-patterns
-        # (card/account number labels like "Card account # XXXX XXXX XXXX 2889")
-        # These are separator rows reconstructed by Fix5, not fused KV rows.
         if re.search(r"#\s*[\dX]", first):
             new_grid.append(row)
             continue
@@ -910,10 +822,9 @@ def _fix_fused_keyvalue_rows(grid):
             new_grid.append(row)
             continue
 
-        label_raw = m.group(1)   # trailing space means value was space-separated, not fused
+        label_raw = m.group(1)
         value = m.group(2).strip()
 
-        # Space before value means NOT fused — "Interest Rate 0.00%" is clean
         if label_raw != label_raw.rstrip():
             new_grid.append(row)
             continue
@@ -936,16 +847,6 @@ def _fix_fused_keyvalue_rows(grid):
 # --------------------------
 
 def _extract_textlayer_rows(pdf_path: str, page_num: int):
-    """
-    Extract structured transaction rows from the PDF text layer using pymupdf.
-    Returns list of {"date": str, "desc": str, "amount": str} dicts, or []
-    if the PDF has no text layer, pymupdf is unavailable, or fewer than 2
-    transaction rows are found (guards against non-transaction pages).
-    Supports multiple date formats:
-      - Numeric:    MM/DD, MM/DD/YY, MM/DD/YYYY  (Chase, TD, BoA, Citi)
-      - Month-name: Jan 16, Feb 7, Mar 05         (Capital One, Amex)
-      - ISO:        YYYY-MM-DD
-    """
     try:
         import pymupdf as fitz
 
@@ -957,7 +858,6 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
         if not words:
             return []
 
-        # Group words by y-bucket (5pt tolerance)
         lines_by_y = {}
         for w in words:
             x0, y0, word = float(w[0]), float(w[1]), w[4]
@@ -975,24 +875,21 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             line_words = sorted(lines_by_y[y], key=lambda t: t[0])
             sorted_lines.append([w for _, w in line_words])
 
-        # Date patterns — numeric (MM/DD, MM-DD variants) or month-name (Jan 16)
         _MONTHS = r"(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)"
         date_re = re.compile(
-            r"^\d{1,2}/\d{2}(?:/\d{2,4})?$"          # MM/DD, MM/DD/YY, MM/DD/YYYY
-            r"|^\d{1,2}-\d{2}(?:-\d{2,4})?$"           # MM-DD, MM-DD-YY (East West Bank)
-            r"|^\d{4}-\d{2}-\d{2}$"                    # YYYY-MM-DD
-            r"|^" + _MONTHS + r"$",                        # "Jan", "Feb" etc (month-name date part 1)
+            r"^\d{1,2}/\d{2}(?:/\d{2,4})?$"
+            r"|^\d{1,2}-\d{2}(?:-\d{2,4})?$"
+            r"|^\d{4}-\d{2}-\d{2}$"
+            r"|^" + _MONTHS + r"$",
             re.IGNORECASE,
         )
-        # When a month-name date appears, the next token is the day number
         month_re = re.compile(r"^" + _MONTHS + r"$", re.IGNORECASE)
         day_re   = re.compile(r"^\d{1,2}$")
 
-        # Amount: -1,234.56  $193.63  -$240.46  - $500.00 (with space after -)
         amount_re = re.compile(
             r"^-?\$?\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$"
             r"|^\$-?\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$"
-            r"|^-\s+\$\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$"  # "- $500.00"
+            r"|^-\s+\$\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$"
         )
 
         rows = []
@@ -1000,17 +897,14 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             if len(line) < 2:
                 continue
 
-            # Detect date and find where description starts
             date_str = None
             desc_start = None
 
             if date_re.match(line[0]):
                 if month_re.match(line[0]) and len(line) > 1 and day_re.match(line[1]):
-                    # Month-name date: "Jan 16" → two tokens
                     date_str = line[0] + " " + line[1]
                     desc_start = 2
                 else:
-                    # Single-token numeric date
                     date_str = line[0]
                     desc_start = 1
             else:
@@ -1019,11 +913,7 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             if desc_start is None or desc_start >= len(line):
                 continue
 
-            # For Capital One: Trans Date and Post Date are both present
-            # Line looks like: Jan 16 Jan 16 CAPITAL ONE MOBILE PYMT - $500.00
-            # After consuming first date, check if next tokens are also a date
             remaining = line[desc_start:]
-            # Capture post_date if next tokens are also a date
             post_date_str = ""
             if remaining and month_re.match(remaining[0]):
                 if len(remaining) > 1 and day_re.match(remaining[1]):
@@ -1038,7 +928,6 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             if len(remaining) < 2:
                 continue
 
-            # Handle "- $500.00" split as two tokens at end
             if (len(remaining) >= 2
                     and remaining[-2] == "-"
                     and remaining[-1].startswith("$")):
@@ -1055,16 +944,11 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
             if not desc:
                 continue
 
-            # Guard: description must contain at least one letter.
-            # Rows where description is purely digits/dates/amounts are
-            # DAILY BALANCE rows (e.g. "170,198.04 01-13 45,442.31 01-24"),
-            # not real transactions — skip them.
             if not re.search(r"[A-Za-z]", desc):
                 continue
 
             rows.append({"date": date_str, "post_date": post_date_str, "desc": desc, "amount": amount_candidate})
 
-        # Guard: require at least 2 rows to avoid false positives on non-transaction pages
         return rows if len(rows) >= 2 else []
 
     except Exception as e:
@@ -1072,11 +956,6 @@ def _extract_textlayer_rows(pdf_path: str, page_num: int):
         return []
 
 def _jb_extract_checks_fallback(text_layer: str) -> List[Tuple[str, str, str]]:
-    """
-    Johnson Bank: when pdfminer column extraction fails (e.g. pymupdf line text or
-    multi-column layout), extract (date, check#, amount) triplets from the Checks
-    subsection. Anchored by 'Checks (' ... 'Daily Account Balance' only.
-    """
     if not text_layer:
         return []
     m = re.search(
@@ -1105,16 +984,6 @@ def _jb_extract_checks_fallback(text_layer: str) -> List[Tuple[str, str, str]]:
     return out
 
 def _extract_jb_summary_sections(text_layer: str, needs_checks: bool = True, needs_dab: bool = True) -> str:
-    """
-    Parse Johnson Bank Checks and Daily Account Balance from pdfminer column-separated
-    page text and return as HTML tables.
-    pdfminer outputs multi-column sections as separate column lists:
-        Date      Number    Amount
-        04-14     3259      45.72
-    becomes:
-        "Date\n04-14\n04-09\n\nNumber\n3259\n3260\n\nAmount\n45.72\n1,628.00"
-    Uses pdfminer (always available) not pymupdf.
-    """
     if not text_layer:
         return ""
 
@@ -1125,7 +994,6 @@ def _extract_jb_summary_sections(text_layer: str, needs_checks: bool = True, nee
     lines = [l.strip() for l in text_layer.splitlines()]
     output = []
 
-    # ── Checks ───────────────────────────────────────────────────────────
     if needs_checks:
         checks = []
         i = 0
@@ -1183,7 +1051,6 @@ def _extract_jb_summary_sections(text_layer: str, needs_checks: bool = True, nee
                 + rows + "\n</table>"
             )
 
-    # ── Daily Account Balance ─────────────────────────────────────────────
     if needs_dab:
         balances = []
         i = 0
@@ -1228,30 +1095,17 @@ def _extract_jb_summary_sections(text_layer: str, needs_checks: bool = True, nee
     return "\n\n".join(output)
 
 def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str:
-    """
-    Compare OCR table rows against PDF text-layer rows and inject any rows
-    the OCR model silently dropped (typically identical consecutive rows).
-    Guard conditions (all must pass for any injection):
-      1. PDF text layer must exist and yield >= 2 transaction rows.
-      2. The OCR table must have DATE + DESCRIPTION + AMOUNT columns.
-      3. At least one (date, desc, amount) key must appear in both OCR and
-         text layer — ensures we are patching the right table.
-      4. Only injects copies of rows ALREADY present in OCR (ocr_count > 0)
-         — never invents new row content.
-      5. Entire function wrapped in try/except — any error returns page_md unchanged.
-    """
     if not page_md or "<table" not in page_md.lower():
         return page_md
 
     try:
         tl_rows = _extract_textlayer_rows(pdf_path, page_num)
         if not tl_rows:
-            return page_md  # scanned PDF or non-transaction page — safe no-op
+            return page_md
 
         def _norm(s):
             return re.sub(r"\s+", " ", str(s or "").strip().lower())
 
-        # Build text-layer frequency map
         tl_freq = {}
         for r in tl_rows:
             key = (_norm(r["date"]), _norm(r["desc"]), _norm(r["amount"]))
@@ -1269,7 +1123,6 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
             if len(grid) < 2:
                 continue
 
-            # Identify DATE, DESCRIPTION, AMOUNT column indices
             header = [str(c or "").strip().upper() for c in grid[0]]
             date_col = desc_col = amt_col = None
             for ci, h in enumerate(header):
@@ -1281,9 +1134,8 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                     amt_col = ci
 
             if date_col is None or desc_col is None or amt_col is None:
-                continue  # not a transaction table
+                continue
 
-            # Build OCR frequency map
             ocr_freq = {}
             ocr_rows_indexed = []
             for ri, row in enumerate(grid[1:], start=1):
@@ -1299,16 +1151,8 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                 ocr_freq[key] = ocr_freq.get(key, 0) + 1
                 ocr_rows_indexed.append((ri, key))
 
-            # Guard: overlap check for Case A (duplicate restore).
-            # For Case A we require at least one row in both OCR and text layer
-            # to confirm we are patching the right table.
-            # For Case B (entirely missing rows) we use a lighter structural check:
-            # if the OCR table has DATE+DESCRIPTION+AMOUNT columns (already verified)
-            # and the text layer date format matches the OCR date format,
-            # we can safely inject — even when zero rows overlap.
             overlap = set(ocr_freq.keys()) & set(tl_freq.keys())
 
-            # Detect date format used in OCR table (MM/DD vs MM-DD vs Mon DD)
             _ocr_dates = [
                 _norm(str(row[date_col] or ""))
                 for row in grid[1:]
@@ -1325,9 +1169,6 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
             tl_fmt  = _date_fmt(_tl_dates)
             date_fmt_match = (ocr_fmt == tl_fmt) or "other" in (ocr_fmt, tl_fmt)
 
-            # Determine missing rows.
-            # Case A: row exists in OCR but count is too low (duplicate dropped)
-            # Case B: row exists in text layer but completely absent from OCR
             to_inject = {}
             to_inject_new = {}
 
@@ -1337,12 +1178,9 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                 if tl_count > ocr_count:
                     missing = tl_count - ocr_count
                     if ocr_count > 0:
-                        # Case A: restore duplicates — requires overlap confirmation
                         if overlap:
                             to_inject[key] = missing
                     else:
-                        # Case B: entirely new row — requires date format match
-                        # (lighter guard: no overlap needed, just structural match)
                         if date_fmt_match:
                             tl_row_data = next(
                                 (r for r in tl_rows
@@ -1355,7 +1193,6 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
             if not to_inject and not to_inject_new:
                 continue
 
-            # Build patched grid — inject missing duplicate rows (Case A)
             new_grid = [grid[0]]
             for ri, row in enumerate(grid[1:], start=1):
                 new_grid.append(row)
@@ -1376,15 +1213,10 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
             new_table_html = _grid_to_html(new_grid)
             result = result.replace(table_html, new_table_html, 1)
 
-            # Case B: completely missing rows → build a SEPARATE new table
-            # appended after the patched OCR table. Using a separate table
-            # preserves the original PDF structure (e.g. Capital One has
-            # separate "Payments" and "Transactions" sections).
             if to_inject_new:
                 ncols = len(grid[0])
                 header_row = grid[0]
 
-                # Find post_date column index if it exists in the header
                 post_date_col = None
                 for ci, h in enumerate(header_row):
                     hh = str(h or "").strip().upper()
@@ -1410,14 +1242,9 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
                 if new_rows:
                     extra_grid = [header_row] + new_rows
                     extra_html = "\n\n" + _grid_to_html(extra_grid)
-                    # Insert immediately after the patched table
                     insert_pos = result.find(new_table_html) + len(new_table_html)
                     result = result[:insert_pos] + extra_html + result[insert_pos:]
 
-        # Case C: page has text-layer rows but NO transaction table in OCR output.
-        # Build a new table from the text layer and prepend it.
-        # Guard: only fires when the page has NO table with both DATE + DESCRIPTION cols.
-        # This is placed AFTER the for-loop so it only runs once per page, not per table.
         if tl_rows:
             _has_txn_table = False
             for _tbl in table_pattern.finditer(result):
@@ -1457,44 +1284,18 @@ def _patch_ocr_with_textlayer(page_md: str, pdf_path: str, page_num: int) -> str
 
     except Exception as e:
         log.warning("_patch_ocr_with_textlayer failed (page %d): %s", page_num, e)
-        return page_md  # always safe — return original on any error
+        return page_md
 
 # --------------------------
 # Fix 5: Mid-table separator row reconstruction
 # --------------------------
 
-# Date pattern for transaction rows: MM/DD, MM/DD/YY, MM/DD/YYYY
 _DATE_RE = re.compile(r"^\d{1,2}/\d{2}(?:/\d{2,4})?$")
-
-# Standalone money amount (may be split off a separator label by OCR)
 _SPLIT_AMOUNT_RE = re.compile(
     r"^-?\$?\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$|^\$-?\d{1,3}(?:,\d{3})*(?:\.\d{1,4})?$"
 )
 
 def _reconstruct_separator_rows(grid):
-    """
-    Detect and reconstruct mid-table separator / label rows that OCR has
-    incorrectly split across columns.
-    The artifact (Bank of America and similar):
-      A section label like "Card account # XXXX XXXX XXXX 2889" sits between
-      transaction rows as a full-width label. OCR splits the trailing digits
-      across columns because they are right-aligned, producing variations like:
-        2-cell split:  ["Card account # XXXX XXXX XXXX 2", "",    "889"]
-        3-cell split:  ["Card account # XXXX XXXX XXXX",   "2",   "889"]
-        clean 1-cell:  ["Card account # XXXX XXXX XXXX 2889", "", ""]
-    Detection criteria (ALL must hold — guards safe for all other PDFs):
-      1. First cell is NOT a date token (transaction rows always start with date).
-      2. First cell contains letters (it is a label, not a bare number).
-      3. All cells except the first either:
-           a. are empty, OR
-           b. are a short pure-digit fragment (1-4 digits, no decimal, no sign)
-              — these are the split-off tails of the account/card number.
-      4. There must be at least one non-empty cell after the first (to detect
-         the split; single-cell rows are also handled as clean label rows).
-    The fix:
-      Concatenate all non-empty cells in order, place the result in col 0,
-      blank all other cells.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -1502,7 +1303,7 @@ def _reconstruct_separator_rows(grid):
     if ncols < 2:
         return grid
 
-    new_grid = [grid[0]]  # keep header unchanged
+    new_grid = [grid[0]]
 
     for row in grid[1:]:
         cells = [str(c or "").strip() for c in row]
@@ -1511,36 +1312,26 @@ def _reconstruct_separator_rows(grid):
 
         non_empty = [(i, c) for i, c in enumerate(cells) if c]
 
-        # Completely empty row — leave as-is
         if len(non_empty) == 0:
             new_grid.append(row)
             continue
 
         first_idx, first_val = non_empty[0]
 
-        # Guard 1: first cell must not be a date token
         if _DATE_RE.match(first_val):
             new_grid.append(row)
             continue
 
-        # Guard 2: first cell must contain letters (label, not a bare number)
         if not re.search(r"[A-Za-z]", first_val):
             new_grid.append(row)
             continue
 
-        # Single non-empty cell — already a clean label row
         if len(non_empty) == 1:
             new_row = [""] * ncols
             new_row[0] = first_val
             new_grid.append(new_row)
             continue
 
-        # Multiple non-empty cells: check that every cell AFTER the first
-        # is a short pure-digit fragment (1-4 digits, no decimal, no sign).
-        # This is the key generic guard — it allows 2, 3, or more cells
-        # as long as all the trailing cells are digit-only fragments.
-        # Real transaction rows always have amounts with decimals (-14.19)
-        # or descriptions with letters, so they never pass this check.
         trailing = [val for _, val in non_empty[1:]]
         all_trailing_are_digit_fragments = all(
             re.fullmatch(r"\d{1,4}", v) for v in trailing
@@ -1550,7 +1341,6 @@ def _reconstruct_separator_rows(grid):
             new_grid.append(row)
             continue
 
-        # Reconstruct: concatenate all non-empty cells in order
         reconstructed = "".join(val for _, val in non_empty).strip()
         new_row = [""] * ncols
         new_row[0] = reconstructed
@@ -1563,27 +1353,6 @@ def _reconstruct_separator_rows(grid):
 # --------------------------
 
 def _merge_split_rows(grid):
-    """
-    Fix OCR artifact where a transaction row is split across two <tr> rows
-    because the description wrapped to a second line in the source PDF.
-    Two patterns detected (Hardin County Bank and similar monospace statements):
-    Pattern A — description + empty continuation:
-      Row N:   [desc, '', '', '', '']          ← description, no date/money
-      Row N+1: ['', debit, credit, date, bal]  ← money/date, no description
-      → Merge into: [desc, debit, credit, date, bal]
-    Pattern B — description + text continuation + money:
-      Row N:   [desc, '', '', '', '']               ← first line of description
-      Row N+1: [desc_cont, debit, credit, date, bal] ← continuation + money
-      → Merge into: [desc + ' ' + desc_cont, debit, credit, date, bal]
-    Guard conditions (no-op unless both rows match the pattern):
-      - Row N must have non-empty col 0 (description).
-      - Row N must have empty date column (col 3 or wherever DATE is).
-      - Row N must have empty debit AND credit columns.
-      - Row N+1 must have non-empty date column.
-      - Row N+1 must have non-empty debit OR credit column.
-      - Row N balance (if present) must be a short fragment without a decimal
-        (1-6 chars, no '.') — confirms it's a split-off ref number, not a balance.
-    """
     if not grid or len(grid) < 3:
         return grid
 
@@ -1591,7 +1360,6 @@ def _merge_split_rows(grid):
     if ncols < 3:
         return grid
 
-    # Identify column indices from header
     header = [str(c or "").strip().upper() for c in grid[0]]
     date_col = desc_col = bal_col = None
     debit_cols = []
@@ -1609,7 +1377,6 @@ def _merge_split_rows(grid):
         if re.search(r"\b(CREDIT|CREDITS|CR|DEPOSIT|ADDITIONS?)\b", h):
             credit_cols.append(ci)
 
-    # Need at least date + description + one money column
     if date_col is None or desc_col is None:
         return grid
     if not debit_cols and not credit_cols:
@@ -1618,7 +1385,6 @@ def _merge_split_rows(grid):
     money_cols = debit_cols + credit_cols
     _MONEY_RE2 = re.compile(r"^-?\$?\d{1,3}(?:,\d{3})*\.\d{1,4}$")
     _DATE_RE2  = re.compile(r"^\d{1,2}/\d{2}(?:/\d{2,4})?$|^\d{1,2}-\d{2}(?:-\d{2,4})?$")
-    # Short fragment guard: 1-6 non-space chars, no decimal point → split-off ref tail
     _FRAGMENT_RE = re.compile(r"^[^\s\.]{1,6}$")
 
     new_grid = [grid[0]]
@@ -1628,7 +1394,6 @@ def _merge_split_rows(grid):
         while len(row) < ncols:
             row.append("")
 
-        # Check if this row is a "description-only" split row candidate
         has_desc    = bool(row[desc_col])
         no_date     = not row[date_col]
         no_money    = all(not row[c] for c in money_cols)
@@ -1644,17 +1409,13 @@ def _merge_split_rows(grid):
             next_has_money = any(bool(next_row[c]) for c in money_cols)
 
             if next_has_date and next_has_money:
-                # Pattern A or B — merge
                 merged = list(next_row)
-                # Append description (and possible continuation from next row)
                 if next_row[desc_col]:
-                    # Pattern B: next row has desc continuation
                     merged[desc_col] = row[desc_col] + " " + next_row[desc_col]
                 else:
-                    # Pattern A: next row has no description
                     merged[desc_col] = row[desc_col]
                 new_grid.append(merged)
-                i += 2  # skip both rows, emit merged
+                i += 2
                 continue
 
         new_grid.append(row)
@@ -1663,17 +1424,6 @@ def _merge_split_rows(grid):
     return new_grid
 
 def _extract_fused_desc_amount(grid):
-    """
-    Fix OCR artifact where a trailing money amount gets fused into the
-    description cell instead of its own debit/credit column.
-    Example (Hardin County Bank):
-      OCR: ['CHASE CREDIT CRD EPAY 8319249882 500.00', '', '', '04/11/25', '16,699.04']
-      Fix: ['CHASE CREDIT CRD EPAY 8319249882', '500.00', '', '04/11/25', '16,699.04']
-    Guard conditions (all must hold):
-      - Description ends with a space + money amount (NNN.NN or N,NNN.NN).
-      - ALL debit AND credit columns are empty for this row.
-      - Date column is non-empty (confirms it is a complete data row).
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -1724,15 +1474,6 @@ def _extract_fused_desc_amount(grid):
     return new_grid
 
 def _is_junk_table(grid):
-    """
-    Detect and suppress known junk tables that are OCR artifacts from
-    page headers/footers, not real transaction data.
-    Current patterns:
-    1. CUSTOMER INFORMATION fragment tables (First Horizon Bank):
-       Single-column table with header "CUSTOMER INFORMATION" and
-       data rows that are short numeric fragments (e.g. "254", "24")
-       from the partial account number / year printed in the header band.
-    """
     if not grid or len(grid) < 2:
         return False
     ncols = len(grid[0])
@@ -1740,24 +1481,12 @@ def _is_junk_table(grid):
         return False
     header_text = str(grid[0][0] or "").strip().upper()
     if "CUSTOMER INFORMATION" in header_text:
-        # All data rows should be short (< 10 chars) numeric/alphanumeric fragments
         data_rows = grid[1:]
         if all(len(str(r[0] or "").strip()) <= 10 for r in data_rows):
             return True
     return False
 
 def _split_fused_multicolumn_header(grid):
-    """
-    Fix OCR artifact where multiple column headers are fused into one <th> cell.
-    Example (First Horizon Bank alternating pages):
-      Fused:   ['DATE DESCRIPTION CARD #', 'DEPOSIT', 'WITHDRAWAL']
-      Correct: ['DATE', 'DESCRIPTION', 'DEPOSIT', 'WITHDRAWAL', 'CARD #']
-    The data rows also have date + description + card# all in col 0:
-      Fused:   ['01/16 PURCHASE - UBER TRIP... 4919', '', '$21.02']
-      Correct: ['01/16', 'PURCHASE - UBER TRIP...', '', '$21.02', '4919']
-    Detection: first header cell contains 2+ keyword terms separated by spaces,
-    AND remaining header cells are valid money column keywords.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -1766,20 +1495,15 @@ def _split_fused_multicolumn_header(grid):
         return grid
 
     first_header = str(grid[0][0] or "").strip()
-    # Split first header cell into parts and check if multiple are keywords
     parts = first_header.split()
     kw_hits = [p for p in parts if _HEADER_KW_EXACT_RE.match(p)]
     if len(kw_hits) < 2:
-        return grid  # not a fused multi-keyword header
+        return grid
 
-    # The remaining header cells must all be keyword columns
     rest_headers = [str(c or "").strip() for c in grid[0][1:]]
     if not all(_HEADER_KW_EXACT_RE.match(h) for h in rest_headers if h):
         return grid
 
-    # Build the new expanded header:
-    # Split the fused first cell into individual keyword columns
-    # Keep non-keyword trailing parts (e.g. "CARD #") as the last new col
     new_header_cols = []
     non_kw_parts = []
     for p in parts:
@@ -1793,12 +1517,10 @@ def _split_fused_multicolumn_header(grid):
     if non_kw_parts:
         new_header_cols.append(" ".join(non_kw_parts))
 
-    # Full new header = expanded first cell + remaining cells
     new_header = new_header_cols + rest_headers
     new_ncols = len(new_header)
-    extra_cols = new_ncols - ncols  # how many new columns were added
+    extra_cols = new_ncols - ncols
 
-    # Identify column roles in the new header
     _DATE_COL_RE   = re.compile(r"\bDATE\b", re.IGNORECASE)
     _CARD_COL_RE   = re.compile(r"\bCARD\b", re.IGNORECASE)
     _DESC_COL_RE   = re.compile(r"\b(DESCRIPTION|DETAILS?|NARRATION|PARTICULARS?)\b", re.IGNORECASE)
@@ -1809,11 +1531,10 @@ def _split_fused_multicolumn_header(grid):
     new_card_col = next((i for i,h in enumerate(new_header) if _CARD_COL_RE.search(h)), None)
 
     if new_date_col is None or new_desc_col is None:
-        return grid  # can't map columns
+        return grid
 
-    # Re-split each data row: col 0 had "date description... card#" fused together
     _DATE_PREFIX = re.compile(r"^(\d{1,2}/\d{2}(?:/\d{2,4})?)\s+(.*)", re.DOTALL)
-    _CARD_SUFFIX = re.compile(r"^(.*?)\s+(\d{4})$", re.DOTALL)  # 4-digit card number at end
+    _CARD_SUFFIX = re.compile(r"^(.*?)\s+(\d{4})$", re.DOTALL)
 
     def pad(row, n):
         r = list(row)
@@ -1828,7 +1549,6 @@ def _split_fused_multicolumn_header(grid):
         fused_cell = cells[0]
         rest_cells = cells[1:]
 
-        # Extract date from front
         date_val = ""
         desc_val = fused_cell
         card_val = ""
@@ -1838,20 +1558,17 @@ def _split_fused_multicolumn_header(grid):
             date_val = dm.group(1)
             desc_val = dm.group(2).strip()
 
-        # Extract card# from end (4-digit number)
         if new_card_col is not None:
             cm = _CARD_SUFFIX.match(desc_val)
             if cm:
                 desc_val = cm.group(1).strip()
                 card_val = cm.group(2)
 
-        # Build new row
         new_row = [""] * new_ncols
         new_row[new_date_col] = date_val
         new_row[new_desc_col] = desc_val
         if new_card_col is not None:
             new_row[new_card_col] = card_val
-        # Fill remaining money columns from rest_cells
         money_col_indices = [i for i,h in enumerate(new_header) if _MONEY_COL_RE.search(h)]
         for mi, ri in enumerate(range(len(rest_cells))):
             if mi < len(money_col_indices):
@@ -1862,17 +1579,6 @@ def _split_fused_multicolumn_header(grid):
     return new_grid
 
 def _normalize_daily_balance_table(grid):
-    """
-    Fix OCR artifact in DAILY BALANCE SUMMARY tables where:
-    - OCR produces N DATE columns followed by N BALANCE columns
-      instead of interleaved DATE|BALANCE|DATE|BALANCE pairs
-    - Multiple dates are stacked in one cell "01/02\n01/03\n01/04\n01/05"
-    Reorders columns and expands stacked date cells into proper rows.
-    Only fires when ALL of these hold:
-    - Header has exactly 2N columns where N >= 2
-    - First N columns are all DATE, last N columns are all BALANCE
-    - At least one data cell contains a newline (stacked dates)
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -1891,13 +1597,11 @@ def _normalize_daily_balance_table(grid):
     if not all(h == "BALANCE" for h in bal_half):
         return grid
 
-    # Build new interleaved header: DATE BALANCE DATE BALANCE ...
     new_header = []
     for i in range(half):
         new_header.append("DATE")
         new_header.append("BALANCE")
 
-    # Expand each data row
     new_rows = []
     _DATE_RE3 = re.compile(r"^\d{1,2}/\d{2}(?:/\d{2,4})?$")
 
@@ -1906,23 +1610,17 @@ def _normalize_daily_balance_table(grid):
         while len(cells) < ncols:
             cells.append("")
 
-        # Dates are all stacked in cells[0]; balances follow immediately in cells[1..N]
-        # (OCR layout: stacked_dates | bal1 | bal2 | bal3 | bal4 | empty...)
-        # NOT: date | date | date | date | bal | bal | bal | bal (despite header order)
         date_cells = [cells[0]]
         bal_cells  = cells[1:]
 
-        # Check if dates are stacked in first cell (newline OR space-separated)
-        # e.g. "01/02 01/03 01/04 01/05" or "01/02\n01/03\n01/04\n01/05"
         first = date_cells[0]
         tokens = re.split(r"[\n\r\s]+", first)
         date_tokens = [t.strip() for t in tokens if t.strip() and _DATE_RE3.match(t.strip())]
         if len(date_tokens) >= 2:
             stacked = date_tokens
         else:
-            stacked = date_cells  # already one date per cell
+            stacked = date_cells
 
-        # Build ONE interleaved row for this group of stacked dates
         new_row = [""] * ncols
         for i, date_val in enumerate(stacked):
             if not _DATE_RE3.match(date_val):
@@ -1938,16 +1636,6 @@ def _normalize_daily_balance_table(grid):
     return [new_header] + new_rows
 
 def _normalize_checks_paid_table(grid):
-    """
-    Fix OCR artifact in CHECKS PAID SUMMARY tables where:
-    - Header col0 = "DATE CHECK # DATE CHECK # DATE CHECK #" (fused repeating groups)
-    - Remaining headers = "AMOUNT AMOUNT AMOUNT"
-    - Data col0 = "01/24 4701 01/18 4704 * 01/18 916831 *" (all groups fused)
-    - Remaining data cells = amount values
-    Expands to proper: DATE | CHECK # | AMOUNT | DATE | CHECK # | AMOUNT | ...
-    Detection: first header cell matches pattern (DATE CHECK #)+ 
-    and remaining headers are all AMOUNT.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -1960,17 +1648,14 @@ def _normalize_checks_paid_table(grid):
     if not _CHECKS_HDR.match(first_h):
         return grid
 
-    # Remaining headers must be AMOUNT (or empty)
     rest_h = [str(c or "").strip().upper() for c in grid[0][1:]]
     if not all(h in ("AMOUNT", "") for h in rest_h if h):
         return grid
 
-    # Count groups from how many times DATE appears in the fused header
     n_groups = len(re.findall(r"\bDATE\b", first_h, re.IGNORECASE))
     if n_groups < 1:
         return grid
 
-    # Build new header: DATE | CHECK # | AMOUNT repeated n_groups times
     new_header = []
     for _ in range(n_groups):
         new_header.extend(["DATE", "CHECK #", "AMOUNT"])
@@ -1986,7 +1671,6 @@ def _normalize_checks_paid_table(grid):
         fused_data = cells[0]
         amounts    = cells[1:]
 
-        # Split fused data into groups by date boundary
         tokens = fused_data.split()
         groups = []
         current = []
@@ -1999,31 +1683,18 @@ def _normalize_checks_paid_table(grid):
         if current and current[0]:
             groups.append(current)
 
-        # Build one output row
         new_row = [""] * len(new_header)
         for i, grp in enumerate(groups):
             if i >= n_groups:
                 break
-            new_row[i * 3]     = grp[0]                       # DATE
-            new_row[i * 3 + 1] = " ".join(grp[1:]).strip()    # CHECK #
-            new_row[i * 3 + 2] = amounts[i] if i < len(amounts) else ""  # AMOUNT
+            new_row[i * 3]     = grp[0]
+            new_row[i * 3 + 1] = " ".join(grp[1:]).strip()
+            new_row[i * 3 + 2] = amounts[i] if i < len(amounts) else ""
 
         new_rows.append(new_row)
 
     return new_rows
 
-# ── Johnson Bank plain-text section converter ─────────────────────────────
-# Johnson Bank formats Deposits, Withdrawals, Checks, and Daily Account Balance
-# as plain text columns (no HTML tables). This converter detects those sections
-# and emits proper <table> HTML.
-#
-# SAFETY GUARDS — all four must hold before any conversion fires:
-#   1. Line matches exact section-header pattern
-#   2. Next non-blank line matches exact column-header pattern
-#   3. First data row uses MM-DD date format (Johnson Bank's unique date separator)
-#   4. The candidate section contains NO existing <table> tags
-# Guard 3 is the key discriminator: all other supported banks use MM/DD.
-
 _JB_SECTION_HDR = re.compile(
     r"^(Deposits?|Withdrawals?|Checks?(?:\s+Paid)?|Daily\s+Account\s+Balance)"
     r"\s*(?:\(cont\.?\))?\s*$",
@@ -2034,7 +1705,6 @@ _JB_COL_HDR = re.compile(
     r"|^Date\s+(?:Number\s+)?(?:Balance|Amount)\s*$",
     re.IGNORECASE,
 )
-# MM-DD date (dash separator) — Johnson Bank's unique format
 _JB_DATE_DD  = re.compile(r"^\d{2}-\d{2}\s+")
 _JB_TXN_ROW  = re.compile(r"^(\d{2}-\d{2})\s+(.+?)\s+(-?\d{1,3}(?:,\d{3})*\.\d{2})\s*$")
 _JB_BAL_ROW  = re.compile(r"^(\d{2}-\d{2})\s+(\d{1,3}(?:,\d{3})*\.\d{2})\s*$")
@@ -2049,24 +1719,15 @@ def _jb_rows_to_html(header_cols, rows):
     return "\n".join(parts)
 
 def _jb_section_has_table(lines, start, end):
-    """Return True if any line in lines[start:end] contains a <table tag."""
     return any("<table" in l.lower() for l in lines[start:end])
 
 def _find_first_data_row(lines, start):
-    """Return index of first non-blank line after start, or None."""
     for i in range(start, min(start + 20, len(lines))):
         if lines[i].strip():
             return i
     return None
 
 def convert_plaintext_bank_sections(text: str) -> str:
-    """
-    Convert Johnson Bank plain-text transaction sections to HTML tables.
-    All four safety guards must pass before conversion fires.
-    Safe no-op for all other banks.
-    """
-    # Quick exit: if the page already has lots of tables, skip entirely
-    # (handles pages that are already properly formatted)
     lines = text.splitlines()
     out   = []
     i     = 0
@@ -2074,7 +1735,6 @@ def convert_plaintext_bank_sections(text: str) -> str:
     while i < len(lines):
         line = lines[i].strip()
 
-        # Guard 1: section header
         if not _JB_SECTION_HDR.match(line):
             out.append(lines[i])
             i += 1
@@ -2083,11 +1743,9 @@ def convert_plaintext_bank_sections(text: str) -> str:
         section_title = line
         j = i + 1
 
-        # Skip blank lines to find column header
         while j < len(lines) and not lines[j].strip():
             j += 1
 
-        # Guard 2: column header
         if j >= len(lines) or not _JB_COL_HDR.match(lines[j].strip()):
             out.append(lines[i])
             i += 1
@@ -2096,16 +1754,13 @@ def convert_plaintext_bank_sections(text: str) -> str:
         col_hdr_line = lines[j].strip()
         data_start   = j + 1
 
-        # Find first non-blank data line
         first_data_idx = _find_first_data_row(lines, data_start)
 
-        # Guard 3: first data row must use MM-DD date format
         if first_data_idx is None or not _JB_DATE_DD.match(lines[first_data_idx].strip()):
             out.append(lines[i])
             i += 1
             continue
 
-        # Find section end (next section header, page separator, or end of text)
         section_end = len(lines)
         for k in range(data_start, len(lines)):
             l = lines[k].strip()
@@ -2115,13 +1770,11 @@ def convert_plaintext_bank_sections(text: str) -> str:
                 section_end = k
                 break
 
-        # Guard 4: no existing <table> tags in this section
         if _jb_section_has_table(lines, i, section_end):
             out.append(lines[i])
             i += 1
             continue
 
-        # All guards passed — parse and convert
         parts_upper = col_hdr_line.upper().split()
         if len(parts_upper) == 2 and parts_upper[1] == "BALANCE":
             header_cols = ["Date", "Balance"]
@@ -2148,7 +1801,6 @@ def convert_plaintext_bank_sections(text: str) -> str:
                     or row_line.startswith("Page:")):
                 break
 
-            # Balance-only row
             mb = _JB_BAL_ROW.match(row_line)
             if mb and header_cols == ["Date", "Balance"]:
                 if current:
@@ -2157,7 +1809,6 @@ def convert_plaintext_bank_sections(text: str) -> str:
                 k += 1
                 continue
 
-            # Full transaction row
             mt = _JB_TXN_ROW.match(row_line)
             if mt:
                 if current:
@@ -2171,7 +1822,6 @@ def convert_plaintext_bank_sections(text: str) -> str:
                 k += 1
                 continue
 
-            # Continuation line
             if current and row_line:
                 desc_key = "Description" if "Description" in header_cols else "Number"
                 if desc_key in current:
@@ -2187,7 +1837,7 @@ def convert_plaintext_bank_sections(text: str) -> str:
         if rows:
             out.append(_jb_rows_to_html(header_cols, rows))
 
-        i = section_end   # jump past the whole section
+        i = section_end
         continue
 
     return "\n".join(out)
@@ -2197,10 +1847,6 @@ def convert_plaintext_bank_sections(text: str) -> str:
 # --------------------------
 
 def _is_nfcu_document(pdf_path: str) -> bool:
-    """
-    Strict detector for Navy Federal statements.
-    Used to gate NF-specific parsing so other banks are unaffected.
-    """
     try:
         import pymupdf as fitz
         doc = fitz.open(pdf_path)
@@ -2210,7 +1856,6 @@ def _is_nfcu_document(pdf_path: str) -> bool:
             text += "\n" + (doc[i].get_text() or "")
         doc.close()
         t = text.lower()
-        # Strict Navy-only guard to avoid affecting other statement types.
         return (
             ("statement of account" in t)
             and ("access no." in t)
@@ -2222,21 +1867,6 @@ def _is_nfcu_document(pdf_path: str) -> bool:
 
 
 def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
-    """
-    Full text-layer parser for Navy Federal Credit Union statements.
-    GLM-OCR fails on NFCU pages because:
-      - The summary table has 5 wide columns with no visible borders
-      - Transaction tables have Amount($) and Balance($) spatially far right
-      - Multi-line descriptions (wrapped lines) confuse OCR table detection
-    This function reads the PDF text layer directly via pdfplumber and produces
-    correct, complete HTML for every section: summary table, transaction tables,
-    items paid, savings, disclosures.
-    Detection guards (all must pass — safe no-op for all other banks):
-      1. Page text must contain 'Access No.' AND 'Statement of Account'
-         (unique to Navy Federal statement format)
-      2. pdfplumber must return extractable text (not a scanned page)
-    Returns complete HTML string for the page, or "" if not NFCU / any error.
-    """
     try:
         import pymupdf as fitz
 
@@ -2260,7 +1890,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
         if not is_nfcu:
             return ""
 
-        # Group words into reading lines by y-position.
         buckets = []
         for w in words:
             if len(w) < 5:
@@ -2322,12 +1951,10 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                     i += 1
                     continue
 
-                # Section/account boundaries
                 if acct_re.match(text) and i != start_idx:
                     break
                 if any(text.startswith(s) for s in stop_markers):
                     break
-                # Some boilerplate lines include prefixes/suffixes; match by containment too.
                 upper_text = text.upper()
                 if (
                     "REMITTANCE RECEIVED" in upper_text
@@ -2364,7 +1991,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                         amount = mvals[-2]
                         balance = mvals[-1]
                     elif len(mvals) == 1:
-                        # Many NFCU lines can be split; single amount token is usually balance.
                         balance = mvals[-1]
 
                     pending = {
@@ -2378,8 +2004,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                         mvals = [tok for tok in toks if money_re.match(tok)]
                         desc_tokens = [tok for tok in toks if not money_re.match(tok)]
 
-                        # Stop if we have already captured a balance and now hit
-                        # known non-transaction boilerplate lines.
                         upper = text.upper()
                         if pending["balance"] and (
                             "REMITTANCE" in upper
@@ -2391,8 +2015,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                         ):
                             break
 
-                        # Beginning/Ending balance rows are single logical rows.
-                        # Do not absorb unrelated continuation text.
                         if (
                             ("BEGINNING BALANCE" in pending["desc"].upper() or "ENDING BALANCE" in pending["desc"].upper())
                             and not mvals
@@ -2408,12 +2030,10 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                                 pending["balance"] = mvals[-1]
                         elif len(mvals) == 1 and not pending["balance"]:
                             pending["balance"] = mvals[-1]
-                    # else ignore stray line
                 i += 1
 
             if pending is not None:
                 rows.append(pending)
-            # keep rows that have at least date+desc, and preferably balance
             cleaned = [r for r in rows if r["date"] and r["desc"] and (r["amount"] or r["balance"])]
             return cleaned, i
 
@@ -2444,7 +2064,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
                 i = i2
                 continue
 
-            # Keep section headers only (avoid noisy paragraph dump)
             if text in ("Checking", "Savings"):
                 out.append(f"<h3>{text}</h3>")
             elif text.startswith("Items Paid"):
@@ -2458,11 +2077,6 @@ def _parse_nfcu_page(pdf_path: str, page_num: int) -> str:
         return ""
 
 def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
-    """
-    Navy-only additive helper:
-    Rebuild "Summary of your deposit accounts" from PDF text-layer words.
-    Returns a clean HTML table or "" (safe no-op on any miss/error).
-    """
     try:
         import pymupdf as fitz
 
@@ -2476,7 +2090,6 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
         if not words:
             return ""
 
-        # Group tokens into reading lines by y-position.
         buckets = []
         for w in words:
             if len(w) < 5:
@@ -2519,8 +2132,6 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
             re.IGNORECASE,
         )
 
-        # NFCU summary table real columns from PDF:
-        # Account | Previous Balance | Deposits/Credits | Withdrawals/Debits | Ending Balance | YTD Dividends
         rows = []
         pending_account = ""
         account_name_re = re.compile(
@@ -2538,7 +2149,6 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
             t0 = toks[0].lower()
             tl = text.lower()
 
-            # Ignore split header lines ("Previous ...", "Balance Credits ...")
             if (
                 "previous" in tl
                 or "deposits" in tl
@@ -2548,7 +2158,6 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
             ) and not any(money_re.match(t) for t in toks):
                 continue
 
-            # Account name can appear on its own line, followed by acct# + numeric columns.
             if account_name_re.match(text.strip()) and not any(money_re.match(t) for t in toks):
                 pending_account = text.strip()
                 continue
@@ -2556,12 +2165,10 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
             mvals = [t.replace("$", "") for t in toks if money_re.match(t)]
             acct_tokens = [t for t in toks if acct_no_re.match(t)]
 
-            # Totals row: "Totals <five amounts>"
             if t0 == "totals" and len(mvals) >= 5:
                 rows.append(["Totals", mvals[0], mvals[1], mvals[2], mvals[3], mvals[4]])
                 continue
 
-            # Account numeric row: "<acctno> <five amounts>"
             if acct_tokens and len(mvals) >= 5:
                 acct = acct_tokens[0]
                 acct_label = (pending_account + " - " + acct).strip(" -") if pending_account else acct
@@ -2585,25 +2192,18 @@ def _extract_nfcu_summary_table(pdf_path: str, page_num: int) -> str:
         return ""
 
 def _replace_nfcu_summary_table(page_md: str, summary_table_html: str) -> str:
-    """
-    Replace the first table after 'Summary of your deposit accounts' with
-    a rebuilt NFCU summary table. Safe no-op if anchor/table is missing.
-    """
     if not page_md or not summary_table_html:
         return page_md
     table_pattern = re.compile(r"<table[^>]*>.*?</table>", re.DOTALL | re.IGNORECASE)
     lower = page_md.lower()
     anchor = lower.find("summary of your deposit accounts")
 
-    # Primary path: replace first table after the summary heading.
     if anchor >= 0:
         for m in table_pattern.finditer(page_md):
             if m.start() > anchor:
                 old_table = m.group(0)
                 return page_md.replace(old_table, summary_table_html, 1)
 
-    # Fallback path: heading may be malformed/absent in OCR text.
-    # For NFCU summary, header should contain balance/deposit/withdrawal/ytd signals.
     for m in table_pattern.finditer(page_md):
         tbl = m.group(0)
         p = TableGridParser()
@@ -2624,11 +2224,6 @@ def _replace_nfcu_summary_table(page_md: str, summary_table_html: str) -> str:
     return page_md
 
 def _split_fused_date_transaction_header(grid):
-    """
-    Fix fused header "Date Transaction Detail" -> separate Date + Transaction Detail.
-    Strict guard: only when header has that exact fused phrase and also has
-    Amount/Balance columns, so unrelated tables are untouched.
-    """
     if not grid or len(grid) < 2:
         return grid
     ncols = max(len(r) for r in grid if isinstance(r, list))
@@ -2679,13 +2274,6 @@ def _split_fused_date_transaction_header(grid):
     return new_grid
 
 def _split_combined_summary_daily_balance_grid(grid):
-    """
-    Some OCR outputs fuse an account summary table and a Daily Balance block
-    into one HTML table. Split them into two tables when structure is clear.
-    Generic guards:
-    - A row containing "Daily Balance" exists
-    - There are rows after it with date-like + amount-like tokens
-    """
     if not grid or len(grid) < 4:
         return None
 
@@ -2719,7 +2307,6 @@ def _split_combined_summary_daily_balance_grid(grid):
         for d, a in zip(dates, amts):
             daily_rows.append([d, a])
 
-    # Need meaningful daily table to split safely.
     if len(daily_rows) < 2:
         return None
 
@@ -2731,20 +2318,12 @@ def _split_combined_summary_daily_balance_grid(grid):
     return [summary, daily]
 
 def _normalize_double_desc_amount_header_table(grid):
-    """
-    Fix 13: Collapse OCR tables with a duplicated column group:
-      Date | Description | Amount | Description | Amount
-    (or 4 cols if trailing Amount column was dropped). UCB-style summary tables
-    put labels like "9 Credit(s) This Period" in the Date column.
-    Output: Date | Description | Amount
-    """
     if not grid or len(grid) < 2:
         return grid
 
     def _hdr_tight(s):
         return re.sub(r"\s+", "", str(s or "").strip().lower())
 
-    # OCR often inserts spaces inside words; compare whitespace-stripped tokens.
     expected5_t = ["date", "description", "amount", "description", "amount"]
     expected4_t = ["date", "description", "amount", "description"]
 
@@ -2794,7 +2373,6 @@ def _normalize_double_desc_amount_header_table(grid):
         if not any(cells[:ncols]):
             continue
 
-        # A) Fused date + label in col0, amount in col1 (Beginning/Ending Balance, etc.)
         mpre = date_prefix_re.match(c0)
         if mpre and _is_money(c1):
             date_s = mpre.group(1).strip()
@@ -2808,7 +2386,6 @@ def _normalize_double_desc_amount_header_table(grid):
             out.append([date_s, desc, c1])
             continue
 
-        # B) Date-only col0; description may contain Beginning Balance + amount; amount col wrong ($0)
         if date_only_re.match(c0.strip()):
             date_s = c0.strip()
             desc = c1
@@ -2828,25 +2405,16 @@ def _normalize_double_desc_amount_header_table(grid):
             out.append([date_s, desc, amt])
             continue
 
-        # D) Summary / label rows: first column is NOT a date, second is money (Credits/Debits/Service)
         if not date_prefix_re.match(c0) and not date_only_re.match(c0.strip()) and _is_money(c1):
             out.append(["", c0, c1])
             continue
 
-        # C) Default: first Date|Description|Amount triple (ignore duplicate pair)
         out.append([c0, c1, c2])
 
     return out if len(out) >= 2 else grid
 
 
 def _normalize_ucb_three_col_beginning_balance_fusion(grid):
-    """
-    Fix 14: UCB sometimes OCRs the summary as 3 columns (not 5), with the real
-    beginning balance inside the description and $0.00 in the amount column:
-      02/01/2025 | Beginning Balance $3,562.49 Average Ledger Balance | $0.00
-    Normalize to: Date | Beginning Balance | $3,562.49
-    Guard: standard Date|Description|Amount header only.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -2854,7 +2422,6 @@ def _normalize_ucb_three_col_beginning_balance_fusion(grid):
         return re.sub(r"\s+", "", str(s or "").strip().lower())
 
     h0 = [str(c or "").strip() for c in grid[0]]
-    # Do not use max row width — OCR sometimes adds an extra empty column on some rows only.
     if len(h0) < 3:
         return grid
     if _tight(h0[0]) != "date" or _tight(h0[1]) != "description" or _tight(h0[2]) != "amount":
@@ -2894,12 +2461,6 @@ def _normalize_ucb_three_col_beginning_balance_fusion(grid):
 
 
 def _fix_three_col_date_desc_amount_left_shift(grid):
-    """
-    Fix 15: OCR shifts transaction rows one column left:
-      col0 = "02/24/2025 ITM DEPOSIT", col1 = "$1,440.00", col2 = empty
-    Expected: Date | Description | Amount
-    Activates only on Date|Description|Amount headers and strict money in col1 with empty col2.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -2948,14 +2509,6 @@ def _fix_three_col_date_desc_amount_left_shift(grid):
 
 
 def _normalize_fused_date_posted_amount_table(grid):
-    """
-    Normalize 2-column OCR tables with fused header like:
-      col1: "Date posted Transaction description [Reference number]"
-      col2: "Amount"
-    into stable 3-column transaction layout:
-      Date | Transaction Description | Amount
-    Strict guard: only activates on that specific fused-header signature.
-    """
     if not grid or len(grid) < 2:
         return grid
 
@@ -2988,7 +2541,6 @@ def _normalize_fused_date_posted_amount_table(grid):
             out.append(["Date", "Transaction Description", "Amount"])
             continue
         if ri < hdr_idx:
-            # Keep pre-header text as non-transaction row.
             joined = " ".join(c for c in cells if c).strip()
             out.append(["", joined, ""])
             continue
@@ -2996,7 +2548,6 @@ def _normalize_fused_date_posted_amount_table(grid):
         c1 = cells[0].strip()
         c2 = cells[1].strip() if len(cells) > 1 else ""
 
-        # Skip duplicated inline sub-headers that often appear in fused OCR tables.
         lower_c1 = c1.lower()
         if (
             lower_c1.startswith("date posted transaction description")
@@ -3012,8 +2563,6 @@ def _normalize_fused_date_posted_amount_table(grid):
                 date = parts[0]
                 desc = " ".join(parts[1:]).strip()
 
-        # Amount may live in any shifted column for malformed OCR tables.
-        # Prefer explicit second-column value, then rightmost strict money token in row.
         amount = c2
         if not amount:
             money_cells = []
@@ -3024,7 +2573,6 @@ def _normalize_fused_date_posted_amount_table(grid):
             if money_cells:
                 amount = money_cells[-1]
 
-        # If still empty, try extracting trailing amount from description.
         if not amount and desc:
             m = re.search(r"(-?\$?\d{1,3}(?:,\d{3})*(?:\.\d{2})-?)\s*$", desc)
             if m:
@@ -3035,23 +2583,11 @@ def _normalize_fused_date_posted_amount_table(grid):
             continue
         out.append([date, desc, amount])
 
-    # Keep at least header + 1 row to avoid degrading clean tables on accidental match.
     if len(out) < 2:
         return grid
     return out
 
 def normalize_html_tables(text: str) -> str:
-    """
-    For every <table>...</table>:
-      1. Parse to grid (expand colspan/rowspan).
-      2. Drop truly-empty columns.
-      3. Merge blank-header text columns into the left column.
-      4. Fix 1 — clean fused header artifacts and recover the lost first data row.
-      5. Fix 2 — promote a misplaced header row when real column headers landed
-                  in a <td> data row instead of the <th> header row.
-      6. Fix 3 — split fused key-value rows in summary sections.
-      7. Emit normalised <table> HTML.
-    """
     if not text or "<table" not in text.lower():
         return text
 
@@ -3067,67 +2603,44 @@ def normalize_html_tables(text: str) -> str:
             p.feed(table_html)
             grid = _build_grid(p.rows)
 
-            # Suppress known junk tables (OCR artifacts from page headers)
             if _is_junk_table(grid):
-                # Append nothing (suppress the table), update last, skip rest of pipeline
                 out.append("")
-                last = m.end()  # must set here before continue since line below is outside try
+                last = m.end()
                 continue
 
-            # Fix 13: MUST run before _drop_truly_empty_columns — otherwise the 5th column
-            # can be removed and the UCB double-pair header no longer matches.
             grid = _normalize_double_desc_amount_header_table(grid)
-            # Fix 14: same bank, 3-col OCR fused Beginning Balance + $0.00 amount column.
             grid = _normalize_ucb_three_col_beginning_balance_fusion(grid)
-            # Fix 15: "Deposits (continued)" etc. — date+desc in col1, amount in col2, empty col3.
             grid = _fix_three_col_date_desc_amount_left_shift(grid)
 
             grid = _drop_truly_empty_columns(grid)
             grid = _merge_blank_header_text_columns(grid)
 
-            # Fix 5: reconstruct mid-table separator rows split across columns by OCR
             grid = _reconstruct_separator_rows(grid)
 
-            # Fix 1: fused-header artifact (two-signal guard — safe on clean PDFs)
             grid, recovered_row = _clean_header_artifacts(grid)
             if recovered_row is not None:
                 grid.insert(1, recovered_row)
 
-            # Fix 2: misplaced header row promotion (keyword-score guard — safe on clean PDFs)
             grid = _promote_misplaced_header_row(grid)
 
-            # Fix NF-2: split fused "Date Transaction Detail" header/cell.
             grid = _split_fused_date_transaction_header(grid)
 
-            # Fix 11: normalize fused "Date posted Transaction description ... / Amount"
-            # two-column tables into stable Date/Description/Amount structure.
             grid = _normalize_fused_date_posted_amount_table(grid)
 
-            # Fix 8: split fused multi-keyword header cell (First Horizon Bank)
             grid = _split_fused_multicolumn_header(grid)
 
-            # Fix 9: normalize DAILY BALANCE SUMMARY tables with stacked dates
             grid = _normalize_daily_balance_table(grid)
 
-            # Fix 10: normalize CHECKS PAID SUMMARY repeating-group tables
             grid = _normalize_checks_paid_table(grid)
 
-            # Fix 6: merge split rows where description wraps to next line
-            # (Hardin County Bank and similar monospace statement formats)
             grid = _merge_split_rows(grid)
 
-            # Fix 7: extract trailing amount fused into description cell
-            # (Hardin County Bank: "CHASE CREDIT CRD EPAY 8319249882 500.00")
             grid = _extract_fused_desc_amount(grid)
 
-            # Fix 19: junk Amount cell but money token still present in row text (Truist MSBILL/PIN).
             grid = _repair_da3_garbled_amount_cells(grid)
 
-            # Fix 3: fused key-value rows in summary sections (no-space guard — safe on all PDFs)
             grid = _fix_fused_keyvalue_rows(grid)
-            # Fix 16/20: duplicate data rows inside same Date|Description|Amount table (OCR repeats).
             grid = _dedupe_duplicate_rows_in_da3_table(grid)
-            # Fix 12: split fused summary+daily-balance combined table, if detected.
             split_tables = _split_combined_summary_daily_balance_grid(grid)
             if split_tables:
                 out.append("\n\n".join(_grid_to_html(g) for g in split_tables if g))
@@ -3142,7 +2655,6 @@ def normalize_html_tables(text: str) -> str:
 
 
 def _transaction_row_dedupe_key(cells):
-    """Stable key for Date|Description|Amount rows (OCR variants, HTML entities)."""
     parts = [str(c or "").strip() for c in cells[:3]]
     while len(parts) < 3:
         parts.append("")
@@ -3150,16 +2662,12 @@ def _transaction_row_dedupe_key(cells):
     date_s = re.sub(r"\s+", "", date_s)
     amt_s = re.sub(r"[\s$,]", "", amt_s).lower().replace("−", "-")
     desc_s = html.unescape(desc_s)
-    desc_s = desc_s.replace("`", "'").replace("’", "'")
+    desc_s = desc_s.replace("`", "'").replace("'", "'")
     desc_s = re.sub(r"\s+", "", desc_s.lower())
     return (date_s, desc_s, amt_s)
 
 
 def _da3_loose_key(cells):
-    """
-    (date, amount) only — for detecting duplicate DA3 tables when description drifts between OCR passes.
-    Uses strict currency + float amount so junk cells do not create keys; 1,234.56 vs 1234.56 match.
-    """
     parts = [str(c or "").strip() for c in cells[:3]]
     while len(parts) < 3:
         parts.append("")
@@ -3181,7 +2689,6 @@ def _da3_loose_key(cells):
 
 
 def _row_counter_loose_da3(g):
-    """Multiset of (date, amount) for DA3 data rows with strict amounts."""
     c = Counter()
     if not g or len(g) < 2:
         return c
@@ -3196,7 +2703,6 @@ def _row_counter_loose_da3(g):
 
 
 def _overlap_ratio_counter(c_num, c_den):
-    """Fraction of c_num's multiset mass matched in c_den (min counts per key)."""
     tot = sum(c_num.values())
     if tot == 0:
         return 0.0
@@ -3228,11 +2734,6 @@ _DA3_LEADING_DATE = re.compile(r"^\d{1,2}/\d{1,2}(?:/\d{2,4})?$")
 
 
 def _repair_da3_garbled_amount_cells(grid):
-    """
-    Fix 19: Amount column shows fused junk (letters, card digits) while the real debit/credit
-    still appears as a money token in the row text — use the rightmost plausible amount.
-    Only for Date|Description|Amount tables and strict MM/DD dates in column 0.
-    """
     if not grid or len(grid) < 2 or not _is_da3_header(grid):
         return grid
 
@@ -3290,11 +2791,6 @@ def _repair_da3_garbled_amount_cells(grid):
 
 
 def _dedupe_duplicate_rows_in_da3_table(grid):
-    """
-    Fix 16 + 20: Drop long runs of OCR-identical rows; allow up to 2 consecutive rows with the
-    same normalized Date|Description|Amount key (legitimate duplicate charges). Non-consecutive
-    repeats are kept (each run capped separately).
-    """
     if not grid or len(grid) < 2:
         return grid
     if not _is_da3_header(grid):
@@ -3331,7 +2827,6 @@ def _dedupe_duplicate_rows_in_da3_table(grid):
 
 
 def _row_text_full_ucb(grid, r: int) -> str:
-    """Lowercased join of all cells in row (for UCB section split heuristics)."""
     if not grid or r < 0 or r >= len(grid):
         return ""
     parts = [str(c or "").strip().lower() for c in grid[r]]
@@ -3346,7 +2841,6 @@ def _is_itm_deposit_row_ucb(grid, r: int) -> bool:
 
 
 def _ucb_verify_electronic_credits_rows(grid, ec_start: int) -> bool:
-    """POS Return, Acct Fund (Apple instant), DDA Transfer IN — PDF Electronic Credits block."""
     if ec_start + 2 >= len(grid):
         return False
     t0 = _row_text_full_ucb(grid, ec_start)
@@ -3359,10 +2853,6 @@ def _ucb_verify_electronic_credits_rows(grid, ec_start: int) -> bool:
 
 
 def _ucb_split_grid_deposits_ec_ed(grid):
-    """
-    Return (grid_dep, grid_ec, grid_ed) each with same Date|Description|Amount header,
-    or None if this does not look like the merged UCB page-3 pattern.
-    """
     if not grid or len(grid) < 2:
         return None
     if not _is_da3_header(grid):
@@ -3390,10 +2880,6 @@ def _ucb_split_grid_deposits_ec_ed(grid):
 
 
 def _try_split_ucb_deposits_table_to_three(prefix: str, table_html: str):
-    """
-    Build HTML: Deposits (continued) + table, Electronic Credits + table, Electronic Debits + table.
-    Returns None if split heuristics do not match.
-    """
     try:
         p = TableGridParser()
         p.feed(table_html)
@@ -3421,11 +2907,6 @@ _UCB_DEP_CONT_BLOCK = re.compile(
 
 
 def _split_ucb_deposits_electronic_sections_html(text: str) -> str:
-    """
-    Fix 17: Split merged UCB Deposits (continued) table into Deposits + Electronic Credits + Electronic Debits.
-    Removes orphan empty Electronic Credits / Electronic Debits div pairs when replaced by real tables.
-    Safe no-op when heuristics do not match.
-    """
     if not text or "deposits (continued)" not in text.lower():
         return text
 
@@ -3441,7 +2922,6 @@ def _split_ucb_deposits_electronic_sections_html(text: str) -> str:
 
 
 def _parse_da3_grid_from_table_html(table_html: str):
-    """Return grid if table is Date|Description|Amount; else None."""
     try:
         p = TableGridParser()
         p.feed(table_html)
@@ -3454,11 +2934,6 @@ def _parse_da3_grid_from_table_html(table_html: str):
 
 
 def _strip_orphan_continued_da3_flatline(text: str) -> str:
-    """
-    Fix 22: Remove one-line OCR dumps like
-    '…(continued) DATE DESCRIPTION AMOUNT($) 04/09 … 04/11 … 239.55' that duplicate the
-    following HTML table and pollute markdown.
-    """
     if not text or "(continued)" not in text.lower():
         return text
     lines = text.splitlines(keepends=True)
@@ -3481,10 +2956,6 @@ def _strip_orphan_continued_da3_flatline(text: str) -> str:
 
 
 def _strip_standalone_continued_banner_line(text: str) -> str:
-    """
-    Fix 27: Remove a single line that is only an account/section “(continued)” banner
-    (no transaction columns), often duplicated between page-separator and the next heading/table.
-    """
     if not text or "(continued)" not in text.lower():
         return text
     lines = text.splitlines(keepends=True)
@@ -3512,10 +2983,6 @@ def _strip_standalone_continued_banner_line(text: str) -> str:
 
 
 def _patch_da3_msft_g_ref_amounts_across_tables(text: str) -> str:
-    """
-    Fix 21: Build MICROSOFT#G… -> strict Amount from any DA3 row, then patch other DA3 rows
-    with the same id whose Amount cell is not strict currency (junk fused into Amount).
-    """
     if not text or "<table" not in text.lower():
         return text
     if "microsoft#" not in text.lower():
@@ -3604,12 +3071,6 @@ def _patch_da3_msft_g_ref_amounts_across_tables(text: str) -> str:
 
 
 def _trim_adjacent_da3_suffix_duplicating_next_table_prefix(text: str) -> str:
-    """
-    Remove trailing data rows from a DA3 table when they are repeated as the opening
-    rows of the immediately following DA3 table (OCR/LLM glued the next section onto
-    the prior table). Requires overlap length >= 2 and leaves at least one data row
-    in the first table. Safe no-op when patterns do not match.
-    """
     if not text or "<table" not in text.lower():
         return text
 
@@ -3662,19 +3123,6 @@ def _trim_adjacent_da3_suffix_duplicating_next_table_prefix(text: str) -> str:
 
 
 def _dedupe_subset_transaction_tables_html(text: str) -> str:
-    """
-    Remove duplicate 3-column Date|Description|Amount tables:
-      - Later table is a strict subset of an earlier one → drop later (original).
-      - Earlier table is a strict subset of a later one → drop earlier (UCB-style
-        "Deposits (continued)" preview before the same rows appear in the full list).
-      - Identical row sets → drop the later table.
-      - Fuzzy overlap (≥92% of later rows match earlier) → drop later (Electronic Debits vs main).
-      - Fix 23: multiset overlap on (date, amount) only — drop when OCR varies description
-        between two copies of the same section (e.g. withdrawals pasted twice).
-      - Runs multiple passes until stable.
-    Row matching uses normalized keys so OCR variants (e.g. "0 27" vs "027", backticks)
-    do not prevent duplicate tables from being detected.
-    """
     if not text or "<table" not in text.lower():
         return text
 
@@ -3729,12 +3177,10 @@ def _dedupe_subset_transaction_tables_html(text: str) -> str:
                 if not sj:
                     continue
 
-                # Earlier j fully contained in later i (preview + full list) → drop earlier.
                 if sj <= si and len(sj) < len(si):
                     drop_idx.add(j)
                     continue
 
-                # Later i strict subset of earlier j → drop later (e.g. Electronic Debits).
                 if si <= sj and len(si) < len(sj):
                     drop_idx.add(i)
                     break
@@ -3744,7 +3190,6 @@ def _dedupe_subset_transaction_tables_html(text: str) -> str:
                     break
 
                 inter = si & sj
-                # Later table's rows almost all appear in earlier table (OCR drift).
                 ri = len(inter) / len(si) if si else 0.0
                 rj = len(inter) / len(sj) if sj else 0.0
                 if ri >= 0.88 and len(si) <= len(sj):
@@ -3754,14 +3199,12 @@ def _dedupe_subset_transaction_tables_html(text: str) -> str:
                     drop_idx.add(j)
                     continue
 
-                # Fix 23: multiset (date, amount) overlap — duplicate tables when description OCR differs.
                 ci = _row_counter_loose_da3(gi)
                 cj = _row_counter_loose_da3(gj)
                 ni, nj = sum(ci.values()), sum(cj.values())
                 if ni >= 1 and nj >= 1:
                     oi_j = _overlap_ratio_counter(ci, cj)
                     oj_i = _overlap_ratio_counter(cj, ci)
-                    # Short re-printed fragment (often 3–5 rows after a repeated header on same PDF page)
                     if (
                         nj >= 8
                         and 3 <= ni <= 14
@@ -3813,10 +3256,6 @@ _RE_CENTER_DIV_TWO_DA3_TABLES = re.compile(
 
 
 def _da3_row_credit_vs_debit_category(desc: str) -> str:
-    """
-    Generic classification for Date|Description|Amount rows (middle column only).
-    Used to detect OCR pasting a withdrawals table after a deposits heading.
-    """
     d = (desc or "").lower()
     if "debit card return" in d:
         return "credit"
@@ -3830,7 +3269,6 @@ def _da3_row_credit_vs_debit_category(desc: str) -> str:
         return "credit"
     if "lrm claims" in d or ("claims" in d and "adj" in d):
         return "credit"
-    # Credits / tax refunds that omit the word "deposit" (common on statement continuations)
     if "adp tax" in d and "customer" in d:
         return "credit"
     if "debit card" in d:
@@ -3853,7 +3291,6 @@ def _da3_row_credit_vs_debit_category(desc: str) -> str:
 
 
 def _da3_table_credit_debit_fractions(grid):
-    """Returns (debit_frac, credit_frac, unknown_frac, scored_row_count)."""
     deb = cre = unk = 0
     if not grid or len(grid) < 2:
         return 0.0, 0.0, 0.0, 0
@@ -3879,12 +3316,6 @@ def _da3_table_credit_debit_fractions(grid):
 
 
 def _drop_da3_misplaced_debit_after_deposit_heading_tables(text: str) -> str:
-    """
-    Fix 25: PDF order is withdrawals → then deposits. OCR sometimes emits two DA3 tables
-    after the deposits heading: real deposits, then a duplicate withdrawals block.
-    Drop the second table when it is overwhelmingly debit-like and the first is credit-like.
-    Repeats until no more matches (multiple duplicate blocks).
-    """
     if not text or "<table" not in text.lower():
         return text
     if "deposit" not in text.lower():
@@ -3916,7 +3347,6 @@ def _drop_da3_misplaced_debit_after_deposit_heading_tables(text: str) -> str:
             if n1 < 5 or n2 < 10:
                 pos = m.start() + 1
                 continue
-            # First table: deposits often have ADP/merchant lines scored as unknown — blend unk toward credit.
             credit_like_first = cr1 + 0.42 * unk1
             if credit_like_first < 0.30 and credit_like_first <= db1 + 0.05:
                 pos = m.start() + 1
@@ -3934,6 +3364,152 @@ def _drop_da3_misplaced_debit_after_deposit_heading_tables(text: str) -> str:
     return text
 
 
+# --------------------------
+# Fix 28: Strip credit-row prefix from debit DA3 table
+# --------------------------
+
+# Keywords that identify a section heading as debit/withdrawal-oriented.
+_DEBIT_HEADING_KEYWORDS = re.compile(
+    r"\b(withdrawal|withdrawals|debit|debits|charges?|subtraction|subtractions|"
+    r"payments?\s+made|money\s+out|paid\s+out|electronic\s+debit)\b",
+    re.IGNORECASE,
+)
+# Keywords that identify a section heading as credit/deposit-oriented.
+_CREDIT_HEADING_KEYWORDS = re.compile(
+    r"\b(deposit|deposits|credit|credits|addition|additions|interest\s+paid|"
+    r"money\s+in|paid\s+in|incoming|received)\b",
+    re.IGNORECASE,
+)
+# How many characters before a table start to scan for a section heading.
+_HEADING_SCAN_WINDOW = 600
+
+
+def _get_heading_before_table(text: str, table_start: int) -> str:
+    """
+    Return the last non-trivial text line in the _HEADING_SCAN_WINDOW chars
+    that precede `table_start`.  HTML tags are stripped before scanning so that
+    centered <div> headings are picked up correctly.
+    """
+    window = text[max(0, table_start - _HEADING_SCAN_WINDOW): table_start]
+    clean = re.sub(r"<[^>]+>", " ", window)
+    lines = [ln.strip() for ln in clean.splitlines() if ln.strip()]
+    for line in reversed(lines):
+        if len(line) > 3 and "---" not in line:
+            return line
+    return ""
+
+
+def _strip_credits_prefix_from_debit_da3_table(text: str) -> str:
+    """
+    Fix 28: When OCR prepends the rows of a preceding credits/deposits DA3 table
+    onto the top of a debits/withdrawals DA3 table, strip that prefix.
+    Safety guards (ALL must pass):
+      1. Both tables must be valid DA3 (Date|Description|Amount header).
+      2. Target (later) table must follow a debit/withdrawal heading.
+      3. Source (earlier) table must follow a credit/deposit heading.
+      4. Only the contiguous leading prefix of matching rows is removed.
+      5. At least 1 data row must remain in the target after stripping.
+      6. Strip count must be >= 1.
+      7. Entire function wrapped in try/except — returns input unchanged on error.
+    """
+    if not text or "<table" not in text.lower():
+        return text
+    try:
+        pattern = re.compile(r"<table[^>]*>.*?</table>", re.DOTALL | re.IGNORECASE)
+        matches = list(pattern.finditer(text))
+        if len(matches) < 2:
+            return text
+
+        # Collect all DA3 tables with their heading context.
+        da3_tables = []
+        for m in matches:
+            g = _parse_da3_grid_from_table_html(m.group(0))
+            if not g or not _is_da3_header(g):
+                continue
+            heading = _get_heading_before_table(text, m.start())
+            da3_tables.append({"match": m, "grid": g, "heading": heading})
+
+        if len(da3_tables) < 2:
+            return text
+
+        def _row_keys(grid):
+            keys = set()
+            for row in grid[1:]:
+                cells = [str(c or "").strip() for c in row]
+                if any(cells):
+                    keys.add(_transaction_row_dedupe_key(cells))
+            return keys
+
+        # For each DA3 table preceded by a debit heading, look backward for the
+        # nearest DA3 table preceded by a credit heading, and measure how many
+        # of the debit table's leading rows appear in that credit table.
+        drop_map = {}  # table match start -> number of rows to strip
+        for i in range(1, len(da3_tables)):
+            ti = da3_tables[i]
+            hi = ti["heading"].lower()
+            if not _DEBIT_HEADING_KEYWORDS.search(hi):
+                continue
+            gi = ti["grid"]
+            data_i = gi[1:]
+            if len(data_i) < 2:
+                continue
+            for j in range(i - 1, -1, -1):
+                tj = da3_tables[j]
+                hj = tj["heading"].lower()
+                if not _CREDIT_HEADING_KEYWORDS.search(hj):
+                    continue
+                gj = tj["grid"]
+                keys_j = _row_keys(gj)
+                if not keys_j:
+                    continue
+                # Count the contiguous leading rows of data_i that are in keys_j.
+                strip_count = 0
+                for row in data_i:
+                    cells = [str(c or "").strip() for c in row]
+                    k = _transaction_row_dedupe_key(cells)
+                    if k in keys_j:
+                        strip_count += 1
+                    else:
+                        break
+                if strip_count < 1:
+                    continue
+                # Guard: at least 1 data row must survive.
+                if len(data_i) - strip_count < 1:
+                    continue
+                drop_map[ti["match"].start()] = strip_count
+                log.info(
+                    "Fix 28: stripped %d leading credit row(s) from debit table at offset %d",
+                    strip_count, ti["match"].start(),
+                )
+                break  # found nearest credit table — stop looking further back
+
+        if not drop_map:
+            return text
+
+        # Apply the strip.
+        out_chunks = []
+        last = 0
+        for m in matches:
+            out_chunks.append(text[last: m.start()])
+            if m.start() in drop_map:
+                strip_n = drop_map[m.start()]
+                g = _parse_da3_grid_from_table_html(m.group(0))
+                if g and _is_da3_header(g) and len(g) - 1 - strip_n >= 1:
+                    new_grid = [g[0]] + g[1 + strip_n:]
+                    out_chunks.append(_grid_to_html(new_grid))
+                else:
+                    out_chunks.append(m.group(0))
+            else:
+                out_chunks.append(m.group(0))
+            last = m.end()
+        out_chunks.append(text[last:])
+        return "".join(out_chunks)
+
+    except Exception as e:
+        log.warning("_strip_credits_prefix_from_debit_da3_table failed: %s", e)
+        return text
+
+
 def stabilize_tables_and_text(page_md: str) -> str:
     if not page_md:
         return page_md
@@ -3958,6 +3534,7 @@ def stabilize_tables_and_text(page_md: str) -> str:
     stabilized = _patch_da3_msft_g_ref_amounts_across_tables(stabilized)
     stabilized = _dedupe_subset_transaction_tables_html(stabilized)
     stabilized = _drop_da3_misplaced_debit_after_deposit_heading_tables(stabilized)
+    stabilized = _strip_credits_prefix_from_debit_da3_table(stabilized)  # Fix 28
     stabilized = _strip_orphan_continued_da3_flatline(stabilized)
     stabilized = _strip_standalone_continued_banner_line(stabilized)
     return close_unclosed_html(stabilized)
@@ -4062,10 +3639,6 @@ def run_ocr(uploaded_file):
 
             parts = []
 
-            # Keep GLM-OCR as the primary parser for all pages.
-            # Navy-specific helpers run only as additive post-processing.
-
-            # Header inclusion: PDF text -> band words -> OCR band
             hdr = ""
             if is_pdf:
                 hdr = extract_zone_text_pdf(path, page_num, 0, he)
@@ -4076,39 +3649,30 @@ def run_ocr(uploaded_file):
             if hdr and hdr.strip():
                 parts.append(normalize_html_tables(fix_account_number(normalize_money_glyphs(hdr.strip()))))
 
-            # Main OCR body: stabilize then patch missing duplicate rows from text layer
             if page_md and page_md.strip():
                 stabilized = stabilize_tables_and_text(page_md.strip())
-                # Fix 4: restore any rows OCR dropped by cross-checking the PDF text layer.
-                # Safe no-op for scanned PDFs (no text layer) and non-transaction pages.
                 if is_pdf and not is_nfcu_doc:
                     stabilized = _patch_ocr_with_textlayer(stabilized, path, page_num)
                     stabilized = _strip_orphan_continued_da3_flatline(stabilized)
                     stabilized = _strip_standalone_continued_banner_line(stabilized)
-                    # _patch_ocr_with_textlayer may append Case B rows from the PDF text layer
-                    # with fused UCB-style cells; run the same HTML table pass again (Fix 13/14).
                     stabilized = normalize_html_tables(stabilized)
                     stabilized = _trim_adjacent_da3_suffix_duplicating_next_table_prefix(stabilized)
                     stabilized = _split_ucb_deposits_electronic_sections_html(stabilized)
                     stabilized = _patch_da3_msft_g_ref_amounts_across_tables(stabilized)
                     stabilized = _dedupe_subset_transaction_tables_html(stabilized)
                     stabilized = _drop_da3_misplaced_debit_after_deposit_heading_tables(stabilized)
+                    stabilized = _strip_credits_prefix_from_debit_da3_table(stabilized)  # Fix 28
                     stabilized = _strip_standalone_continued_banner_line(stabilized)
                 elif is_pdf and is_nfcu_doc:
-                    # Navy-only additive fix: rebuild malformed summary table from
-                    # text layer, but keep GLM body/header/footer unchanged.
                     nfcu_summary = _extract_nfcu_summary_table(path, page_num)
                     if nfcu_summary:
                         stabilized = _replace_nfcu_summary_table(stabilized, nfcu_summary)
 
-                # Fix 4D: append Johnson Bank Checks + Daily Account Balance
-                # Run OUTSIDE _patch_ocr_with_textlayer so exceptions there don't block it.
                 if is_pdf:
                     try:
                         needs_checks = 'class="jb-summary-checks"' not in stabilized
                         needs_dab    = "daily account balance" not in stabilized.lower()
                         if needs_checks or needs_dab:
-                            # Try pdfminer first (column-aware), fall back to pymupdf
                             _txt4d = ""
                             try:
                                 from pdfminer.high_level import extract_text as _pm_extract
@@ -4128,12 +3692,10 @@ def run_ocr(uploaded_file):
                                 if jb_extra:
                                     stabilized = stabilized.rstrip() + "\n\n" + jb_extra
                     except Exception:
-                        pass  # safe no-op — never break other PDFs
+                        pass
 
                 parts.append(stabilized)
 
-            # Footer extraction: PDF text layer -> band words -> OCR crop
-            # Deduplication guard prevents double-printing when body OCR already captured it.
             if ENABLE_FOOTER_OCR and page_num < len(page_images):
                 ftr = ""
                 if is_pdf:
@@ -4145,7 +3707,6 @@ def run_ocr(uploaded_file):
                 if ftr and ftr.strip():
                     ftr_clean = normalize_money_glyphs(ftr.strip())
 
-                    # Guard 1: first-line dedup (original check)
                     ftr_first_line = next(
                         (ln.strip().lower() for ln in ftr_clean.splitlines() if ln.strip()),
                         ""
@@ -4154,11 +3715,6 @@ def run_ocr(uploaded_file):
                         ftr_first_line in part.lower() for part in parts
                     )
 
-                    # Guard 2: suppress footer that is a raw text-layer dump of
-                    # transaction rows (e.g. East West Bank two-column layout).
-                    # Detected when the footer contains 3+ date tokens (MM/DD or MM-DD)
-                    # AND the footer itself contains amounts — meaning it is transaction
-                    # data, not a legitimate page footer like an address or disclaimer.
                     _footer_date_re = re.compile(r"\b\d{1,2}[-/]\d{2}\b")
                     _footer_amt_re  = re.compile(r"\b\d{1,3}(?:,\d{3})*\.\d{2}\b")
                     _date_hits = len(_footer_date_re.findall(ftr_clean))
@@ -4172,7 +3728,6 @@ def run_ocr(uploaded_file):
                 all_pages.append("\n\n".join(parts))
 
         merged = "\n\n---page-separator---\n\n".join(all_pages) if all_pages else "(No content)"
-        # Cross-page: UCB section split + same transaction tables dedupe once on full doc.
         if all_pages and merged and not merged.startswith("Error") and "<table" in merged.lower():
             merged = _strip_orphan_continued_da3_flatline(merged)
             merged = _strip_standalone_continued_banner_line(merged)
@@ -4181,6 +3736,7 @@ def run_ocr(uploaded_file):
             merged = _patch_da3_msft_g_ref_amounts_across_tables(merged)
             merged = _dedupe_subset_transaction_tables_html(merged)
             merged = _drop_da3_misplaced_debit_after_deposit_heading_tables(merged)
+            merged = _strip_credits_prefix_from_debit_da3_table(merged)  # Fix 28
             merged = _strip_orphan_continued_da3_flatline(merged)
             merged = _strip_standalone_continued_banner_line(merged)
         return merged