pdf: first data row of a continuation table is silently dropped when the header is not repeated #214

Main previously rechecked 2026-10-29 at e7178e0 (Cargo.lock unchanged, pdf-inspector 1.14.2). Current PyPI release: firecrawl-anydoc 0.2.4. The current-main result below was reproduced with the Rust core example, not a rebuilt Python wheel.

Latest release independently rechecked 2026-10-30: firecrawl-anydoc 0.2.4 installed into a new isolated Python 3.14.4 venv. Running check_rowdrop.py reproduced both noheader modes missing 008/016, while both repeated-header controls retained all 24 rows. Main remains e7178e0; its earlier source-build evidence was not rerun.

Problem
A table across three table-only PDF pages has its header on page 1 only. Conversion silently removes the first data row on pages 2 and 3. Repeating the header on each page prevents the loss. Both ruled and unruled tables show the same result.

Minimal reproduction
Save the inline stdlib-only generator below as mkpdf.py. All fixture data are synthetic.

python3 mkpdf.py noheader.pdf noheader
python3 mkpdf.py header.pdf header
python3 mkpdf.py noheader_unruled.pdf noheader_unruled
python3 mkpdf.py header_unruled.pdf header_unruled
cargo build --locked --example convert
target/debug/examples/convert noheader.pdf

Expected: all 24 rows, Widget 000 through Widget 023, exactly once.
Observed on current main:
noheader: 22 rows; missing Widget 008 and Widget 016
header: 24 rows; none missing
noheader_unruled: 22 rows; missing Widget 008 and Widget 016
header_unruled: 24 rows; none missing
No warning or conversion error was emitted.

Likely owning component
src/formats/pdf.rs delegates Markdown extraction to pdf-inspector; this PDF path bypasses anydoc's shared table/header model. In pdf-inspector 1.14.2, src/markdown/convert.rs, merge_continuation_tables (lines 598-607) unconditionally skips the first two Markdown lines of each continuation table. The first line can be data rather than a repeated header. This matches the deterministic row loss above.
A possible repair is to keep the first row when it does not match the original table header, while removing the Markdown separator. A dependency update may be the right integration route if this is already repaired upstream; newer pdf-inspector versions were not tested here.

Duplicate check (2026-10-30)
Reviewed the complete paginated anydoc issue/PR inventory: 208 unique records, including 126 open. No matching continuation-row-loss report was found in titles/bodies. Issue #173 concerns multi-column reconstruction, not this first-row loss. Merged PR #196 changes shared CSV/xlsx/ods header voting and does not affect the PDF path. Open PRs #175 and #181 update pdf-inspector for RTL extraction; they were not built or claimed to fix this bug.

Synthetic PDF generator

"""Write a synthetic multi-page ruled table PDF (stdlib only, no PII).
usage: mkpdf.py OUT MODE   MODE: noheader | header | noheader_unruled | header_unruled"""
import sys
out, mode = sys.argv[1], sys.argv[2]
ruled = 'unruled' not in mode; rep_header = mode.startswith('header')
COLS = [50, 170, 260, 400, 520]; W = 560; RH = 24
HEAD = ['Item', 'Version', 'Region', 'Start', 'End']
def row(n): return [f'Widget {n:03d}', f'{1+n%3}.{n%7}', 'XX', f'{1+n%12}/{1+n%28}/2025', f'{1+(n+5)%12}/{1+n%28}/2026']
pages=[]; n=0
for p in range(3):
    rows=[HEAD] if (p==0 or rep_header) else []
    for _ in range(8): rows.append(row(n)); n+=1
    ops=['BT /F1 10 Tf ET']; y=740
    for r in rows:
        if ruled:
            for i,x in enumerate(COLS):
                x2 = COLS[i+1] if i+1<len(COLS) else W
                ops.append(f'{x-4} {y-7} {x2-x} {RH} re S')
        for x,c in zip(COLS,r): ops.append(f'BT /F1 10 Tf {x} {y} Td ({c}) Tj ET')
        y-=RH
    pages.append('\n'.join(ops).encode())
objs=[]
def add(b): objs.append(b); return len(objs)
font=add(b'<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>')
pages_id=len(objs)+1+2*len(pages)
kids=[]
for c in pages:
    cid=add(b'<< /Length %d >>\nstream\n'%len(c)+c+b'\nendstream')
    kids.append(add(b'<< /Type /Page /Parent %d 0 R /MediaBox [0 0 612 792] /Resources << /Font << /F1 %d 0 R >> >> /Contents %d 0 R >>'%(pages_id,font,cid)))
assert add(b'<< /Type /Pages /Kids [%s] /Count %d >>'%(b' '.join(b'%d 0 R'%k for k in kids),len(kids)))==pages_id
cat=add(b'<< /Type /Catalog /Pages %d 0 R >>'%pages_id)
buf=bytearray(b'%PDF-1.4\n'); offs=[]
for i,o in enumerate(objs,1):
    offs.append(len(buf)); buf+=b'%d 0 obj\n'%i+o+b'\nendobj\n'
x=len(buf); buf+=b'xref\n0 %d\n0000000000 65535 f \n'%(len(objs)+1)
for o in offs: buf+=b'%010d 00000 n \n'%o
buf+=b'trailer\n<< /Size %d /Root %d 0 R >>\nstartxref\n%d\n%%%%EOF\n'%(len(objs)+1,cat,x)
open(out,'wb').write(buf)

Python release reproduction: install firecrawl-anydoc==0.2.4, generate each fixture as above, then call anydoc.to_markdown on each PDF.

Update 2026-10-30: pdf-inspector main bcf387d (latest release v1.25.2) still has the same unconditional skip in merge_continuation_tables (src/markdown/convert.rs, the line_idx >= 2 loop near line 670). This is from reading the code only. I did not build or run the newer version.