Problem

predict_merges merges at most one successor per TEXT element. After recording merges[A] = [B] it sets curr_ind = B and moves on, so C is processed as a fresh element rather than being appended to A's merge list.

If A continues into B and B continues into C by the same criteria, all three belong to the same paragraph. The current code produces two short paragraphs (A+B and C+D) instead of one.

This becomes visible with PDFs that have many small orphan clusters — for example scanned pages with a native text layer, where each line of text is its own cluster. The pairwise limit leaves every other line as a separate paragraph, causing unnatural blank lines throughout the output.

Fix

Replace the single if-block with a while loop that keeps extending the merge chain from the current tail (check_ind) as long as the continuation condition holds:

merge_list: List[int] = []
check_ind = ind

while True:
    # find next non-skippable element from check_ind
    ...
    m1 = re.fullmatch(r".+([a-z,\-])(\s*)", sorted_elements[check_ind].text)
    m2 = re.fullmatch(r"(\s*[a-z])(.+)", sorted_elements[ind_p1].text)
    if m1 and m2:
        merge_list.append(sorted_elements[ind_p1].cid)
        curr_ind = ind_p1
        check_ind = ind_p1
    else:
        break

if merge_list:
    merges[elem.cid] = merge_list

The existing _readingorder_elements_to_docling_doc already iterates over the full merge list per element, so no changes are needed there.

Also fixes the skip-loop element-vs-label comparison (sorted_elements[ind_p1].label in skip_labels instead of sorted_elements[ind_p1] in [...] which always evaluated to False — also submitted separately as #153).