merge_pages drops text lines — multi-page DetectDocumentText aggregation loses content

[cschanhniem]

Bug

kernel.merge_pages only merges key_values, not lines:

def merge_pages(pages: list[Page]) -> Page:
    merged_kvs = [kv for p in pages for kv in p.key_values]
    return Page(width=pages[0].width, height=pages[0].height, key_values=merged_kvs)

If any caller (Pro or future OSS code) passes a multi-page text extraction result through merge_pages, all the lines from every page are silently dropped. Only key_values survive, so the merged Page would return zero text blocks for plain OCR use cases.

Expected

merge_pages should flatten both lines and key_values (and any other list fields on Page) from all input pages.

Suggested fix

def merge_pages(pages: list[Page]) -> Page:
    if len(pages) == 1:
        return pages[0]
    return Page(
        width=pages[0].width,
        height=pages[0].height,
        lines=[line for p in pages for line in p.lines],
        key_values=[kv for p in pages for kv in p.key_values],
    )

If Page gains more list fields in future, they should be merged here too — or the function could iterate over Page's dataclass fields automatically.

Impact

Currently merge_pages is called from the Pro plugin path. The OSS endpoints pass the full pages list to the response builders directly, so OSS core is not currently affected. But a future aggregation endpoint or external caller using the kernel API would silently lose text.