extract_attachment returned "unsupported" for .epub, so CPT 2021/2022
and CPT Changes 2023 had no text at all, and 2018/2019/2024 indexed
only from their PDF siblings. Added rex.comments.epub_text — stdlib
only, shared with pfs.cpt_epub — that resolves an EPUB's own reading
order via META-INF/container.xml -> the OPF's manifest + spine (falling
back to sorted .xhtml/.html names when container.xml is missing), and
strips tags/entities into plain text. extract_attachment's new .epub
branch returns the same ExtractResult shape the PDF branch does; no
change needed in llm.source._attachment_sections, which already tries
every bib attachment regardless of extension.
pfs.cpt_epub used to hard-code "OPS/" as the content-file prefix in
three places; it now resolves the same way via epub_text.content_root,
so a differently-templated EPUB would still locate its content instead
of silently parsing to nothing.
77fd48eb31