fix: heading leakage in flush — clear pending_headings unconditionally; strip trailing stars from list-item headings

fix: parser fidelity — pending buffer, refs capture, heading chain, hygiene, caption gate, length basis

feat: top-10 retrieval pool; plaintext/markdown manuscript input

fix: filter picture text, strip markdown debris, merge open-sentence continuations

fix: bypass pymupdf4llm layout analysis on vector-heavy pages

fix: drop author/affiliation front matter; show first three authors et al.

feat: paragraph segmentation and parse_pdf