fix(text): keep unique /PlacedPDF bodies instead of suppressing them

[ajbufort]

Problem

InDesign wraps placed artwork in /PlacedPDF marked content, and the text extractor suppresses that text to drop decorative overlays and duplicated draft galleys (the PMC8100493 case). The keep/suppress decision in placed_pdf_text_dominates used only a character-count ratio:

placed_chars >= 800 && other_chars.saturating_mul(3) < placed_chars

i.e. keep placed text only when it is substantial and the non-placed text is a small (< 1/3) minority of it.

That heuristic misfires on spreads where the publisher places the entire page body inside /PlacedPDF (a floor-plan / marketing PDF: room labels, dimensions, disclaimer) but a comparable amount of unique text also sits outside the region (headers, callouts). The 3:1 dominance test fails, so the whole page body is suppressed - extract_text / extract_spans return nothing, while pdftotext and PyMuPDF extract the full page.

Concretely, on a 6-page InDesign floor-plan spread, two whole pages returned 0 spans where poppler returns 768 and 825 words. extract_chars already recovered ~3400 chars on those pages, so the loss was purely in this suppression gate.

Fix

The char-count ratio is a crude proxy for the real signal, which is duplication: a draft-galley overlay repeats the surrounding text, whereas a placed page body is mostly unique. This keeps the two cheap gates and replaces the ratio's second job with an actual duplication check:

1. placed_chars < 800 -> suppress (decorative figure; unchanged).
2. placed text dominates the page (other*3 < placed) -> keep (whole-body-placed, MATEC; unchanged).
3. otherwise -> keep unless a majority (>= 50%) of the placed word tokens also appear in the non-placed text (a duplicate overlay).

Gate 3's tokenisation runs only after gates 1 and 2, so the common single-column path stays allocation-free.

Measured

On a 468-doc consensus content-F1 sample: 3 docs recovered (a placed floor-plan 0.51 -> 0.99, another 0.89 -> 1.00, a third +0.01), no regressions, mean +0.0013. The change only ever affects pages carrying the /PlacedPDF tag.

Tests

• test_placed_pdf_kept_when_unique_body_amid_comparable_outside - the recovered case.
• test_placed_pdf_suppressed_when_large_duplicate_overlay - a full-size placed duplicate still suppressed.
• Existing placed_pdf tests (MATEC keep, minority-overlay suppress, no-op without tag) unchanged and passing.

Full gate on this branch off main: cargo test --lib (5744 passed), cargo test --doc (139 passed), cargo clippy --lib clean, cargo fmt --check clean.

[yfedoseev]

Approved — reviewed and regression-passed; in the merge queue. (Formal approval to supersede the earlier review state and record sign-off.)