papers: expand crawl with CSV/EML missed-paper seeds

[AmitMY]

What

Cross-referenced two external Google Scholar exports against the papers crawl to find sign-language papers we never reached, then fed the recoverable ones back in.

• Sources: csv-files.zip (Publish-or-Perish title exports for "sign language" + recognition/translation, pre-2000→2023) and eml-files.zip (223 "sign language" Scholar alert emails, 2025–26) — 5,585 unique candidate titles, ~44% not in the crawl.
• New scripts/resolve_seeds.py: resolves titles → Semantic Scholar paperIds via the title-match endpoint and injects new ids into state/frontier.json (resumable).
• Took the 1,568 missed post-2014 SL titles (≈257 non-SL false positives — semiotics, animal communication, generic gesture, audiology, spam — excluded via an LLM title judge), resolved 634 to SS ids, added 602 new seeds, and drained the crawl.

Result (state.tar.gz refreshed)

	Before	After	Δ
Total rendered papers	15,575	16,137	+562
Papers ≥ 2014	11,282	11,823	+541

Notes

• The expansion confirmed the existing crawl had already captured essentially the entire reachable SL citation closure: of 555 newly-expanded nodes, only 24 surfaced any new SL neighbor (37 total), and the BFS hit a fixed point in 9 iterations. The added papers were peripheral leaves nothing previously pointed to.
• The remaining ~933 unresolved titles are not indexed on Semantic Scholar (recent theses, local/regional journals) — unreachable by an SS-based crawl, left out by design.

🤖 Generated with Claude Code