v0.9.2 — arXiv-ID↔title consistency + subtle wrong-verdict fixes

Fixed

• Misattributed arXiv IDs silently survived (and could rewrite an entry into an unrelated paper): an entry's cited arXiv ID was trusted without checking that the ID's actual paper matches the entry, so a wrong ID (copy-paste / lookup error) survived because title/author search VERIFIED the entry against the real paper from Crossref/DBLP/S2 — and the resolver's arXiv-ID-keyed stages would rewrite the entry into the unrelated paper at confidence = 1.0 (real cases: onebench2024's correct 2412.07689 "ONEBench…" replaced by 2412.06745 "RoboTron-Drive…"). bibtex-check now runs a pre-search _check_arxiv_id_consistency() that fetches the entry's own arXiv ID and reports FactCheckStatus.ARXIV_ID_MISMATCH when the fetched title diverges (gated by FactCheckerConfig.check_arxiv_consistency, default on, with arxiv_consistency_min_title=0.50). bibtex-update's Resolver._verify_arxiv_match() gates Stage 1 / 1b on MATCH_THRESHOLD like the search-based stages, so a wrong cited ID falls through to title-based resolution instead of corrupting the entry.
• DBLP homonym disambiguation suffix produced false author_mismatch: DBLP appends a 4-digit suffix to homonymous author names ("Yu Sun 0020", "Chuan Guo 0001"); dblp_hit_to_record took the number as the family name. The suffix is now stripped so the surname is the real family name (the digits move into the given name).
• Particle surnames produced false AUTHOR_MISMATCH / false HALLUCINATED and false resolver rejections: author comparison was asymmetric. The BibTeX entry side reduced a name to its last token, while the API-record side used the raw multi-word family field verbatim ("van den Oord" → van den oord), so Jaccard against the entry's oord was 0 and a correctly-cited paper by an author with a nobiliary particle (von, van der, de la, dos, …) was flagged as an author mismatch — depressing the score enough to risk a false HALLUCINATED in bibtex-check, and (combined with the new arXiv-match gate, below) falsely rejecting a correctly-resolved record in bibtex-update. last_name_from_person now reduces any family — including particles — to its final, most distinctive token, and every comparison site runs both the entry side and the API-record side through it: FactChecker._compare_all_fields / _score_candidate, and the resolver's five match-score sites via a new _record_surnames helper. Both citation styles now produce the same key symmetrically.
• extract_arxiv_id_from_text mistook numbers that merely look like a modern arXiv ID: any NNNN.NNNNN substring matched, so a DOI fragment (10.1234/5678.9012 → 5678.9012) or an arbitrary number was treated as an arXiv identifier, driving false preprint detection in bibtex-update and bogus arXiv lookups in bibtex-check / Zotero sync. A new is_valid_arxiv_id rejects modern IDs whose YYMM month is not 01–12, and extraction now scans all candidates (finditer) returning the first valid one. FactChecker._arxiv_id_from_entry applies the same month check to eprint.
• Legacy / .pdf arXiv URLs were truncated: ARXIV_HOST_RE stopped at the first /, so arxiv.org/abs/hep-th/9901001 yielded hep-th and arxiv.org/pdf/2602.01031v2.pdf leaked the v2.pdf suffix. The host pattern now captures the full path segment, and extraction strips a trailing .pdf and vN version and re-normalizes through ARXIV_ID_RE, so legacy category IDs survive intact and modern IDs are returned bare.
• Generic single-word venue aliases collapsed distinct sibling journals: get_canonical_venue("Nature Physics") substring-matched the alias "nature" and returned canonical "nature", conflating Nature with Nature Physics / Science with Science Robotics / PNAS with PNAS Nexus and masking a genuine VENUE_MISMATCH. Only the generic single-word journal names (nature, science, pnas) now require an exact match; acronym venues keep substring matching, so a track/suffix form ("NeurIPS Track" → neurips, "ICML 2021" → icml) and multi-word siblings ("Nature Communications" → nature_comm) still canonicalize correctly.

Internal

• De-duplicated the arXiv-record lookup: _check_arxiv_id_consistency and _query_arxiv_by_id now share a memoized FactChecker._arxiv_record(id) instead of fetching + parsing the same arXiv Atom feed twice per entry. Factored the record-surname extraction (updater._record_surnames) and the arXiv-ID normalization (extract_arxiv_id_from_text._normalize) out of duplicated call sites.

Tests

• +22 tests in tests/test_subtle_failures.py (particle-surname symmetry on both the fact-checker and resolver paths, impossible-month arXiv-ID rejection, legacy/.pdf URL extraction, generic-journal vs acronym venue matching); plus arXiv-ID↔title consistency, resolver _verify_arxiv_match, and DBLP-suffix/CoRR-venue regression tests. Full suite: 785 passed, 1 skipped (zero regressions).