fix: fail closed on incoherent monthly citations

[gabrielspadon]

Monthly audit finding

The August refresh converged mechanically, but its generated PR still contained semantic citation conflicts and unsafe merge decisions. This source-only change fixes the decision paths. No generated .bib file is edited.

Source fixes

• Keep rejected candidate DOIs out of disk deduplication.
• Reprocess published DOI entries whose URL still points to a different preprint.
• Remove unconfirmed Scholar container metadata after a validated arXiv DOI.
• Treat Crossref proceedings volumes as books rather than papers.
• Match abbreviated author initials without conflating different authors.
• Force every author directory involved in a same-DOI conflict to re-enrich.
• Reject finalization if material same-DOI conflicts remain anywhere in the corpus or the a2i2 aggregate.

Audit evidence

• Monthly run 32494455377 succeeded and converged in three passes.
• Generated PR data: monthly pipeline refresh 2026-08 #48 has 503 BibTeX path changes and all structural CI green.
• Semantic review found rejected-candidate DOI leakage, stale published metadata, a contaminated arXiv record, one proceedings volume misclassified as a paper, and unresolved same-DOI conflicts.
• The deletion of Zero-Shot Out-of-Distribution Detection with Feature Correlations is retained because authoritative metadata identifies it as an unpublished 2019 record, outside the configured 2020 contribution window.

Verification

• ruff check citeforge/ tests/ main.py
• mypy citeforge/ main.py
• pytest -m "not live"
• 2,074 passed, 2 skipped, 7 deselected
• No output/ changes