fix: fold duplicate records of one work into one

[hedmana]

Closes #61

What

• graph.build.clean folds duplicate records into one. It keeps the most cited record, together with every citation and any abstract the kept record lacks. The report counts these as merged_nodes.
• Versions are detected automatically: one title, one first author's family name, and years at most two apart after re-dating. A title alone is not enough, because Deep learning is both LeCun 2015 and Goodfellow 2016. A given name is too strict, because DiffPool's first author is both Rex and Zhitao Ying.
• A new field-config key, duplicate_works, names records filed under another work's metadata. In gnn.yaml these are GCN, GraphSAGE, GIN, VGAE, Attention Is All You Need, Adam, BERT, Hamilton et al. 2017, APPNP (under its first author's later name), and a 2020 book chapter that OpenAlex credits with GAT's citations. Each one was identified by what its citers cite on Semantic Scholar.
• GCN's record W2964015378 moves from exclude_works to duplicate_works. Pitfalls and Neural Combinatorial Optimization get their titles back; NCO's authors are still wrong, since no author correction exists. W4292779060 is excluded, because it has too few citers to identify.
• The Deep Graph Infomax gold target moves to its kept record.

Why

About 450 crawled works sat in the graph twice, which split each work's citations and could rank a work twice. For example, ChebNet had 1446 + 435 citations inside the graph and DiffPool 415 + 414. Excluding GCN's mis-filed record discarded 2458 citations.

On the rebuilt snapshot gnn-20260930, the graph goes from 10051 works to 9599, and GCN's citations inside the graph go from 2453 to 3779.

	dev found	dev nDCG@10	test found	test nDCG@10
learned, before	23/32	0.695	17/31	0.542
learned, after	23/32	0.690	16/31	0.509

This is a metadata fix, not tuning, and the ranking does not improve. The proxy gate still holds, with 15/15 paired MRR wins over logistic on both samplers (+0.023 uniform, +0.014 year-matched). Adam, no longer split across two records, enters GGSNN's top 5, which is the known problem of tools competing with originators.

Verify

uv run pytest tests/test_build.py
uv run python -m graph.build --field gnn   # merged_nodes 463
uv run python -m gnn.lineages --field gnn

Generated with Claude Code