perf: break citation cycles one tangle at a time

[hedmana]

Closes #47

What

acyclic breaks each strongly connected component in one pass. It lays the component's works out newest first and keeps only the citations running forward in that order, so a citation of a later work, the one most at odds with publication order, is still what goes first. Among works of the same year, the one its peers cite most is placed as the earliest.

Why

The old loop ran find_cycle over the whole ancestry once per removed edge. Every ranking on the expanded snapshot paid for that, including each /lineage request.

Ancestry	before	after
GraphSAGE, served snapshot	5.6 s	0.10 s
GCN, served snapshot	6.1 s	0.08 s
GraphSAGE, with recovered citations	12.2 s	0.11 s

It removes slightly more edges than before, 237 against 202 of GraphSAGE's roughly 40,000. The old loop's picks among same-year cycles depended on the order find_cycle happened to walk. The result is a DAG either way.

Verify

uv run pytest tests/test_score.py tests/test_originators.py tests/test_api.py
curl localhost:8000/lineage/gnn/W2624431344   # well under a second

Generated with Claude Code