graph_node_classification: a complete integer-keyed answer scores about 2 #46

Task

graph_node_classification

Claim

problem.md says prediction keys "can be strings or values convertible to strings", and test_nodes arrives as integers. The verifier looks up string IDs only, so a complete answer keyed by integers matches no node. Every quality metric is 0, and the run is still valid because the coverage check only counts keys.

Impact

The same predictions from an Opus 5.5 solution score 90.54-90.63 with string keys and 1.95-2.01 with integer keys (n=4). We know of no natural case: the starter and 105 solutions we hold all key by str(nid).

Evidence and how to reproduce

Change a solution's output keys from str(nid) to nid, and grade both with tools/judge_local.py. overall goes from 0.885740 to 0.

Proposed fix

Draft PR: #37.