fix: give training the patience to outlast its early popularity peak

[hedmana]

Closes #43

What

• TrainConfig defaults go from 400 epochs and patience 40 to 600 epochs and patience 100.
• PLAN.md's expanded-snapshot results now come from the served snapshot, trained with these defaults.

Why

On the expanded snapshot, one step from a random start ranks by popularity alone, and validation AP reads 0.899. Learning more dips below that before rising past it: 0.86 at epoch 2, 0.90 at 50, 0.92 at 150, 0.93 at 300. With patience 40, training kept epoch 1 and stopped at 41. The served model was effectively untrained.

With the new defaults, gnn.evaluate on the served snapshot, paired over 15 runs against the logistic baseline:

	uniform	year-matched
MRR	+0.024 ± 0.008, 15/15	+0.013 ± 0.003, 15/15
AUC	+0.015 ± 0.002, 15/15	+0.030 ± 0.006, 15/15

The plan had recorded that the model only tied popularity under year-matched negatives. That came from the undertrained model. On the old snapshot, training already settled around epoch 100 and is unaffected.

Verify

uv run python -m gnn.train --field gnn        # best_epoch well past 1 in metrics.json
uv run python -m gnn.evaluate --field gnn     # about 8 minutes

Generated with Claude Code