Level-0 Opponent-Model Fitting¶

ImaginedOpponent.fit_level0 now reserves a validation holdout before training so level-0 performance is measured on unseen data instead of only the optimization set.

• Default split: 90% train, 10% holdout.
• Split timing: the partition is sampled once at the start of fit_level0, then each epoch trains only on the training subset.
• Return value: Level0FitMetrics, with per-split cross_entropy, accuracy, error_rate, and sample counts.
• Display helper: print_level0_fit_metrics(...) renders the metrics as a rich.Table so the training notebooks can show the split-wise results consistently.
• Small-buffer behavior: if the buffer has only one sample, training uses that sample and the holdout metrics are reported as NaN because a true holdout is impossible.

This is mainly meant to quantify level-0 imitation error honestly before higher-level imagined best-response training builds on top of it.