Adaptive per-call test set row limit at predict time

[klemens-floege]

What

The API now accepts much larger test sets, and the static row/cell limits the client enforces already come from get_model_limits. What was missing client-side is the adaptive part: prediction compute scales with n_train_rows * n_test_rows, so the effective per-call test row limit shrinks as the fitted training set grows.

• Add PREDICT_ROW_PAIRS_BUDGET (250k × 1M): at predict time the test row limit is min(test_set_max_rows, budget // n_train_rows) — e.g. 1M training rows → 250k test rows per call, ≤ 250k training rows → the static 1M cap binds.
• validate_test_set takes an optional train_rows; both TabPFNClassifier and TabPFNRegressor pass the fitted training set size, so oversized test sets fail fast client-side with a clear "split across multiple calls" message instead of erroring after upload.

train rows	test row limit per call
1M	250k
500k	500k
≤ 250k	1M (static cap)

Tests

Unit tests cover the static cap, the adaptive limit at the 1M-train anchor, the adaptive limit never exceeding the static cap, and the offline no-op when limits are unavailable.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 208dc36. Configure here.}