Fix comprehensiveness/sufficiency crash on code-sequence inputs

[solarsys]

Why

ComprehensivenessMetric and SufficiencyMetric (and evaluate_attribution / Evaluator) crash for any model whose inputs are plain code-sequence tensors: MLP, Transformer, RETAIN and so on.

RuntimeError: Expected tensor for argument #1 'indices' to have one of the following scalar types: Long, Int; but got torch.FloatTensor instead (while checking arguments for embedding)

RemovalBasedMetric._apply_ablation handles StageNet's (time, values) tuples with an integer-safe path, but plain tensors are ablated with x * (1 - mask). The float mask turns the code ids into floats, so the embedding lookup fails. The "mean" strategy has the same problem, and averaging code ids would be meaningless even if it ran.

@fbonc proposed a fix for this in #1170, which was auto-closed as stale. This PR revives it. Credit to them for the diagnosis.

What

• pyhealth/metrics/interpretability/base.py:
  • New helper _ablate_codes sets the removed codes to the padding id 0 with torch.where, so the dtype is kept. It also keeps the existing guard: a sample whose codes would all be removed keeps its first code.
  • Both the tuple path and the plain-tensor path now use the helper for integer inputs, whatever ablation_strategy is set to.
  • Float inputs behave exactly as before.
• docs/api/metrics/pyhealth.metrics.interpretability.rst: explains how integer inputs are ablated.
• examples/interpretability/comprehensiveness_code_sequences.py: trains an MLP on synthetic code sequences and scores Integrated Gradients. It runs in seconds on CPU with no downloads. It prints comprehensiveness 0.987 and sufficiency -0.006, because the label is driven by a single code.

Tests

• New tests/core/test_interp_ablation_codes.py (5 tests):
  • the helper's dtype and padding behaviour;
  • the keep-one-code guard;
  • float inputs still follow the strategy;
  • evaluate_attribution with MLP and Transformer under all three strategies gives finite scores.
• The end-to-end tests crash with the error above on master and pass with this change.
• Existing interpretability tests pass: test_interp_metrics, test_interpret_ensemble, test_ig_gim, test_stagenet, test_stagenet_mha, test_shap, test_lime (178 tests, OK).
• tools/check_pr_rules.py passes.

🤖 Generated with Claude Code

[jhnwu3]

lgtm