Fix encoding OOV category.

[trivialfis]

Close #12189

[Copilot]

Pull request overview

Fixes a correctness bug in the ordinal (auto) re-coder for native categorical features where out-of-vocabulary (OOV) categories could be silently mapped to an in-range training category when lower_bound returned an insertion point rather than an exact match.

Changes:

• Add exact-match verification after lower_bound in CPU and CUDA SearchSorted implementations (string + numeric) to properly detect OOV categories in-range.
• Add CPU and GPU Python tests covering OOV categories whose sorted insertion point falls within the training set.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/encoder/ordinal.h	Adds post-lower_bound equality check to avoid returning an insertion point as a match (CPU path).
src/encoder/ordinal.cuh	Mirrors the CPU fix on the CUDA path for both string and numeric category searches.
python-package/xgboost/testing/ordinal.py	Introduces run_cat_oov_in_range test helper validating error behavior for in-range OOV categories.
tests/python/test_ordinal.py	Wires the new OOV-in-range test into the CPU test suite.
tests/python-gpu/test_gpu_ordinal.py	Wires the new OOV-in-range test into the GPU test suite.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[RAMitchell]

I am still not super happy about us doing recoding.

[trivialfis]

I am still not super happy about us doing recoding.

Yeah, at some point I gave up arguing against that.