Overview

This PR addresses Issue #53: #53

Problem: roc_curve can return a threshold > 1 when y_score are probability estimates in [0,1], due to prepending a sentinel threshold using max(y_score) + 1.

Fix

• In sklearn/metrics/_ranking.py::roc_curve, replace the prepended threshold logic with dtype-aware handling:
  • For float scores: use np.nextafter(thresholds[0], np.inf, dtype=thresholds.dtype) to set a value strictly greater than the max score, minimally exceeding the domain when max == 1.0.
  • For integer scores: retain max + 1 behavior.
• Docstring updated accordingly.

Non-regression tests

Added to sklearn/metrics/tests/test_ranking.py:

1. test_roc_curve_thresholds_probabilities_below_one — when max(y_score) < 1, the first threshold is the next representable float above the max and thresholds remain within [0,1].
2. test_roc_curve_thresholds_probability_one — when max(y_score) == 1.0, the first threshold is np.nextafter(1.0, +inf); mapping remains consistent.
3. test_roc_curve_thresholds_integer_scores — for integer/non-probability scores, the first threshold is max(score) + 1 and monotonic decrease holds.

Reproduction steps and observed failure (pre-fix)

• Command used (minimal selection):

pytest sklearn/metrics/tests/test_ranking.py -k "thresholds_probabilities_below_one or thresholds_probability_one or thresholds_integer_scores"

• Pre-fix observed failure excerpt:

============================= test session starts ==============================
platform linux -- Python 3.11.14, pytest-9.0.2, pluggy-1.6.0
collected 207 items / 204 deselected / 3 selected

test_ranking.py FFF                                                      [100%]

=================================== FAILURES ===================================
______________ test_roc_curve_thresholds_probabilities_below_one _______________
...
E       assert 1.9 == 0.9000000000000001 ± 9.0e-07
...
__________________ test_roc_curve_thresholds_probability_one ___________________
...
E       assert 2.0 == 1.0000000000000002 ± 1.0e-06
...
___________________ test_roc_curve_thresholds_integer_scores ___________________
...
E       assert 4 == (np.int64(3) + 1)
...
====================== 3 failed, 204 deselected in 0.65s =======================

Validation (post-fix)

• The same test selection passes locally after the fix.

Notes

• CI will not be manually triggered per task rules; this PR targets branch scikit-learn__scikit-learn-26194 and will remain unmerged pending review.
• This change preserves API semantics: thresholds remain strictly decreasing; the mapping score >= thresholds[i] remains valid.

PR created per request; title includes the required token.