✔️ Linting Passed

All linting checks passed. Your pull request is in excellent shape! ☀️

_{Generated for commit:44f9242. Link to the linter CI:here}

Thanks for this. The introduction ofcollinear_free_mask() is a particularly elegant solution. Using the geometric cross product to detect collinearity is a robust and generalizable method, and it correctly avoids the limitations of the previous second-difference heuristic, which could miss redundant points in many realistic ROC segments.

One minor suggestion: ensure thatfps is also prepended with0 before thedrop_intermediate logic (as is done fortps andthresholds), in case it's not already handled earlier in the flow. This keeps all arrays aligned and avoids subtle mismatches during masking.

Some minor linting issues remain; once those are resolved, this should be good to go. Do squash your commits also...

Hi !

This PR fixes two key bugs insklearn.metrics.roc_curve whendrop_intermediate=True:

1. Incorrect point retention due to prepending (0,0) after dropping intermediates
2. Faulty collinearity checks that only removed perfectly midway points

Bug 1: Incorrect Order of Operations
Before Fix:

Thresholds: [inf, 7, 4, 0]FPR: [0.0, 0.0, 0.0, 1.0]  ❌TPR: [0.0, 0.25, 1.0, 1.0] ❌ROC Curve:TPR1.0 |             ● (1,1)     |          /     |        ● (0,1)     |      /      |     ● (0,0.25)   ← Redundant point     |    /     |   ● (0,0)     +------------------ FPR        0.0          1.0

After Fix:

Thresholds: [inf, 4, 0] ✓FPR: [0.0, 0.0, 1.0] ✓TPR: [0.0, 1.0, 1.0] ✓ROC Curve:TPR1.0 |          ● (1,1)    |         /    |        ● (0,1)   ← Direct jump from (0,0)    |           |   ● (0,0)    +------------------ FPR     0.0              1.0

Problem: Intermediate points were dropped before adding (0,0), causing redundant points to remain.
Fix: We now add (0,0) FIRST before checking for intermediates.

Bug 2: Faulty Collinearity Detection
Before Fix (Horizontal Segment):

TPR1.0 |          ● (1,1)     |          |      |          ● (0.6,1)  ← Should be removed     |          |      |          ● (0.3,1)  ← Should be removed     |          |      |          ● (0.1,1)  ← Should be removed     |          |      |       ● (0,1)     |   ● (0,0)     +------------------ FPR      0.0  0.1 0.3 0.6 1.0

After Fix:

TPR1.0 |          ● (1,1)   |          |   |          |   |          |   |       ● (0,1)   ← Only essential points remain   |   ● (0,0)   +------------------ FPR    0.0              1.0

Problem: Old method only removed points exactly midway between neighbors.
Fix: New geometric test detects ANY collinear points using vector cross-product:

Before Fix (Diagonal Segment):

TPR1.0 |                  ● (1,1)     |               ● (0.9,0.9)     |           ● (0.7,0.7)     |       ● (0.4,0.4)     |   ● (0,0)     +------------------ FPR      0.0              1.0

After Fix:

TPR1.0 |                  ● (1,1)    |               /    |             /    |           /    |         /    |       /    |     /    |   ● (0,0)    +------------------ FPR     0.0              1.0

Thanks !