silhouette_samples: ragged/NaN-padded input handled inconsistently across metric branches

[SomSamantray]

Problem

tslearn.clustering.silhouette_samples handles ragged (NaN-padded) time-series input inconsistently across metric branches:

• metric="euclidean" raises a raw ValueError: Input X contains NaN.
• metric="dtw" silently truncates NaNs and returns results.
• metric="softdtw" silently miscalculates over the NaN padding (mean shifts from 0.1795 to 0.1518 on ragged data) with no warning.

Suggested fix

Add a documented policy: either validate that all series are full-length (and raise a clear error for ragged input in every branch), or strip NaN padding consistently before all branches (including euclidean), and add tests covering ragged/NaN-padded input for every metric.

Metadata

Severity: P1
Confidence: 75
Reviewer(s): adversarial (ce-code-review run 20260821-221628-4e205155 on PR #703)
Finding ID: tslearn/clustering/utils.py:358 ragged and NaN-padded time-series input

[charavelg]

@SomSamantray, It is expected that

metric="euclidean" raises a raw ValueError: Input X contains NaN.

whereas dtw and softdtw support variable length datasets.

Regarding,

metric="softdtw" silently miscalculates over the NaN padding (mean shifts from 0.1795 to 0.1518 on ragged data)

I cannot reproduce this behaviour. Can you share some code that showcases the problem plz?