silhouette_samples(metric='precomputed') silently ignores metric_params, n_jobs, verbose

[SomSamantray]

Problem

tslearn.clustering.silhouette_samples(X, labels, metric="precomputed", metric_params=..., n_jobs=..., verbose=...) silently drops all of metric_params, **kwds, n_jobs, and verbose. The docstring promises metric_params and **kwds are passed to the distance function, and silhouette_score raises on unexpected kwargs via sklearn forwarding, so the silent no-op is a contract inconsistency on the new API surface.

Suggested fix (design call)

Either (a) raise TypeError for any non-None metric_params/n_jobs/verbose/**kwds on the precomputed branch (loud rejection, parity with silhouette_score), or (b) explicitly document that these are ignored when metric == "precomputed".

Metadata

Severity: P3
Confidence: 75
Reviewer(s): api-contract, adversarial (ce-code-review run 20260821-221628-4e205155 on PR #703)
Finding ID: tslearn/clustering/utils.py:330 silhouette_samples(metric="precomputed") silently drops metric_params

[charavelg]

Neither tslearn silhouette_score or sklearn one raise on unexpected kwargs with metric=precomputed. When precomputed is used, no distance function is used.

[ApparentlyTejas]

I'd like to take this. Looking at the code, silhouette_samples(metric="precomputed") returns early before metric_params/n_jobs/verbose/**kwds are ever used, so they're silently dropped — no distance function runs on a precomputed matrix, so there's nothing for them to configure. This matches sklearn's own silhouette_samples(metric="precomputed"), which behaves the same way (per the discussion on #713).

Given that, I'm going with option (b): document the ignored behavior rather than raising, for parity with sklearn and with silhouette_score's existing precomputed handling. Concretely:

• Docstring notes on metric_params, n_jobs, verbose, and **kwds stating they're ignored when metric="precomputed".
• An inline comment at the early-return explaining why.
• A regression test asserting that passing metric_params/n_jobs/verbose alongside metric="precomputed" doesn't change the result.

Opening a PR shortly.