fix(knn): accept k=1 in KNNClassifier

[NotAFlightRisk]

Checklist

• My branch is up-to-date with main branch.
• Everything works and tested on latest stable Rust.
• Coverage and Linting have been applied (fmt and clippy)

Current behaviour

KNNClassifier::fit with with_k(1) comes back with k should be > 1, k=[1], so there's no way to get a 1-nearest-neighbour classifier, and a loop over k from 1 upwards falls over on the first one. KNNRegressor has taken k = 1 since 92dad01 relaxed its copy of the check, the classifier just never got the same change.

New expected behaviour

Same check as KNNRegressor, so k = 1 works and k = 0 is still an error. The new test fits with k = 1 and checks taht every training point gets its own label back.

It's two characters of fix really, the test is most of the diff.

Change logs

Changed

• KNNClassifier accepts k = 1

[Mec-iS]

Thanks for the fix, @NotAFlightRisk. I reviewed the diff (1 file, +12/−2) against main.

Summary: The change is correct and minimal. parameters.k <= 1 becomes parameters.k < 1, and the error text becomes k should be > 0. This is the same guard KNNRegressor::fit already uses (k < 1, "k should be > 0"), so the two estimators are now consistent. The downstream search code accepts k = 1 too: LinearKNNSearch requires k >= 1 && k <= len(data), and CoverTree::find only rejects k == 0. I don't see a reason for the old <= 1 restriction.

Suggestions (non-blocking):

1. Test the boundary. The new knn_fit_predict_k1 test covers k = 1, but nothing pins down k = 0 still failing. The description says "k = 0 is still an error", so a short test would guard that: assert!(KNNClassifier::fit(&x, &y, KNNClassifierParameters::default().with_k(0)).is_err());.
2. Strengthen the k=1 test. Five points with labels [2,3,2,3,2] only check that each training point returns its own label. That is fine, but it would also be worth running it against the non-default algorithms (KNNAlgorithmName::CoverTree and LinearSearch), since the two backends handle k independently. A predict_proba check at k = 1 (one-hot rows) would also be cheap and cover the second public entry point.
3. CHANGELOG. Other recent fixes (predict method on DecisionTreeRegressor panics when tree is not fit #469/Decision tree panic #470) have entries in CHANGELOG.md. The PR description has a "Change logs" section but the file isn't touched. Please add a Fixed line under Unreleased. Because the error message string changes (> 1 → > 0), a downstream user matching on it would be affected, so it's worth a mention, as was done for MultiClassSVC.
4. Docs. Please check whether the KNNClassifierParameters::k doc comment or the module docs say "k > 1" anywhere, and update them if so.
5. Nit. There is a typo in the PR description: "taht" → "that".

CI: lint, MSRV (1.85), the check_features jobs and i686-linux tests have passed. The remaining test jobs (linux, macOS, Windows, wasm) and coverage were still running when I looked. Please confirm they are green, in particular that no existing test asserted the old k <= 1 failure.

With the k = 0 test and the CHANGELOG entry added, this looks good to merge from my side. cc @Mec-iS

[Mec-iS]

Thanks for the fix!