NDArray M2: multi-dimensional slicing with :: syntax

[Copilot]

Extends NDArray[A] with a varargs apply(selectors: RangeExtender*) that brings the same ergonomic slicing already available on Matrix to N-dimensional arrays.

API

val cube = NDArray(data, Array(4, 5, 3))

cube(::, 1 until 3, ::)       // shape [4, 2, 3] — zero-copy view
cube(0 until 2, ::, ::)       // shape [2, 5, 3] — zero-copy view
cube(::, ::, Array(0, 2))     // shape [4, 5, 2] — gather (copy)

RangeExtender = Range | Array[Int] | ::.type is reused as-is from rangeExtender.scala.

Implementation (vecxt/src/ndarrayOps.scala)

• Zero-copy path — when every selector is ::, a step-1 Range, or a contiguous Array[Int]: returns a strided view sharing the backing array, adjusting only offset and shape; strides are unchanged.
• Gather path — when any selector is non-contiguous: resolves all selectors to Array[Int] and iterates output indices in col-major order to fill a fresh contiguous array.
• Bounds validation always on (varargs precludes transparent inline): validates selector count matches ndim, range start/end within shape, and each Array[Int] element within shape.

Note: a single Array[Int] on a 1-D NDArray (arr(Array(0,3,7))) resolves via Scala's overload rules to the existing N-D element-access apply(indices: Array[Int]), not the new slicing overload. Multi-argument calls (e.g. arr(Array(0,2), ::)) are unambiguous.

Tests (vecxt/test/src/ndarraySlicing.test.scala)

Covers 1-D, 2-D, and 3-D cases across all selector types; zero-copy mutation visibility; consistency with slice; and error paths (wrong selector count, out-of-bounds range, out-of-bounds gather index).