[tannevaled]

What

The go-ndarray playground shows each pipeline as Go | Ruby | NumPy and runs the Ruby column in the browser through cmd/wasm. This PR implements the Ruby NDArray API its code generator emits (the playground contract), in internal/vm/ndarray.go.

Area	Added or changed
Creation	arange(stop) / (start, stop) / (start, stop, step), linspace(start, stop, num = 50), eye(n, m = n, k = 0), zeros()/ones() (0-d with no shape; a shape Array is accepted too), full(value, *shape)
Views	reshape (with -1), T, squeeze(*axes), expand_dims(axis), dup/copy (a copy of the data)
Indexing	a[...], NumPy basic indexing: Integer, true for :, a Range (1...3, 1.., ...3, inclusive 1..2 / 1..-1) or an Enumerator::ArithmeticSequence ((0..) % 2, (1...7) % 3, (-1..) % -1, (5...1) % -2, 1.step(7, 3)). All-Integer, one per axis, returns the element as a Float.
Elementwise	square, ** (scalar; array with broadcasting; 2 ** a through NDArray#coerce), clip, NDArray.maximum/minimum (a number on either side broadcasts)
Reductions	sum/mean/max/min/prod/argmax/argmin take (axis, keepdims:); with no axis they return a Float, or an Integer flat index for argmax/argmin. cumsum / cumsum(axis)
Products and joins	NDArray.outer, NDArray.concatenate([...], axis:), NDArray.stack([...], axis:)
Inspecting	p a / inspect / to_s print go-ndarray's NumPy-style Repr (array([[0., 1., 2.],\n [3., 4., 5.]])) instead of Array(shape=[2 3], data=[...])

Bugs fixed

• NDArray.arange(5) raised ArgumentError: arange: runtime error: index out of range [1] with length 1, which was a recovered Go panic. NDArray.full() failed the same way. Every method now checks its arguments. A wrong count raises ArgumentError (MRI's wrong number of arguments (given n, expected m)), and a wrong type raises TypeError. A go-ndarray error is re-raised with go-ndarray's own message: IndexError for an index or axis out of range, ArgumentError for the rest.
• TestNativeFaultBecomesRubyError used NDArray.arange(6) as its example of a faulting native. It now lives in native_fault_internal_test.go and plants its own faulting native, so the safety net no longer depends on a binding staying broken.
• The README said in three places that the browser target GOOS=js GOARCH=wasm "does not build". It has built since internal/vm, ci: fix the js/wasm build, and make the guard that caught it reachable #684. readme_wasm_lane_test.go now compares a flattened README, so the reworded claims are caught across line breaks and bold markers. Checked: the test fails on the old README.

Dependency

github.com/go-ndarray/ndarray goes from v0.3.0 to v0.11.0, and github.com/go-fft/fft from v0.1.14 to v0.23.0 with it. This is done here rather than through the grouped Renovate PR #792, which is still dirty and is not touched.

How the expectations were derived

• Every expected value in ndarray_test.go was worked out by hand from NumPy's definition of the operation. None was taken from the binding's own output.
• TestNDArraySlicing walks the contract's indexing table row by row, plus inclusive ranges, negative steps, clamping, empty slices and multi-axis indexing. Each row holds three things that must agree: the Ruby index, the hand-derived NumPy elements and shape, and the same slice taken through go-ndarray's Go API (Slice(A/All/R/Rng/From/To/Step)).
• TestNDArrayMatchesGoOracle compares p output with go-ndarray's Repr of the same computation done in Go. go-ndarray's Go API is the oracle the playground's Go column uses.
• array ** array has no go-ndarray counterpart. It is checked against elementwise math.Pow after broadcasting.
• TestNDArrayArgumentChecks calls every method with too few arguments, too many, and wrong types. It asserts the exception class and message, and that no message contains runtime error.
• ndarray.go is at 100% coverage. The ToS row leaves both BELOW100 records.
• web/smoke.mjs now runs an NDArray program in the JS host, which is the playground's use case.

Known differences from NumPy (go-ndarray's, not the binding's)

• Repr formats floats in their shortest form, so linspace(0, 1, 5) prints array([ 0., 0.25, 0.5, 0.75, 1.]). NumPy prints array([0. , 0.25, 0.5 , 0.75, 1. ]). Large arrays also get a , shape=(...) suffix.
• clip(lo, hi) with lo > hi raises ArgumentError (go-ndarray's error). NumPy returns hi everywhere.
• argmax(axis) returns an int64 array, as NumPy does. to_a still yields Floats, as the contract requires.

🤖 Generated with Claude Code