NumPy parity: Sum/Prod over an empty axis, Slice with fewer indices; differential fuzzer

[tannevaled]

Parity fixes (each checked against NumPy 2.5.3 on cfarm421)

• SumAxis/ProdAxis along a zero-length axis now return the identity: np.zeros((3,0)).sum(axis=1) is [0,0,0] and .prod(axis=1) is [1,1,1]. Before, both returned an error, which was inconsistent with Sum()/Prod() of an empty array (0 and 1). Max/Min/ArgMax/ArgMin still error, matching NumPy's ValueError. MeanAxis still errors, matching Mean(): NumPy returns NaN with a RuntimeWarning, and Go has no warnings. The old docstring said the whole-array reductions "error on empty input", which was false for Sum/Prod; it is corrected. A MeanAxis error branch that can no longer be reached was removed, since the coverage gate flagged it.
• Slice with fewer indices than axes takes the remaining axes whole, as m[1] does in NumPy. This used to be an error. The padding never writes into the caller's index slice; a test covers this and fails when the guard is removed.
• MaskSelect: NumPy rejects a mask that only broadcasts to the array (verified). Here that is kept as a documented extension, and the docstring now states exactly where the two differ (leading-axes masks, 0-d masks).

Differential fuzzer (FuzzOps)

Arbitrary bytes drive slices, reshapes, transposes, squeezes, broadcasting arithmetic, axis reductions, Take, MaskSelect, Concatenate, MatMul and CumSum. Each result is compared element by element with an oracle written from NumPy's definitions (Python's slice.indices, the broadcasting rule, and so on). The library must reject exactly the calls the oracle rejects.

• 47M executions, no divergence after the two fixes above, which it found.
• Positive control: a first version only compared each array with its own Copy, and it missed a deliberate Squeeze stride bug in 31M executions. The oracle version catches it in 11 seconds (At = Copy = -3, oracle 3). That input is kept as a seed.

Docs

• The README and SECURITY.md now state the memory-amplification caveat. The library checks that a shape is representable, not that it fits in memory, and in Go running out of memory kills the process. Broadcasting, Outer, MatMul and summing the empty axis of (0, n) produce more than their operands hold.

Version: v0.4.0. Calls that used to error now return results.

🤖 Generated with Claude Code