security: validate result shapes, bound empty-array work, guard every asm call; go 1.26.6

[tannevaled]

Security audit, 2026-10-04. Full write-up in SECURITY.md.

Found and fixed

• Concatenate summed axis lengths without an overflow check: 17 arrays (2^59, 0) gave a negative dimension. Fixed with a checked sum, plus validateShape on the result. This also covers Stack, which used to return (9, 2^58, 0), a shape New rejects.
• MatMul/Dot/Inner of (m, 0) @ (0, n) never validated m*n. (2^32,0)@(0,2^32) returned a (2^32, 2^32) array with Size() == 0; a product wrapping to −1 panicked in makeslice. Both are now ErrShapeMismatch.
• Linspace(…, 1<<62) panicked in makeslice; it now returns ErrShapeMismatch.
• Effective hangs (DoS) on empty arrays with a huge axis. Reducing (2^58, 0) over axis 0, concatenating it along its last axis, or CumSum of (0, 2^58) walked 2^58 iterations doing nothing. The kernel is now skipped when the result is empty.

Hardened

• Every Go wrapper around an assembly kernel (amd64, arm64, ppc64le, loong64) indexes the last element the kernel will touch in each operand. A caller's mistake becomes a Go index panic instead of an out-of-bounds access in assembly. TestWrappersRefuseShortOperands checks this; sabotaging one guard makes it fail.
• New fence_test.go: each operand is mmap'd flush against a PROT_NONE page, both head and tail, for n = 1…70 and GEMM kc = 1…9, and the FMA/LASX fallbacks run too. It passes on real amd64 (Zen 3), arm64 (cfarm185 + M-series), ppc64le (POWER9) and loong64 (3C5000L, LASX). Positive control: a kernel reading one element too far faults on all four page sizes (4K/16K/16K/64K).

Other

• go 1.26.6 (root and generator module). govulncheck at 1.26.4 listed 10 stdlib vulns (none reachable); at 1.26.6 there are none. Regenerated .s are byte-identical on all 4 arches.
• Private vulnerability reporting is enabled on the repo, as SECURITY.md says.
• Untracked 4 coverage profiles (cg*.out, cover2.out) that had been committed; *.out is now ignored.
• README: SIMD now ships on amd64/arm64/ppc64le/loong64 (the text said amd64+arm64); new "Untrusted input" section.

Cost

On a single Zen 3 core, an A/B of 10 interleaved rounds: MatMul 256², Dot1D and AddInto 16Ki all fall within ±2%, which is inside the spread between runs.

Version: v0.3.0. The minimum Go version goes up, and some calls now return errors where they used to return arrays.

🤖 Generated with Claude Code