Calling kernelmatrix on a vector means its size becomes fixed (e.g. push! to it no longer works)

This might be a bug of Julia but here goes:

This does not work

X = ColVecs(rand(3, 20))
Z = collect.(X)
kernelmatrix(SqExponentialKernel(), [X[1]], Z)
push!(Z, X[1])
ERROR: cannot resize array with shared data

but this does (note the copy on Z)

X = ColVecs(rand(3, 20))
Z = collect.(X)
kernelmatrix(SqExponentialKernel(), [X[1]], copy(Z))
push!(Z, X[1])

also this does not work

X = [rand(3) for _ in 1:20]
Z = collect.(X[1:5])
kernelmatrix(SqExponentialKernel(), [X[1]], Z)
push!(Z, X[1])

Also removing the call to kernelmatrix solves the problem, and
When going to the stack trace this just leads to some ccall, so it's hard to go further....

Reproduce the issue using the Julia examples in the report, comparing the calls with Z and copy(Z). Start at kernelmatrix and follow the stack trace into the reported ccall boundary to find why Z becomes non-resizable. Done means the original Z remains resizable after kernelmatrix, while the existing kernel-matrix examples continue to work.