Behaviour discrepancy between the REPL and the testing environment round 2: second reduction boogaloo

[Hamiltonian-Action]

This is a sequel to #339 that now transcends the boundaries of loop unrolling, depending only on the configuration of the kernel launch. The setup is more or less the same and can be found in this updated repository. Unlike last time around, this issue is observed even in the latest development branch (b725007 at the time of writing).

Screenshots

REPL

Testing

[maleadt]

Oh, this may be because we're missing the REPL AST transform. Probably needs something like https://github.com/JuliaGPU/Metal.jl/blob/9b8c6043c223433a9b601ffc3bf2dcc060017cd9/src/initialization.jl#L70-L87