Stage1 bug: runtime copy of statically-dereferenced *[0:S]T doesn't hold sentinel

[rohlem]

Slicing with length of 0 returns a slice with .len = 0, .ptr = undefined. See #4771 for discussion on that.
This issue is probably related, but when testing I found out that sentinel-terminated slices actually work, so I can't find a source of UB anywhere in what I'm doing. Test case:

const expect = @import("std").testing.expect;
test "sentinel-terminated 0-length slices" {
    var u32s: [4]u32 = [_]u32{0, 1, 2, 3};

    var index: u8 = 2;
    var slice = u32s[index..index:2];
    var array_ptr = u32s[2..2:2];
    const comptime_known_array_value = u32s[2..2:2].*;
    var runtime_array_value = u32s[2..2:2].*;

    expect(slice[0] == 2);
    expect(array_ptr[0] == 2);
    expect(comptime_known_array_value[0] == 2);
    expect(runtime_array_value[0] == 2); //fails
}

Inspecting the addresses with std.debug.print, comptime_known_array_value and runtime_array_value have the same address... whatever that means.
My best explanation would be that the compiler gets confused over whether it needs to separately allocate and copy the sub-array or not, and so that memory is left uninitialized.