evalv3: "explicit error in source" regression when using bottom in repeating disjunctions

[mvdan]

# With the old evaluator.
env CUE_EXPERIMENT=evalv3=0
exec cue vet -c=false

# With the new evaluator.
env CUE_EXPERIMENT=evalv3=1
exec cue vet -c=false

-- input.cue --
package p

#Action: {
	action: {
		read!:  string
		write?: _|_
	} | {
		write!: string
		read?:  _|_
	}
}

#Minio: {
	container: #Action & {
		action: {
			read!:  string
			write?: _|_
		} | {
			write!: string
			read?:  _|_
		}
	}
	doWrite: bool | *true
	container: action: defaultAction
	defaultAction: {
		if doWrite { write: "somefile.txt" }
		if !doWrite { read: "somefile.txt" }
	}
}

As of c29fd02:

# With the old evaluator. (0.010s)
> env CUE_EXPERIMENT=evalv3=0
> exec cue vet -c=false
# With the new evaluator. (0.027s)
> env CUE_EXPERIMENT=evalv3=1
> exec cue vet -c=false
[stderr]
explicit error (_|_ literal) in source:
    ./input.cue:17:12
explicit error (_|_ literal) in source:
    ./input.cue:20:12
[exit status 1]

This seems like a straight up regression to me. Note that the original source used regular fields in place of required fields, but the regression reproduces with either of them, and required fields are the more appropriate here I think.

[mpvl]

Simplified reproducer:

{ a?:  _|_ } | { a:  string, b?: _|_ }
{ a?:  _|_ } | { a:  string }
if false { a: "x" }

(with help of the graphical CUE debugger that showed quite quickly where the problem lies.)