evaluator: allow source value for `for` comprehension to be partially evaluated

[myitcv]

What version of CUE are you using (cue version)?

$ cue version
cue version v0.0.0-20220824153543-547c4390430c

       -compiler gc
     CGO_ENABLED 1
          GOARCH arm64
            GOOS linux
             vcs git
    vcs.revision 547c4390430c9ab4274a8f5aa3548dacfc1f4981
        vcs.time 2022-08-24T15:35:43Z
    vcs.modified false

Does this issue reproduce with the latest release?

Yes

What did you do?

# z.cue - unrolled comprehension
exec cue eval x.cue z.cue
cmp stdout stdout.golden

# y.cue - comprehension
exec cue eval x.cue y.cue
cmp stdout stdout.golden

-- x.cue --
package x

#AllOutputs: {
	reject:   string
	resource: string
	retry: {
		output: #Output
	}
}

o: #Output & {
	retry: {
		output: {
			reject: "ok"
		}
	}
}
-- y.cue --
package x

#Output: or([ for name, config in #AllOutputs {
	(name): config
}])
-- z.cue --
package x

#Output: {
	reject: string
} | {
	resource: string
} | {
	retry: {
		output: null | #Output
	}
}
-- stdout.golden --
#AllOutputs: {
    reject:   string
    resource: string
    retry: {
        output: {
            reject: string
        } | {
            resource: string
        } | {
            retry: {
                output: null
            }
        }
    }
}
#Output: {
    reject: string
} | {
    resource: string
} | {
    retry: {
        output: null
    }
}
o: {
    retry: {
        output: {
            reject: "ok"
        }
    }
}

What did you expect to see?

Passing test.

What did you see instead?

# z.cue - unrolled comprehension (0.017s)
# y.cue - comprehension (0.014s)
> exec cue eval x.cue y.cue
[stderr]
2.retry: structural cycle:
    ./y.cue:3:13
[exit status 1]
FAIL: /tmp/testscript1560619719/repro.txtar/script.txt:6: unexpected command failure

Note that z.cue is the unrolled version of y.cue, so I think these two should be identical in behaviour.

Note this is note fixed by https://review.gerrithub.io/c/cue-lang/cue/+/542674.

Also, this appears to have been an issue for some time. At least as far back as v0.3.2.

[mpvl]

See

#Output: or([ for name, config in #AllOutputs { (name): config }])

#AllOutputs: {
	reject:   string
	resource: string
	retry:  output: #Output
}

The source for a for comprehension must be fully evaluated before it is used. This is due to how "void" arcs are handled. I guess theoretically this could be done "lazily", but this more complicated and currently it is defined as such.

As a result, evaluating #Output triggers the evaluation of #AllOutput, which in turn triggers the evaluation of #Output (a cycle), before the comprehension can be evaluated and thus before a call to or can be made.

[mpvl]

See https://review.gerrithub.io/c/cue-lang/cue/+/533224 for feasibility.

[myitcv]

Just noting per earlier discussion, that in the repro above, z.cue is the "unrolled" version of the comprehension in y.cue. exec cue eval x.cue z.cue shows there is not a cycle, so arguably exec cue eval x.cue y.cue should give the same result. Therefore, this issue is tracking ensuring that the comprehension version behaves like the "unrolled" version.