evalv3 structural cycle regression compared to evalv2 (again)

[mvdan]

# evalv2
env CUE_EXPERIMENT=evalv3=0
exec cue export in.cue

# evalv3
env CUE_EXPERIMENT=evalv3=1
exec cue export in.cue

-- in.cue --
out: #Exporter & {
	_imports: [_branch]
	#conf: global: region: "Europe"
}

_branch: {
	objects: branch: {
		#conf: {...}
		region: #conf.global.region
	}
	imports: [_leaf]
	#export: #Exporter & {
		_objects: objects
		_imports: imports
	}
}
_leaf: {
	objects: leaf: {
		#conf: {...}
		region: #conf.global.region
	}
	#export: #Exporter & {
		_objects: objects
	}
}
#Exporter: {
	CONF=#conf: {...}
	_objects: {...}
	_imports: [...]

	objects: [
		// local objects
		for name, obj in _objects & {[string]: #conf: CONF} {
			(name): obj
		},

		// transitively imported objects
		for import in _imports
		for obj in ({#conf: CONF} & import.#export).objects {
			obj
		},
	]
}

As of 9a44314:

# evalv2 (0.008s)
> env CUE_EXPERIMENT=evalv3=0
> exec cue export in.cue
[stdout]
{
    "out": {
        "objects": [
            {
                "branch": {
                    "region": "Europe"
                }
            },
            {
                "leaf": {
                    "region": "Europe"
                }
            }
        ]
    }
}
# evalv3 (0.006s)
> env CUE_EXPERIMENT=evalv3=1
> exec cue export in.cue
[stderr]
structural cycle:
    ./in.cue:39:15
[exit status 1]
FAIL: repro-evalv3.txtar:7: unexpected command failure

Reported by @nxcc via a private reproducer. Split off from #3903, which was a smaller reduction of the same code; this reduction is more faithful to what the original private code was actually doing.