"incomplete value" regression in evalv3 which appears to be order dependent

[mvdan]

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

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

-- input.cue --
package p

routes: default: {}
routes: [string]: spec: [{
	refs: [{"default"}]
}]
#Fields: {
	input: _
	output: [for x in input { x }]
	...
}
#Kubernetes: {
	#Fields & {
		Routes: route: [_]: spec!: HTTPROUTESPEC.#x
	}
	#Fields // Commenting out this line also fixes evalv3.
	Routes: _
	input: res1: Routes
}
let HTTPROUTESPEC = {
	#x: [...{
		refs: [...{string}]
	}] | *[{
		other: []
	}]
}

// Order dependent! Breaks if the lines below are moved to the top of the file.
entrypoint: used.output
used: #Kubernetes & {
	Routes: route: routes
}

As of 46fc54aa9caf95393dcd3fe0ac066c8112da986d:

# With the old evaluator. (0.010s)
> env CUE_EXPERIMENT=evalv3=0
> exec cue vet -c
# With the new evaluator. (0.030s)
> env CUE_EXPERIMENT=evalv3=1
> exec cue vet -c
[stderr]
used.output.0.route.default.spec: incomplete value [{refs:["default"]}] | [{refs:["default"],other:[]}]
[exit status 1]

See the comments; removing one repetitive definition embedding, or moving the last two fields to the top of the file, both make the error go away. So it seems like there's some rather nasty bug in evalv3 that causes order-dependent or inconsistent evaluation.

[mvdan]

Reduced from a failure in Holos as reported by @jeffmccune. Note that #3637 was a spin-off from the example above, but a different bug entirely.