https://go.dev/play/p/THKjlNkot6b

package main

import (
	"testing"
)

func Test(t *testing.T) {
	defer func() { println("defer", t.Failed()) }()
	t.Cleanup(func() { println("Cleanup", t.Failed()) })
	panic(1)
}

Changing the panic to t.Fatal("foo") does say it failed, so panic probably ideally would do the same.
The problem is it really can't, because the only way to find out about a panic is a recover in a deferred function,
and the handler that will mark the test failed runs after the more recently deferred stuff.
We might be able to special case t.Cleanup, but it wouldn't help the plain defer.

On further thought, there's a clear distinction between defer and cleanup. Deferred func can recover the panic and stop the failure. The cleanup cannot. So it makes sense for cleanup to see t.Failed and the defer not to. I see the bug in the Cleanup case and will send a CL (but not for Go 1.18).

Thanks for looking into this, @rsc!

So it makes sense for cleanup to see t.Failed and the defer not to

I agree. Also, currently we are being a bit inconsistent since t.Failed does return true from a cleanup function if the test has parallel subtests:

https://go.dev/play/p/G_BE1zxkcEG

package main

import (
	"testing"
)

func Test(t *testing.T) {
	t.Cleanup(func() { println("Cleanup", t.Failed()) })
	t.Run("", func(t *testing.T) { t.Parallel() })
	panic(1)
}

Prints: Cleanup true.

Rolling forward to 1.20. Please comment if you disagree. Thanks.

I believe CL 506755, which was a fix for #61034, also fixes this. That CL was reverted due to an issue it introduced, but hopefully it will be resubmitted eventually.