Falling through control flow should also be counted.
In other words I think this should also be warned on:

if x {
 r := Grab()
 use(r)
 defer r.Close()
}
return

I'm afraid that this at the same time causes friction especially when writing tests and the tests get fleshed out step by step. In those cases govet will block running the test, even if it is fully correct.

rsc, err := openrsc()
if err ...
defer rsc.Close()
// test test so far, more to come later

@thediveo I tried covering this with:

Ideally this would be counted as per SSA rules so

It's very unclear (and incorrect).
I guess a more correct analysis could be to do this analysis on the CFG while ignoring control flow.
So if err != nil and it's content would be ignored because your error checking block would finish with a panic, return or other terminating block.
defer rsc.Close() would be the final statement, if would be skipped, rsc, err := openrsc() would be the penultimate statement and this wouldn't match since one of the return values from the penultimate value is an argument to the defer.

how frequent is this that it needs to be a vet check?

How does this satisfy vet's 'Correctness' criteria? cmd/vet/README:

Correctness:

Vet's checks are about correctness, not style. A vet check must identify real or
potential bugs that could cause incorrect compilation or execution. A check that
only identifies stylistic points or alternative correct approaches to a situation
is not acceptable.

Unless one or more of their arguments (this include receivers, function arguments and variables captured by closures) are coming from the penultimate statement, this is so code like this remains ok:

I don't understand this line. Do you mean whether or not arguments exist will affect the logic? (It certainly doesn't).

I searched for matches in opensource code and there were a couple too much false positive, I thought it would be pretty rare, but yes this probably falls under styling.