The runtime race tests during run.bash do:

if err := t.dirCmd(".", "go", "test", "-race", "-i", "runtime/race", "flag", "os/exec").Run(); err != nil {
    return err
}
if err := t.dirCmd(".", "go", "test", "-race", "-run=Output", "runtime/race").Run(); err != nil {
    return err
}
if err := t.dirCmd(".", "go", "test", "-race", "-short", "flag", "os/exec").Run(); err != nil {
    return err
}

At some point during this, most likely during "go test -race -i runtime/race flag os/exec", the installed cmd/cgo binary becomes a race-enabled cmd/cgo with tsan linked into it.

The most obvious effect of linking tsan is that cmd/cgo becomes unable to run in under 1 second. It looks like there is a 1-second sleep in the tsan finalizers (for a good reason I assume, but cc @dvyukov).

To be clear, the bug here is not the 1-second sleep but the fact that the go command is installing a race-enabled cgo.