...that the AnyCRLF module is automatically putting the crlf-layer on the stack if it is not already available.

Good question.  Like you, I couldn't figure out a way to get at the file handle of the lower layer in either OPEN or PUSHED.  OTOH, as you do have access to the handle in the FILL method, you could use binmode to compose the following hack:

package PerlIO::via::AnyCRLF;

sub PUSHED {
    my ($class) = @_;
    my $have_crlf;
    return bless \$have_crlf, $class;
}

sub FILL {
    my ($self, $fh) = @_;

    binmode $fh, ":crlf" unless $$self;
    $$self = 1;

    my $len = read $fh, my $buf, 4096;
    if (defined $buf) {
        $buf =~ s/\r/\n/g;
    }
    return $len > 0 ? $buf : undef;
}
1;
[download]

Although this works, it doesn't feel right. For one, it violates the principle of least surprise...  as you can see with the debug prints before and after reading from the file handle.

#!/usr/bin/perl

use PerlIO::via::AnyCRLF;

open my $f, "<:via(AnyCRLF)", "le.txt" or die $!;

# debug
# print "layers before reading :", join(':',PerlIO::get_layers($f)), "
+\n";  # :unix:perlio:via

print while <$f>;

# print "layers after reading  :", join(':',PerlIO::get_layers($f)), "
+\n";  # :unix:perlio:crlf:via
[download]

The other (similarly suboptimal) way would be to take care of opening a handle yourself:

package PerlIO::via::AnyCRLF;

sub PUSHED {
    my ($class) = @_;
    my $fh_ref;
    return bless \$fh_ref, $class;
}

sub FILL {
    my ($self) = @_;
    my $len = read $$self, my $buf, 4096;
    if (defined $buf) {
        $buf =~ s/\r/\n/g;
    }
    return $len > 0 ? $buf : undef;
}

sub OPEN {
    my ($self, $path) = @_;
    
    open $$self, "<:crlf", $path or die $!;

    # debug
    # print "AnyCRLF layers :", join(':',PerlIO::get_layers($$self)), 
+"\n";

    return 1;
}
1;
[download]

but in this case, the layer can no longer be properly stacked, as it always becomes the bottommost layer. (Might not be a problem in your particular case, though.)

Unfortunately, the PerlIO::via docs leave a number of questions unanswered (e.g. what is OPEN supposed to return; why does it fail to pass a handle, even if there is in fact a lower layer? etc.), and code samples using OPEN are hard to find.  Well, maybe the "is subject to change" note for the *OPEN methods is meant seriously — and noone has yet worked out in what way they should be changed :)