Just for the record, a "unixish" implementation of BrowserUk's idea could look something like this:

#!/usr/bin/perl -slw
use strict;

our $N //= 12;
our $I //= 0;

my $cmd = qq[exec $^X -E'sleep $N; \$_=<STDIN> if $I; sleep 2; printf 
+"Kid done (read %d bytes)\\n", length'];

my $inInputState = 1;
my $timedOut = 0;
$SIG{PIPE} = sub { $inInputState = 0; };
$SIG{ALRM} = sub { $inInputState = 0; $timedOut = 1; die };
$SIG{CHLD} = 'IGNORE';

my $pid = open CMD, '|-', $cmd or die $!;

alarm 10;
eval { syswrite CMD, " \b"x(2**15+1) };
alarm 0;
    
if( $timedOut ) {
    print "Command timed out";
    kill 15, $pid;
}
if( $inInputState ) {
    print "Child waiting for input";
}
else {
    print "Kid never entered input state";
}
print 'Parent done';


__END__
$ ./detectChildInputState -I=0 -N=2
Kid done (read 0 bytes)
Kid never entered input state
Parent done

$ ./detectChildInputState -I=1 -N=2
Child waiting for input
Parent done
$ Kid done (read 65538 bytes)

$ ./detectChildInputState -I=1 -N=12
Command timed out
Kid never entered input state
Parent done

$ ./detectChildInputState -I=1 -N=9
Child waiting for input
Parent done
$ Kid done (read 65538 bytes)

$ ./detectChildInputState -I=0 -N=9
Command timed out
Kid never entered input state
Parent done
[download]

The main difference revolves around the SIGPIPE signal which on Unix would be delivered to a process if it attempts to write to a broken pipe (this is the case when the child terminates before having read anything).

By default, this signal would terminate the writing process, so it would have to be handled one way or another, anyway (e.g. $SIG{PIPE} = 'IGNORE'). OTOH, we can take advantage of this error notification, in which case we don't need an extra thread (or process) doing the blocking write.  The logic is kind of reversed now: we assume things went ok, unless we know otherwise, which is when

• the child terminated before it went into an input state (in which case we get a SIGPIPE)
• a timeout occurred before the child went into an input state

in those cases, $inInputState is set to zero in the respective signal handler.

A couple of more notes:

• the timeout is implemented via the usual alarm mechanism (instead of status polling in a loop)
• I'm using syswrite to circumvent Perl's own buffering without having to fiddle with autoflush
• Unix pipes typically use a rather large buffer (64k in my case), so the chunk written needs to be significantly larger than on Windows, in order to get the write operation to block
• the backspace cancellation trick would only work under rare circumstances (AFAICT) — simply reading from stdin would, for example, not treat the backspaces in any special way.
• last but not least, as a consequnce of the above, the tested program should be able to handle 64k of junk in case it puts up an innocent prompt (whether this is in fact an issue, of course depends on the type of program...)