Sorry - no, I leave all the parallelization still to prove. It runs all test files in parallel (well, up to a limit). prove is called by make test, as you correctly surmised.

The run_across_instances just loops (single-tasked) through the Chrome instances.

I see. In that case I'm not sure how I could apply your setup to my problem (multiple .t files that should run with different shells)?

No - per prove-run, my approach would use the same shell for all runs. Maybe you can/want to use a shim in between your .t file and prove, that is something like:

#!/bin/bash
TESTSCRIPT=$1
for SHELL in ./test-shells/* ; do
    $SHELL $TESTSCRIPT
    if $? ; then
        exit $?
    fi
done
[download]

If you set this script as the interpreter of your test files, that should run each test with all shells (but still, single threaded). That way, you could at least achieve parallelism across your test files.

I'm sure there is a bash-way of stopping the interpreter script when any subprocess fails. Maybe it's set -e or something, but I'm sure you know that far better than I do.