Don't lose log messages when two Catalyst apps share a process

[fleetfootmike]

[found by real live @fleetfootmike, explored and fixed driving Claude Code with /superpowers, /paad (with thanks to @obra, @Ovid)]

The problem

Mount two Catalyst apps in one process, which Plack::Builder makes easy enough, and each builds its own Log::Log4perl::Catalyst from the same config. Log4perl's appender registry is global, and the constructor doesn't cope with running over it twice: log messages go missing entirely, with no warning.

With autoflush off, the constructor sticks a buffer in front of every appender and tells the loggers to write to the buffer instead. Run it a second time and it builds a fresh set of buffers, drops them into the registry on top of the first set, and then only moves the loggers still writing to the raw appender:

foreach my $logger (values %$Log::Log4perl::Logger::LOGGERS_BY_NAME){
    if(defined $logger->remove_appender( $appender->{name}, 0, 1)) {
        $logger->add_appender( $buf_app );
    }
}

Any logger the first run already moved isn't on the raw appender any more, so remove_appender returns undef and it's left where it is: writing to a buffer that's no longer the one in the registry. _flush() walks the registry, so it empties the new buffers while the message sits in the orphaned one, and the line is never written.

Traced with refaddrs, the end of a request looks like this:

after pass 1: buffer#1 = ...605120
after pass 2: registry Api_catalyst_buffer = ...905928  (replaced, #1 orphaned)
  buffer#1 holds 1 msg(s)      <- the logger is still writing here
  buffer#2 holds 0 msg(s)      <- _flush() drains this one
  Api appender after flush: [] <- message gone

_flush() is never skipped, for what it's worth. It runs, and it's category-agnostic. It just drains the wrong objects.

Why it hasn't bitten more people

It needs a category named in the config, because those loggers exist in time for the first run to move them onto a buffer. A category that first appears later escapes it: it resolves its appender by name at the point it's created and finds the newer buffer. So whether you lose messages depends on whether the category is declared, which is a horrible thing to debug.

autoflush => 1 also hides it, because the second constructor then skips the wrapping block entirely and only one set of buffers is ever built. That's how I found this. An app of ours has an API and an admin UI mounted together, and the request log only worked because one app passed the flag. Take the flag away and the whole log silently empties.

The fix

The constructor already half expects to be run twice, in that it skips appenders that are themselves buffers. It just never checks whether an appender already picked one up. One next sorts it:

next if exists
    $Log::Log4perl::Logger::APPENDER_BY_NAME{ $buf_app_name };

t/072Catalyst.t covers it. It fails on master and passes with the fix. There were no tests for Catalyst.pm before, so it's a new file. Full suite is green: 76 files, 752 tests.

Deliberately kept minimal. The alternative is to have the second run adopt the existing buffers, but nothing needs that, and skipping is the smaller change.

Unrelated, while I was in here

t/071ScreenStdoutStderr.t isn't in MANIFEST, so it isn't shipped in the tarball. Left alone, since it isn't mine to fix in this PR, but you probably want it.