As I said, I started from a wrong end. I'm kind of busy working from home and staying there with a wife and three children, so I didn't have time to fix it immediately. Here's a much simpler and faster version, which parses melpa's archive-contents in less than 5 seconds on my machine:

#! /usr/bin/perl
use warnings;
use strict;

use Marpa::R2;

my $dsl = << '__DSL__';

:default ::= action => ::first
lexeme default = latm => 1

List     ::= ('(') Elements (')')
Elements ::= Element+               action => [values]
Element  ::= List
           | Vector
           | Atom
           | String
           | Pair

Vector   ::= ('[') Elements (']')
Atom     ::= identifier
String   ::= ('"') Quoteds ('"')
Quoteds  ::= Quoteds Quoted         action => concat
           | Quoted
Quoted   ::= backslash
          || qq
          || plain
Pair     ::= Element (dot) Element  action => pair


:discard   ~ whitespace
whitespace ~ [\s]+

dot        ~ '.'
backslash  ~ '\\'
qq         ~ '\"'
identifier ~ [-\w@:+]+
plain      ~ [^\\"]+

__DSL__

sub concat { $_[1] . $_[2] }
sub pair   { +{ $_[1] => $_[2] } }

my $grammar = 'Marpa::R2::Scanless::G'->new({source => \$dsl});

my $lisp = do { local $/; <> };
my $value_ref = $grammar->parse(\$lisp, {semantics_package => 'main'})
+;
use Data::Dumper; print Dumper $value_ref;
[download]

Thanks for this, choroba. It finishes in about 2 seconds on my computer, pretty impressive. I'll see what I can use to improve my SExpression::Decode::Marpa.

And here's my stab at creating a Marpa-based parser, based on JSON::Decode::Marpa: https://github.com/perlancar/perl-SExpression-Decode-Marpa/. It's unfinished (its number and string rules, particularly, are still not adjusted), but can already parse the original archive-contents file, a bit faster than Data::SExpression:

% time perl -Ilib -MSExpression::Decode::Marpa=from_sexp -MFile::Slurper=read_text -E'from_sexp(read_text "archive-contents")'

real    0m4.023s
user    0m3.818s
sys     0m0.204s