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

use Text::CSV_XS;

my $csv = 'Text::CSV_XS'->new ({ binary         => 1,
                                 auto_diag      => 1,
                                 keep_meta_info => 1 });
open my $in, '<:encoding(utf8)', shift or die $!;
while (my $row = $csv->getline($in)) {
    next if $row->[0] =~ m/^#/ && ! $csv->is_quoted(0);

    $csv->say(*STDOUT, $row);
}
[download]

Tested with

#x,y,z skip
abc,"d
#e
f",ghi keep
#comment skip
a,b,c,#xyz keep
"#foo",x,y,z keep
[download]

The meta info knows whether the field was quoted or not.

True, though AFAICT the meta_info doesn't seem to keep track of escaped characters:

use warnings;
use strict;
use Data::Dump;
use Text::CSV;

my $csv = Text::CSV->new({ binary=>1, auto_diag=>2,
    keep_meta_info=>1, escape_char=>"\\" });
while ( my $row = $csv->getline(*DATA) ) {
    dd $row, $csv->meta_info;
}
$csv->eof or $csv->error_diag;

__DATA__
foo,bar
"#foo","bar"
#foo,bar
\#foo,bar
[download]

In this special case it looks like there won't be portuguese words starting with a "#", so it would work for OP, as long as he is aware of it

True as well :-) (I guess this is more about the generic case of filtering comments from CSV files.)