Perl: a bare pattern match scores zero outside a boolean slot

[dekobon]

Summary

Perl's bare pattern match /^#/ — a match against the implicit $_ —
scores one ABC condition inside a boolean slot and zero outside one,
while the bound form $x =~ /^#/ scores one in both. Found while
landing #1461, which fixed the same asymmetry in five other languages
but could not use the same route here.

Measured

bca metrics --no-config, abc.conditions per sub, at 336ed709:

sub bare_in   { if (/^#/) { return 1; } return 0; }        # 1
sub bare_out  { my $r = /^#/; return $r; }                 # 0   <-- the gap
sub bound_out { my ($x) = @_; my $r = ($x =~ /^#/); $r }   # 1
sub ctrl_out  { my ($x) = @_; my $r = ($x == 1); $r }      # 1

bare_out is the only row that disagrees with its controls.

Why #1461's route does not work here

#1461 moved five languages' relational productions out of
<lang>_bool_terminal_kinds!() into an unconditional arm, so the
construct scores wherever it is written. PatternMatcher (337) and
PatternMatcherM (336) cannot take that route unchanged: the ABC walk
visits every node, so an unconditional arm would also fire on the
pattern inside $x =~ /^#/, whose EQTILDE token is already counted —
bound_out would go to 2. That is the §5 double count
(.claude/rules/grammar-dispatch.md), and it is the same collision that
made Kotlin's as go through the wrapper peel rather than the terminal
set in #1459.

The existing note on perl_bool_terminal_kinds! explains why listing
the patterns double-counts nothing today: every walker that consumes
the set either breaks on the enclosing binary_expression or requires
the chain's list node to be the parent, so the pattern node is never
reached beside its own =~. An unconditional arm has no such
protection.

The shape a fix would take

A gated arm: count PatternMatcher / PatternMatcherM unless the
node's parent is a binary_expression whose operator token is =~ or
!~. That is a different shape from the seven unconditional arms #1461
landed, which is why it was left out rather than folded in.

Two things to check before writing it:

• Whether substitution_pattern_s (s///) and
  transliteration_tr_or_y (tr///) should move with it. Both are
  absent from the terminal set today on the grounds that they evaluate
  to a count rather than a bool — a judgement perl_bool_terminal_kinds!
  records and decide: four uncounted boolean operands, and whether terminal sets should be token arms #1461 did not revisit.
• Whether the same gap exists in Ruby, whose =~ is likewise a token
  arm gated on a binary parent. Ruby has no bare-match-against-$_
  spelling, so it probably does not, but that is an assumption rather
  than a measurement.

Corpus cost

None expected: no corpus carries a .pl / .pm file.

[dekobon]

Split out of #1461, which landed the same rule for C#, Java, Groovy, Kotlin and Ruby. The decision record there carries the full per-language classification table; this is the one row in it that could not take the unconditional-arm route.