parser & IO-layer grab-bag: prefix ~~ in return, chained hash subscripts, PerlIO::via stub

[fglock]

Summary

Grab-bag PR picked up while running jcpan -t against CPAN modules. Three independent fixes, each motivated by a different module; unified here per request.

1. Prefix ~~ at start of return arguments

Fixes prefix ~~ (double bitwise complement, idiomatic for "force numeric scalar context", e.g. ~~@array to get a count) in return statements.

Two bugs, both triggered by the same Perl idiom in SQL::Beautify::_is_keyword:

return ~~ grep { $_ eq uc($token) } @{$self->{keywords}};

• In return ~~ EXPR, parseZeroOrMoreList's looksLikeEmptyList saw ~~ (registered as an INFIX_OP for smartmatch) as an infix operator and treated the list as empty, silently dropping EXPR. parseReturn now detects a leading ~~ and parses the operand as a normal expression. We can't change looksLikeEmptyList globally, because constructs like undef ~~ undef (where undef is a ;$-prototype op) legitimately parse as undef() followed by binary smartmatch — an earlier version of this fix broke that and regressed op/smartmatch.t.
• ParsePrimary case "~~" tried to split ~~ into two ~ tokens by rewriting the single ~~ lexer token to ~ and re-parsing — but the second ~ token never existed (the lexer emits ~~ as a single token), so ~~@x was effectively ~@x. Rewrote to directly build ~(~EXPR).

Found via jcpan -t SQL::Beautify, which failed 5 subtests because _is_keyword always returned undef, so uc_keywords never uppercased anything.

2. Multi-element subscripts in chained hash deref

Previously, chained hash access like $h{a}{-word => 'ou'} (implicit arrow deref) evaluated the multi-element subscript in scalar context, keeping only the last element ('ou'). The initial (non-deref) level already joined keys with $; (SUBSEP) to form 'a$;b'-style keys, so the two paths disagreed:

$h{-word => 'ou'}       -> FETCH("-word\x1cou")  OK
$h{a}{-word => 'ou'}    -> FETCH("ou")            wrong

Fixed the chained-deref path to join multi-element subscripts with $; too, matching both upstream Perl semantics and PerlOnJava's own behavior on the non-deref case.

Found via jcpan -t Regexp::Common.

3. PerlIO::via stub + loud-fail for :via(...) layer

Unblocks jcpan -t Redis (and the wider PerlIO::via::Timeout → IO::Socket::Timeout → Redis dependency chain). Before this change, PerlIO::via was missing entirely, so CPAN flagged it as a failed prerequisite and the whole chain was marked NA, causing the compile tests of the downstream modules to error with Can't locate ....

• src/main/perl/lib/PerlIO/via.pm — new stub (same pattern as the existing PerlIO::encoding stub). Lets use PerlIO::via; succeed so CPAN's prerequisite resolver is happy and downstream modules can be installed. No layer dispatch happens here — that's out of scope for this change.
• LayeredIOHandle — teach splitLayers that via(...) is paren-grouped (same treatment as encoding(...)), so a spec like :via(Foo::Bar) is not split at the :: inside the class name. addLayer now recognizes via(...) and throws PerlJavaUnimplementedException. binmode catches that exception type specifically and emits a Perl-level warning instead of silently swallowing it — so open($fh, "<:via(Foo)", ...) now says "PerlIO layer :via(Foo) not implemented" rather than returning a silently layer-less handle.
• dev/modules/perlio_via.md — design doc covering the full plan for a functional PerlIO::via implementation (bridging layer dispatch into user-supplied Perl callbacks). Not attempted here; tracked separately.

Result:

./jcpan -t Redis
...
Result: PASS  (PerlIO::via::Timeout)
Result: PASS  (IO::Socket::Timeout)
Result: PASS  (Redis)

All Redis tests that can run without fork / a live redis-server now pass.

Test plan

• make passes
• jcpan -t SQL::Beautify passes
• jcpan -t Regexp::Common passes
• jcpan -t Redis passes (all three modules in the chain: PerlIO::via::Timeout, IO::Socket::Timeout, Redis)
• Existing layer specs (:encoding(UTF-8):crlf, etc.) still work — regression-checked with a small round-trip

Generated with Devin