ssl: raise ArgumentError for a negative length in SSLSocket#sysread

[Shaisolaris]

Since bf1a419 ("ssl: implement blocking SSLSocket methods in Ruby"), a negative length passed to SSLSocket#sysread or #read_nonblock reaches SSL_read():

ssl.sysread(-1)
# master:      OpenSSL::SSL::SSLError: SSL_read: bad length
# before:      ArgumentError: negative string size (or size too big)

The new string_new_capacity clamps the capacity to 0, so the old ArgumentError from rb_str_new() no longer happens. With a caller-provided buffer the negative length already reached SSL_read() before that commit.

This adds a check in ssl_read that raises ArgumentError for a negative length, like IO#sysread, in all of these paths. The change has not been released yet (it is not in v4.0.2).

Tests:

• Added test_sysread_negative_length to test/openssl/test_pair.rb. It fails on master and passes with this change.
• test_pair.rb (33 tests), test_ssl.rb (92), test_ssl_session.rb (11) and test_buffering.rb (5) pass locally on Ruby 3.3.8 with the system OpenSSL. I did not run the full rake test or the CI matrix of OpenSSL/LibreSSL/AWS-LC versions.

Follow-up validation — October 16

Follow-up validation (October 16): independently detect rb_len2long rather than assuming that HAVE_RB_LEN_T implies the conversion helper exists. A syntax-only compile probe with a present length type and absent helper fails before and passes after the guard repair. The negative-length regression now lives in the SSL test class rather than the shared native Socket mixin. Built on Ruby 3.4.8/OpenSSL 3.6.2; full socket-pair file: 103 tests, 531 assertions, no failures. Windows/Ruby-head CI remains to be verified.

[rhenium]

We should make #read, #readpartial, and #read_nonblock reject this error as well, not just #sysread.

(As a bonus - IO#sysread also handled this inconsistently when an output buffer is passed. I ran into that a few days ago and filed https://bugs.ruby-lang.org/issues/22433.)

The second commit doesn't look necessary. defined HAVE_RB_LEN_T is indeed supposed to imply that rb_len2long() is available. Both currently exist only in Ruby's master branch.

[rhenium]

windows mingw head looks outdated:

ruby 4.1.0dev (2026-10-10 master 17aeb89200) +PRISM [x64-mingw-ucrt]

Ruby < 3.2 on Windows appears to be failing due to https://bugs.ruby-lang.org/issues/18880