[nateberkopec]

This is AI-assisted. Model: gpt-6.1-sol

Summary

Fixes #335.

libev's allocator contract treats realloc(ptr, 0) as "free this and return NULL", and libev discards the return value of those size-zero calls during teardown. nio4r currently passes Ruby's xrealloc directly to ev_set_allocator, but it does not implement that contract: both xrealloc(NULL, 0) and xrealloc(ptr, 0) return a new one-byte allocation (allocating the replacement before freeing the original pointer when non-NULL). libev throws that pointer away, so every selector teardown leaks a little memory.

This PR adds a small adapter in ext/nio4r/nio4r_ext.c:

• size == 0: xfree(pointer) and return NULL, matching libev's expectation.
• otherwise: xrealloc(pointer, size), so allocations still go through Ruby's allocator and GC malloc accounting.

The vendored libev sources are unchanged.

Reproduction: https://github.com/nateberkopec/nio4r-leak

Validation

• Linux aarch64, Ruby 3.3.10, Valgrind 3.19, using the existing repro:

  • stock nio4r, 0 / 100 / 1000 selector closes: 0 / 2,300 / 23,000 bytes definitely lost from selector teardown
  • with this patch, 1000 closes: 0 bytes definitely lost from selector teardown
  • a fresh Valgrind run against the exact PR candidate also reported 0 definitely-lost selector teardown bytes after 1000 epoll closes
  • Valgrind still reports other findings in Ruby itself. Those are unrelated to this change, so the run is not globally clean.

• macOS, Ruby 4.0.7: native extension builds. 1000 cycles each of register IO / select / deregister / double-close passed for :poll, :kqueue and :select.

• Apart from this change, the ext and lib gem trees match upstream HEAD exactly.

• Full upstream RSpec suite, against this PR's native extension with RSpec 3.13 (within the declared ~> 3.7 requirement):

  • macOS arm64, Ruby 4.0.7, kqueue: 112 examples, 0 failures, 2 pending (seed 23748)
  • Linux aarch64, Ruby 3.3.10, epoll: 112 examples, 0 failures, 2 pending (seed 44240)
  • Pending examples use the suite's existing UDP/TCP/TLS socket-fixture guards; no tests were excluded or modified.

No new tests were added.