ReadIOAdapter#read returns too many bytes for Enumerator bodies

[fallintoplace]

Summary

OpenAI::Internal::Util::ReadIOAdapter#read(max_len) can return more than max_len bytes when wrapping an Enumerator. The buffered enumerator path currently slices with an inclusive range, so read(3) can return 4 bytes and read(0) can consume data.

Reproduction

adapter = OpenAI::Internal::Util::ReadIOAdapter.new(["abc", "def"].to_enum) {}

p adapter.read(0)
p adapter.read(1)
p adapter.read(3)
p adapter.read(99)
p adapter.read(1)

Expected

""
"a"
"bcd"
"ef"
nil

Actual

""
"ab"
"cdef"
""
nil

Impact

This adapter is used for streamed request bodies, including multipart uploads. IO#read(max_len) is byte-oriented, so returning more than max_len bytes can violate callers' expectations when an upload body is backed by an Enumerator. The same code path should use byte lengths rather than character lengths for buffered reads.