fix(pg-protocol): copy the bytes a copyData message keeps out of the parser buffer

[nigrosimone]

The parser hands copyData messages a chunk that is a view into its own buffer (BufferReader.bytes uses slice). When the next socket chunk arrives and what is left of the previous one fits in front, the parser moves those bytes to the start of the same buffer, so a chunk delivered earlier changes under whoever kept it. A copyData of 64 bytes followed by a commandComplete cut at its last byte is enough: the chunk then starts with COPY 1\0. pg itself ignores copyData, but a Submittable that keeps the chunk, for example to push it into a stream, gets silently corrupted data. The MD5 salt goes through the same method.

The fix returns a copy from bytes().

The added test parses that sequence in two chunks and checks the chunk is intact: it fails on master and passes here.

[charmander]

It’s a risky API design, but I think it was intentional that parsed messages can contain buffers that reference the parser’s working buffer. Is anything in pg actually broken as a result?

Then again, with binary mode being unimplemented, since bytes is only used by those two small messages, it’s not as if it’s much of a performance benefit.