fix(noise): size handshake read buffer with correct overhead

[dimartiro]

Description

read_handshake_message allocated its decryption output buffer as message.len() + 200 — an arbitrary overhead flagged with a TODO.

Closes #332

[lexnv]

Could you please double check for running nodes in Kusama and Polkadot? 🙏 I would simply place info logs for before and after with the actual length.

IIRC the actual message we receive is bigger than the message.len() hence the added 200, which would otherwise trigger a reallocation and affect performance

[dimartiro]

The out buffer holds the decrypted plaintext, not the received frame — and in Noise the plaintext is always smaller than the ciphertext, so message.len() can't be exceeded. I traced this in snow 0.9.6:

• _read_message writes to out only in the final decrypt_and_mix_hash(ptr, payload) (handshakestate.rs:445); the ephemeral/static DH keys go to internal buffers (self.re/self.rs), never to out. ptr is a suffix of message, so ptr.len() <= message.len().
• The size check is out.len() >= ciphertext.len() - TAGLEN with TAGLEN = 16 (cipherstate.rs:53), so out.len() = message.len() is sufficient with margin.
• read_message takes a fixed &mut [u8]; an undersized buffer returns Error::Decrypt, never a reallocation. So the old +200 was pure over-allocation, not realloc-avoidance.

Happy to add temporary before/after info logs on a Kusama/Polkadot node if you'd still like empirical confirmation, but the bound is provable from the snow source.