Verify in-place AEAD decrypt is in use — adopt if not

[stanvx]

Parent

Architecture review (2026-06-21), "Latency tricks worth lifting wholesale". Independent of the deepening series. Smallest slice in this batch — could land in under an hour.

What to build

Both libsodium ChaCha20-Poly1305 and ring's AES-GCM support overlapping input/output buffers (m == c). Saves one stack-allocated buffer per encrypted packet. The architecture review claims AnyPlug's crypto "already uses ring"; this slice verifies whether the in-place pattern is actually in use, and adopts it if not.

Audit shared/usbip-core/src/crypto.rs and the call sites in usbip-server/src/crypto_stream.rs (and the matching client side). For each decrypt call, check whether the plaintext buffer is freshly allocated, or whether it's aliased to the ciphertext buffer. If aliased, document with a comment; if not, refactor to alias.

Acceptance criteria

• Audit comment in usbip-core/src/crypto.rs lists each decrypt call site with a one-line verdict: "in-place" or "allocates"
• Every "allocates" verdict is either (a) refactored to in-place with a unit test proving byte-equality, or (b) explicitly marked "must allocate because X" with reasoning
• cargo test --workspace --release passes (especially cargo test -p usbip-core --release for the crypto round-trip tests)
• cargo clippy --workspace -- -D warnings clean
• No public API change to SealedSession::open / seal
• A new micro-benchmark (criterion, default sample size) shows the latency improvement on a 512-byte packet

Blocked by

None — can start immediately.