test(noise): add interop tests, listener binary, and criterion benchmarks for ML-KEM-768 HFS

[paschal533]

Summary

Additive contribution on top of this PR's implementation — royzah's source files (src/) are untouched.

What's added

• tests/interop_hfs.rs — SeededResolver infrastructure + two structural compliance tests:

  • handshake_structural_properties: verifies msg1=1216B, msg2=1200B, msg3=64B and both sides compute the same handshake hash within a single run
  • second_run_produces_coherent_results: second independent run with different seeds
  • print_vectors (ignored): diagnostic helper that prints message hex for manual cross-language inspection

• examples/noise_hfs_listener.rs — standalone TCP listener binary for cross-language interop testing:

  cargo run --example noise_hfs_listener --features mlkem-hfs -- 9999
  # prints: READY 9999
  # after connection: PEER 12D3KooW...
  

• benches/noise_hfs.rs — Criterion benchmarks comparing classical Noise XX vs XXhfs handshake latency, plus 1KB post-handshake transport throughput:

  cargo bench --features mlkem-hfs
  

Live Rust ↔ Python interop — confirmed ✓

I wrote a standalone Python dialer speaking Noise_XXhfs_25519+ML-KEM-768_ChaChaPoly_SHA256 directly (using kyber-py), then dialed this PR's listener binary live:

msg1 sent: 1216 bytes (e_pk=32, e1_pk=1184)
msg2 received: 1304 bytes
msg2: responder identity verified, peer=12D3KooWHvA2VB6DBguv9i27jy41SbYexFub8bXRmucFUevAGaWJ
msg3 sent: 168 bytes
PEER 12D3KooWHvA2VB6DBguv9i27jy41SbYexFub8bXRmucFUevAGaWJ
HANDSHAKE COMPLETE

Both sides completed mutual authentication (identity signatures verified) and the three-message handshake succeeded end-to-end. Message geometry confirmed: msg1=1216B, msg2=1304B (includes libp2p identity payload), msg3=168B.

Note on X-Wing variant: Our existing JS and Python implementations use X-Wing (ML-KEM-768 + X25519 bundled, protocol /noise-pq/1.0.0) — a valid but distinct approach from this PR's raw ML-KEM-768 pattern. The two are not wire-compatible, which is worth calling out for libp2p/specs#723.

Finding: snow's ML-KEM entropy bypasses the seeded resolver

The snow fork's generate() in DefaultResolver calls system entropy directly, ignoring the injected CryptoResolver. This prevents fully deterministic cross-language test vectors (relevant to specs#723 question 2 on KDF mixing order). Worth considering a seeded generation path in the fork for interop testing.

Suggestion: RustCrypto ml-kem swap (follow-up)

Resolver::resolve_kem() delegates to snow's DefaultResolver, whose ML-KEM backend is marked unaudited. The RustCrypto ml-kem crate (FIPS 203 compliant, audited by NCC Group) could replace it by implementing snow::types::Kem — same pattern as the existing x25519-dalek DH wrapper. Happy to write that as a follow-up PR if you're open to it.

Test plan

• cargo test --features mlkem-hfs --test interop_hfs → 2 passed, 1 ignored
• cargo build --example noise_hfs_listener --features mlkem-hfs → compiles
• cargo bench --bench noise_hfs --features mlkem-hfs --no-run → compiles
• Live Rust listener ↔ Python dialer handshake → HANDSHAKE COMPLETE
• cargo clippy --features mlkem-hfs → zero new warnings in noise crate
• cargo fmt --check → passes
• No modifications to src/lib.rs, src/protocol.rs, or src/io/

[paschal533]

Update: Full Rust ↔ JS ↔ Python triangle — 3/3 PASS

We migrated both js-libp2p-noise (PR #665) and py-libp2p (PR #1310) from X-Wing to raw ML-KEM-768 (matching this PR's protocol) and ran a live 3-way interop test:

Pair	Protocol	Result
Rust listener + Python dialer	Noise_XXhfs_25519+ML-KEM-768_ChaChaPoly_SHA256	✅ PASS
Rust listener + JS dialer	Noise_XXhfs_25519+ML-KEM-768_ChaChaPoly_SHA256	✅ PASS
Python listener + JS dialer	Noise_XXhfs_25519+ML-KEM-768_ChaChaPoly_SHA256	✅ PASS

All three runtimes (Rust/JS/Python) now speak the same wire protocol on /noise-mlkem768-hfs/0.1.0. The triangle test runner is at py-libp2p/scripts/interop_all.sh.

Peer IDs from the live run:

Rust listener + Python dialer: PEER 12D3KooWGPozMDorYpmMRUy1GsuSmtP8Di3jsZ4dWSo5jZYZKPwQ
Rust listener + JS dialer:     PEER 12D3KooWJckPQ3D88XbsHUJ4N9eGVd1Fwm2cfUKGaFjfQHQgQodc
Python listener + JS dialer:   PEER 12D3KooWHAggagiFE4DY71juR9TY54dwKJbBECkcPv285vaNoy5h

(Previously our JS and Python used X-Wing / /noise-pq/1.0.0 — a distinct protocol. We switched both to raw ML-KEM-768 to match this PR and confirm the triangle.)