ssh: fix mlkem768x25519 hybrid shared secret encoding

[cole-christensen]

Closes #11208

Summary

mlkem768x25519-sha256 key exchange intermittently fails (~1 in 512 handshakes)
with incorrect signature on the peer. hybrid_common/4 reconstructed the
classical X25519 shared secret by taking the last ?X25519_PUBLICKEY_SIZE bytes
of its mpint encoding. Because mpint is minimal-length, a shared secret whose
most significant byte is 0x00 and whose next byte is < 0x80 encodes to a
31-byte payload, and "the last 32 bytes" then includes a byte of the 4-byte
length prefix (0x1f) in place of the genuine leading 0x00. The two peers hash
different octet strings, derive a different exchange hash H, and the signature
check fails.

This encodes the ECDH shared secret as a fixed-width big-endian octet string,
preserving leading zero bytes.
draft-ietf-sshm-mlkem-hybrid-kex
§2.4 specifies this encoding — the component secrets are hashed as fixed-length
big-endian byte arrays — and OpenSSH's kexmlkem768x25519.c matches it.

 hybrid_common(K_pq_secret, Curve, PeerPublic, MyPrivate) ->
     K_cl_secret = compute_key(ecdh, PeerPublic, MyPrivate, Curve),
-    K_cl_secret_mpint = <<?Empint(K_cl_secret)>>,
-    K_cl_secret_mpint_trim =
-        binary:part(K_cl_secret_mpint, byte_size(K_cl_secret_mpint), -?X25519_PUBLICKEY_SIZE),
-    crypto:hash(sha(Curve), <<K_pq_secret/binary, K_cl_secret_mpint_trim/binary>>).
+    K_cl_secret_fixed = <<K_cl_secret:(?X25519_PUBLICKEY_SIZE*8)/big-unsigned-integer>>,
+    crypto:hash(sha(Curve), <<K_pq_secret/binary, K_cl_secret_fixed/binary>>).

The new encoding is byte-for-byte identical to the old one in every case except
the previously-broken one, so it only changes behaviour where the handshake was
already failing.

Probability / impact

P(trigger) = P(secret[0]==0x00) × P(secret[1] < 0x80) = 1/256 × 1/2 =
1/512 ≈ 0.195% per mlkem768x25519-sha256 handshake. At any non-trivial
connection volume this is a frequent, hard-to-diagnose interop failure against
OpenSSH (which is correct).

Introduced in 95d0848c60 ("ssh: MLKEM kex ssh"), released in OTP-28.4.

Verification

• Deterministic unit check (<<0,1,0:240>> → old code yields 1F01…, fixed
  yields 0001…). See linked issue.

• Monte-Carlo: 300k random secrets, old-vs-fixed mismatch rate 0.194%, every
  mismatch satisfies secret[0]==0x00 ∧ secret[1] < 0x80.

• End-to-end against OpenSSH 10.3p1, BOTH roles, with the hybrid code
  instrumented to count how many handshakes actually hit the trigger ("case C")
  so success is not luck-of-the-draw:

  Role	Mode	Trigger handshakes	Failures
  OTP daemon ← OpenSSH client	stock	20	20 (1:1)
  OTP daemon ← OpenSSH client	fixed	45	0
  OTP client → OpenSSH sshd	stock	21	21 (1:1)
  OTP client → OpenSSH sshd	fixed	31	0

• Regression test in the suite (ssh_algorithms_SUITE:mlkem768x25519_hybrid_secret_encoding):
  passes with the fix, fails against the stock encoding. Run via the supported
  framework: ts:run(ssh, ssh_algorithms_SUITE, mlkem768x25519_hybrid_secret_encoding, [batch])
  → TEST COMPLETE, 1 ok, 0 failed.

• Full ssh_algorithms_SUITE: TEST COMPLETE, 930 ok, 0 failed of 930

[CLAassistant]

All committers have signed the CLA.

[github-actions]

CT Test Results

    2 files     29 suites   28m 54s ⏱️
  506 tests   499 ✅  7 💤 0 ❌
1 715 runs  1 686 ✅ 29 💤 0 ❌

Results for commit 20ce349.

♻️ This comment has been updated with latest results.

To speed up review, make sure that you have read Contributing to Erlang/OTP and that all checks pass.

See the TESTING and DEVELOPMENT HowTo guides for details about how to run test locally.

Artifacts

• Complete CT logs (Download Logs)
• HTML Documentation (Download HTML Docs)
• No Windows Installer found

// Erlang/OTP Github Action Bot