BLE P0 (release-blocking): authenticate the BLE-delivered ML-KEM key against the pinned AK before caching (ADR 0002 rows 8/9)

[cryptskii]

RELEASE-BLOCKING — do not merge until two-phone hardware validation passes.
This is a BLE wire-format change (new proto fields). DSM beta redeploys both peers together; there is no back-compat. Both phones must be flashed with this build and a real bilateral transfer must complete before merge.

What this closes

ADR 0002 matrix rows 8/9 — the last open rows. The bilateral BLE prepare handlers cached the peer's wire-supplied ML-KEM (Kyber) key with no authentication: no signature the receiver verified against the pinned AK covered that key. (The commit-phase bilateral-sign transcript signs only domain ‖ commitment_hash — it excludes the Kyber key, and it is transaction authorization, not peer-identity authentication.)

Exploit it closes: anyone who could deliver a prepare frame naming a victim's device_id could seed an attacker Kyber key first-write-wins; every subsequent §11.1 per-step-EK receipt then encapsulated to the attacker's key.

The fix — a detached binding, reusing the one identity primitive

Chosen deliberately over folding the key into the bilateral-sign transcript (that would couple peer-key authentication to the commit-phase signature — wrong layer boundary). This reuses the same verify_kyber_identity_binding primitive already used by storage-fetch and repair, so DSM has one way to authenticate a Kyber key.

• proto (proto/dsm_app.proto): sender_kyber_binding_sig (BilateralPrepareRequest field 17), responder_kyber_binding_sig (BilateralPrepareResponse field 10).
• sender: build_local_kyber_identity_binding() signs binding_digest(device_id, genesis, kyber_pk) under the local AK, riding the request, the response, and the non-Android host response path.
• receiver (verify_and_cache_peer_kyber): verifies the detached binding against the PINNED contact AK — never the wire signing key, never the wire device_id/genesis — before bind_contact_kyber_key_if_absent.

pinned QR/BLE AK
   └─ verifies kyber_binding_sig over binding_digest(device_id, genesis, kyber_pk)
        └─ only then first-write cache

bilateral-sign is unchanged.

Behavior

condition	result
valid binding under pinned AK	Kyber key cached (first-write-wins)
empty key / empty sig	rejected, nothing cached
forged sig	rejected, nothing cached
wrong AK (valid binding, non-pinned signer)	rejected, nothing cached
substituted key (valid sig, different wire key)	rejected, nothing cached
non-canonical / old-domain digest	rejected, nothing cached
no contact record	rejected, no TOFU
contact with no pinned AK	rejected, no TOFU

Every rejection leaves the prior Kyber cache untouched. Skeleton/echo/test constructors of the prepare messages carry an empty binding (fail-closed at the receiver).

Proof

verify_and_cache_peer_kyber_verifies_before_caching (in bilateral_ble_handler.rs) drives all eight rows above against a real test DB with a real SPHINCS+-signed binding: every rejection asserts the cache stayed empty; the valid case asserts it cached. Mutation-verified — disabling the binding verification turns the test red (then restored).

Gates (local)

• cargo fmt --all -- --check ✅ · cargo clippy --all-targets -- -D warnings ✅ · scripts/guard_protos.sh ✅ (single source of truth) · ci/production_safety_checks.sh ✅
• cargo test -p dsm_sdk / cargo test -p dsm — results below.

Still required before merge

Two-phone hardware validation — flash both devices, complete a real bilateral BLE transfer, confirm the receiver caches the peer Kyber key (binding verifies) and that a tampered/absent binding fails closed. Not runnable in CI.