fix(genesis): write the canonical authority root on the existing-head branch and finish first-run identity publication in-session

[cryptskii]

Every freshly created wallet persisted a device head whose genesis authority root was [0u8; 32] while AppState held the real seed-rooted v3.g; the ERA faucet's authority evidence (which re-derives the true G) then failed closed on every first-run device. This PR fixes the root at its ordering and makes first-run identity publication complete in-session.

Root cause (reproduced by the new tests)

install_v2_genesis calls StateMachine::set_state BEFORE CoreSDK::write_genesis_device_head. set_state materialised a head with a zero genesis when none existed, and write_genesis_device_head only wrote genesis on its construct-new branch — so the existing-head branch (the one genesis install always takes) left the persisted head claiming a zero root. On top of that, AppRouterImpl::new ran the synthetic-genesis path on every router build, which once the write became authoritative would have overwritten a restored canonical root.

What changes

• dsm/types/device_state.rs: DeviceState::set_genesis_digest — install G on an already-constructed head.
• dsm/core/state_machine/mod.rs: set_state re-seeds an EXISTING head only; it does not know G and no longer manufactures a zero-root head.
• dsm_sdk/sdk/core_sdk.rs: write_genesis_device_head sets genesis unconditionally on both branches and installs the head in memory; the synthetic-genesis path refuses to run over an identity that already has a canonical head. Three tests reproduce the defect at its ordering: the existing-head branch, set_state not fabricating, and the router build not overwriting a restored root.
• dsm_sdk/sdk/kyber_identity.rs: a cold LOCAL_KYBER_PUBKEY slot is recovered from the canonical derivation (current_smaster() + DSM/kyber\0) instead of refused — first-run devices no longer park in PublicationPending until an app restart.
• dsm_sdk/handlers/system_routes.rs: system.createGenesisV2 writes the publication row and drives publish_identity_now in-session (spawned; genesis never blocks on the network; the startup retry stays the backstop, not the only driver).
• dsm_sdk/sdk/identity_publication.rs: after the Published row is written, post dsm-wallet-refresh so the Android host recomputes the session phase instead of showing "PUBLISHING IDENTITY…" until restart.

Hardware

On the 4-phone beta fleet (2026-09-13, wiped devices): system.createGenesisV2: identity PUBLISHED … (5/5 verified, quorum 3) in-session on all four; the ERA faucet claim then admits (100 ERA each) with no restart.

Verification

• cargo test --release -p dsm --lib core::state_machine types::device_state
• cargo test --release -p dsm_sdk --lib sdk::core_sdk sdk::kyber_identity sdk::identity_publication handlers::system_routes init:: (--test-threads=1)
• fmt / clippy -D warnings / ci/production_safety_checks.sh on the branch
• Board: CI on this PR.