consensus: rejoin a lagging archive by syncing proposals from peers

[dazthecorgi]

A global-consensus participant that falls behind — e.g. an archive isolated by a network partition — could not rejoin consensus. Orphan resolution and finalized-rank advancement happen only on the consensus message path (on_receive_proposal -> forks.add_validated_state -> drain orphans); a plain frame-store write never triggers them, and freshness gates drop a lagging node's traffic until its finalized rank advances. With no path to backfill the missed proposals, the node would orphan every later proposal forever — its frame store could catch up (via the read-only poller) while its consensus engine stayed permanently stuck.

This ports the Go node's catch-up path (SyncProvider / GetGlobalProposal -> AddProposal): when the engine orphans a proposal for a missing parent, pull the missing proposals from a peer and submit them into the consensus loop so the node finalizes them and resumes voting — rather than merely mirroring frames.

Changes:

• Serve full proposals: implement GlobalService.GetGlobalProposal, assembling state + parent QC + prior TC + proposer vote from the clock store (FrameLookup::get_global_proposal). Previously stubbed to return nothing for non-genesis frames.
• Persist the proposer vote at proposal ingest (keyed filter,rank,selector) so it can be served back; the store trait had the accessor but no producer wrote it. Add ProposalVote/TimeoutCertificate proto<->wire conversions and proto_proposal_to_signed (only QC had one).
• Trigger: add Consumer::on_missing_parent, fired at the orphan-cache site, surfaced through GlobalConsumer and a required SyncTriggerHook on ConsensusActivationParams — required (not optional) so the recovery path cannot be silently left unwired.
• Catch-up task (node): a Notify-driven task pulls proposals via ArchiveClient::get_global_proposal ascending from the engine's finalized frame (tracked separately from the poller-shared head) and submits them (submit_quorum_certificate / submit_timeout_certificate / submit_proposal).

Verification: the devnet rank-1 partition scenario (archive-4 isolated) now passes with the victim genuinely rejoining consensus — it orphans during the partition, then on heal the catch-up supplies the one missing parent frame and the engine finalizes frames 1..4 (state finalized, persisted candidate frames), orphaning stops, and 4/4 nodes reach the stop frame with the chain-safety check clean. Before this change the victim stayed at finalized=0, orphaning every rank.