Rewrite replay worker for per-author DAG buffering

[intendednull]

Context

The replay node (crates/replay/src/role.rs) currently uses a flat VecDeque<Event> FIFO buffer with apply_lenient(). It needs to be rewritten for the per-author DAG model.

What needs to change

ServerData struct

// Old
struct ServerData {
    state: ServerState,
    events: VecDeque<Event>,  // bounded FIFO
}

// New
struct ServerData {
    dag: EventDag,
    state: ServerState,       // cached, maintained incrementally
    pending: PendingBuffer,
    max_events_per_author: usize,
}

Key changes

• Per-author bounded chains instead of single FIFO (prevents chatty peer from evicting quiet peer's history)
• Sync via HeadsSummary comparison instead of StateHash linear scan
• on_event() uses dag.insert() + apply_incremental() + pending.resolve()
• Snapshot fallback when peer is too far behind
• Handle InsertError variants (SeqGap → buffer, Duplicate → skip)

Depends on

• Wire protocol update (WorkerRequest/WorkerResponse changes)

References

• Spec: docs/specs/2026-04-01-per-author-merkle-dag-state-design.md (Section 8)

[intendednull]

Fixed. crates/replay/src/role.rs has been rewritten to the design in Section 8 of docs/specs/2026-04-01-per-author-merkle-dag-state-design.md.

ServerData now holds an EventDag, a cached ServerState, and a PendingBuffer (no more VecDeque). on_event routes events to the correct server via CreateServer hash / known-prev / known-author / default, then try_insert calls dag.insert + apply_incremental, resolving pending events on success and buffering on SeqGap/NotGenesis. Sync requests compare HeadsSummary and return events_since delta, with a Snapshot fallback when the peer is behind. All InsertError variants are handled. Wire protocol updated to WorkerRequest::Sync { server_id, heads } and WorkerResponse::Snapshot { snapshot, post_snapshot_events }.