nx_rollback_session.nx

nx_rollback_session.nx -- GGPO-class rollback netcode session handle. Composes the five foundation primitives: - nx_sim_snapshot : ring of last-N state snapshots - nx_state_delta_codec : 16-byte input packets via bit-packed varint - nx_frame_hash : per-frame state hash for desync detection - nx_zigzag_varint : signed-integer wire encoding - (caller's sim primitives -- whatever advances the game state) ...into a single session handle that game code holds for one multiplayer match. Modelled directly on GGPO's session callbacks (Cannon 2006 / pond3r/ggpo) but adapted to NishiLang's callback- less calling convention: storage-only, caller drives the sim loop. Caller pattern per tick: 1. read local input -> save via session_save_local_input 2. snapshot state -> save via session_save_state 3. advance sim using local + last-known remote input 4. compute frame_hash -> save via session_save_hash 5. transmit (tick, local_input, frame_hash) to peer 6. when (peer_tick, peer_input) arrives: a. session_set_remote_input(peer_tick, peer_input) b. if peer_tick < current_tick: - session_restore_state(peer_tick) into sim state - re-advance from peer_tick to current_tick with freshly-corrected remote input each tick 7. when (peer_tick, peer_hash) arrives: a. our_hash = session_get_hash(peer_tick) b. verdict = nx_frame_hash_compare(...) c. if MISMATCH -> trigger desync recovery Session layout (i64 cells in a flat block): h[0] = state_size h[1] = max_rollback_frames (snapshot/input ring depth) h[2] = current_tick (highest local frame saved) h[3] = oldest_pending_remote_frame (= highest local frame we have NOT yet received remote input for; before we can rollback to frame F we need remote input at F) h[4] = snapshot_ring_ptr (nx_snapshot_ring_alloc handle) h[5] = local_input_ring_ptr (parallel i64 array, depth slots)