[farhan-syah]

What changed

Breaking

Change	Where
quic_keep_alive_secs and quic_idle_timeout_secs are removed. Both derive from rpc_timeout_secs.	nodedb-types/src/config/tuning/network.rs
Array tiles route by fmix64(tile_key). ShardBucket replaces the Hilbert-range shard filter.	nodedb-types/src/array_route.rs
The WAL logs resolved rows for KV and vector predicate writes. The old KV predicate record is removed.	nodedb/src/control/write_resolve

Cluster waits

• nodedb-cluster/src/transport: QUIC idle timeout = rpc_timeout_secs. Keep-alive = a fifth of it.
• Startup refuses an RPC timeout below MIN_RPC_TIMEOUT (500 ms).
• One statement deadline runs through the gateway, Raft forwarding, Calvin submit, CRDT admission, DDL and sync.
• The gateway retries a leader change until the statement deadline.
• A remote read gets the full statement budget.

Correctness

• A held redo entry stops its group's apply lane until a snapshot covers it.
• A replica that owes a snapshot cannot lead. Membership keeps one voter that can lead.
• A panicking apply holds its group and answers its waiter. Other groups keep applying.
• Data-group snapshots carry backup cut barriers and cut floors (nodedb/src/types/snapshot_calvin.rs).
• Calvin barrier rows cost one fsync per apply pass, not one per event.
• A departed vShard's Calvin state is retried until removed, after boot recovery finishes.
• Row reads take LockMode::IntentShared on their collection. TRUNCATE invalidates point and index reads.
• The Calvin validator orders collection-level conflicts (nodedb-cluster/src/calvin/sequencer/validator/).

Behaviour you will notice

Before	After
A silent peer kept its connections for 30 s.	It loses them after the RPC timeout (default 5 s). In-flight writes report an unknown outcome.
Background metadata proposals waited 5 s.	They wait up to the statement deadline (default 30 s).
A read on a partitioned node was refused at 0.75 s.	It is refused at the RPC timeout.

Why

• Fixes the stage-1 CI errors on the Replicate Calvin commit, barrier, and read state through the data group #414 merge.

Test	Cause
crash_redo_chunk_stream	A fixed 750 ms read wait. The boot gate opened before replay finished.
calvin_multi_part_node_delete	The same 750 ms read wait.
sql_transactions_kv_overlay	A fixed 5 s metadata wait, reported as XX000.

• Closes the four known limits from Replicate Calvin commit, barrier, and read state through the data group #414: barrier fsync per event, departed-vShard barrier rows, read reply wait, array tile spread.

How to check it

cargo nextest run --workspace --exclude nodedb-cluster-tests --all-features --no-fail-fast --retries 0
cargo nextest run -p nodedb-cluster-tests --all-features --no-fail-fast --retries 0

• Stage 1: 20782 of 20782 pass. Stage 2: 502 of 502 pass.
• Clippy is clean with all features and with none. Every scripts/ci gate passes.

Not covered here:

• nodedb-lite must drop the MetaOp::CalvinFlush arm (adapter/meta.rs).
• nodedb-lite must rename hilbert_range: None to shard_bucket: None (query/visitor/array/query.rs, two sites).
• cluster_pitr_restore_point once failed on CI with 57014 after restore. The root cause is unconfirmed. It passed in the final run.