[citizen-stig]

With rollback enabled, Session::finish calls ReverseDeltaBuilder::finalize before it starts the merkle update. finalize waits for the speculative prior-value loads and fetches the priors of any written keys that weren't preserved in advance. All of that happens before any merkle work begins.

Changes

• finalize now runs after update_and_prove has been kicked off, so the delta is built while the merkle workers run.

This is safe because both sides only read: the merkle update reads pages and leaves, and the delta worker reads values through its own beatree read transaction (rollback/reverse_delta_worker.rs). The store isn't modified until the FinishedSession is committed, and a sync waits for open read transactions. The resulting delta is the same.

Unrelated to #947, which moves where the delta is persisted (inside the store sync). This PR changes when it is built (during Session::finish).

Touches the same lines of Session::finish as #951. Whichever lands second needs a trivial rebase.

Performance impact

after benchmarking from #944 (rollback enabled, e.g. NOMT_BENCH_ROLLBACK=<len>).

With rollback enabled (NOMT_BENCH_ROLLBACK=100, transfer workload, 16 paired rounds on a dedicated NVMe), Session::finish drops from 92.1 ms to 71.4 ms (−22.6%), mean block time drops by 11.2%, and throughput rises by 13% (63.7k → 72.5k ops/s). With an overlay window of 4, the gains are similar: finish −26.8% and block time −12.9%. This removes almost all of rollback's cost from finish, which now matches the rollback-disabled time of 73.5 ms. With rollback disabled the change is a no-op, and the durable commit time is unaffected.

Combined with #947 (rollback delta persisted inside the store sync), the durable commit also drops by 13.6%, so mean block time falls by 17.6% and throughput rises by 21.7% (63.9k → 75.8k ops/s) compared with master.

  ┌────-──────┬─────────────┬─────────────────────┬───────────────────────────┬────────────────────────────┐
  │  vs base  │   finish    │ durable_commit mean │     Block time, mean      │      Block time, p50       │
  ├───-───────┼─────────────┼─────────────────────┼───────────────────────────┼────────────────────────────┤
  │  952 only │ −22.9%, 4/4 │ +0.7% (CI ±37)      │ −11.0% (CI ±16)           │ −7.9% (CI ±23)             │
  ├─────────-─┼─────────────┼─────────────────────┼───────────────────────────┼────────────────────────────┤
  │ 952 + 947 │ −23.6%, 4/4 │ −14.0% (CI ±19)     │ −18.0% [−26.6, −9.4], 4/4 │ −17.7% [−18.7, −16.7], 4/4 │
  └─────────-─┴─────────────┴─────────────────────┴───────────────────────────┴────────────────────────────┘