Record a base copy's checkpoint bound in the same transaction as its audit-less rows #1134

Problem

A base copy writes rows without log entries (isCopyApply). Checkpoint admission must therefore treat a route's certifiedFrom as void when it is at or below the highest version a copy wrote for that origin. Today the copying connection keeps that bound in its own [copyBound, peer] row in __dbis__. It commits the bound before handing a copied row to core's apply loop: a write-ahead that runs about once per origin per copy pass (harper-pro#1092).

Recording the bound in the same transaction as the copied rows would remove that write-ahead entirely. It is not crash-safe today:

• A rocksdb-js transaction takes its WAL setting from the store it was opened on (src/binding/transaction/transaction_handle.cpp, writeOptions.disableWAL = dbHandle->disableWAL).
• Table stores open with the WAL off (core/resources/databases.ts, openRocksDatabase); __dbis__ opens with it on.
• Column families flush independently. After a crash, a table column family can hold a flushed copied row while the __dbis__ memtable that held its bound is gone. Admission would then accept a stale certifiedFrom.

What it would take

One of these:

• copy transactions written with the WAL on;
• atomic flush across the database's column families;
• a bound key stored in each table's own column family, which admission would then read per table.

Each one changes the audit-less copy durability contract, so it is a separate design.

Value

It removes the per-origin write-ahead (one awaited __dbis__ commit per origin per pass) and allows an exact per-origin bound instead of one reserved through the copy start. Low priority: the current design is correct, and its cost is about one commit per origin per pass.

Refs harper-pro#1092, harper-pro#922.

Review-Deferral: d22a669f4b94c44b severity=nit category=performance