[McHersheys]

Fixes #8.

Change

Only pending() changes at runtime: complete one bounded .get() before each yield instead of suspending .iterate() across the client's awaited object PUT. This releases the WAL read snapshot before acknowledge() writes.

• BigInt rowid keyset preserves insertion order and integer precision without materializing the graph.
• CROSS JOIN keeps the rowid seek outermost; the query-plan regression rejects a per-page whole-job sort.
• ACK validation, retries, wire format, schema, and cache custody are unchanged. No catch/retry workaround, cache reset or long transaction.
• The cursor relies on the cache's existing native positive-rowid inserts; it does not add support for arbitrary manual rowid rewrites or promise general parallel-mutation safety.

Runtime diff: 12 added / 3 removed lines in one function. The remaining changes are eight focused real-file cache regressions and one real paged-client regression, reusing the existing client fixture and encrypted page/journal implementation.

Validation

Based on current main aeaf2ca017f2539c24a8a4e1b076fdd2160a589f; the affected runtime closure is unchanged from 0.9.2 a6a17edba700a595faa7acc281a58146cb625e23. Linux, Node 24.21.0, embedded SQLite 3.53.4; fresh synthetic homes, two actual WAL connections, no external requests or credentials.

Exact focused boundary	Pass	Fail	Skip
New race/client tests + causal controls, unchanged main cache	3	2 expected (database is locked at ACK)	0
Patched cache + all nine new regressions	9	0	0

node --test --test-name-pattern='^pending |^paged upload ACK' \
  test/relay-page-cache.test.js test/relay-page-client.test.js

Also passed syntax checks for all three changed files and the existing wall-clock source check for both changed test files. The cache regressions cover the three causal contrasts, lazy/source-bound order, ACK persistence/reopen, another uploader's ACK, rowids beyond Number's safe range, and rowid-seek/no-sort query planning. The client regression commits on a second connection during the awaited real object PUT, then checks upload completion and exact native restore.

Earlier exact-runtime validation is reused rather than rerunning unaffected suites. The full repository gate and browser tests were not rerun for this PR. Upstream review/merge remains separate from these focused results.