GC retention: production size refactor#

This append-only record supplements the three GC retention reviews. It evaluates an internal refactor of the published baa7d7d730329fdbf1fab30cc25a0932abc757da implementation, after a normal merge of origin/main at 45c33292b2c4f09a63f856a829c1e15ca2b470aa. The evaluated sources are bound by sizeRefactor.sha256 in the source record. The commit that first adds this record is the evaluated revision. Earlier source bindings and experiments remain historical records.

Preserved laws and smaller machinery#

No oracle model, history grammar, expected result, run budget, or public contract changes in this refactor. The original violated GC law and authority, original production RED, and hostile controls remain in the earlier records. The size request does not identify another product defect: the starting implementation passes the focused behavioral checks but exceeds the requested shipped weight.

The governing laws remain persisting-transaction retention, renewal after the last hold, absolute startup/preload handoff, exact last-unsubscribe delay, pending-rollback deadline preservation, explicit-cleanup precedence, and no restoration of discarded optimistic rows. The optimistic transaction owners also require the same net mutation payload, original snapshots, metadata, settlement outcome and public rows. Cleanup retains its three observation cuts and ordered error diagnostics, including an error whose value is undefined.

The refactor removes duplicated control flow and temporary structures:

• Empty and nonempty transaction commits share completion and publication. Empty commits still complete synchronously without invoking the handler.
• Mutation merging dispatches on the incoming operation without allocating a composite operation string on supported paths. Insert/update and update/update share field and metadata merging while preserving their distinct envelopes. Applying mutations shares one insert-or-replace path.
• The GC timer has one renewal entry point. Settlement preserves remaining handoff time, while last unsubscribe clears it. Explicit cleanup reuses timer cancellation. Automatic cleanup already consumes its queue entry.
• Idle delivery calls cleanup directly after the same eligibility and budget checks. The cleanup promise owns reentrant callers, and sync cleanup selects exactly one synchronous or asynchronous finish path, so no additional finished flag is needed.
• The shared timeout executes due queue work directly, without a single-caller dispatch method or a second empty-queue guard.
• Size calculation counts directly instead of allocating filtered arrays. Collection iteration delegates to its existing entries iterator. Settled transaction pruning uses the existing sorted map's deferred-delete operation instead of a scratch list and a second pass.

Deferred deletion is essential. Ordinary deletion splices the sorted key array while it is being iterated and skips the following transaction. The overlay loop has no callback or ordered transaction-map read between deferred deletions. Restoration of the ordered view therefore occurs at the next ordered read, after the overlay is complete.

Executed behavioral evidence#

The final focused run passes 608 tests in 15 files, with no type errors. It includes GC retention, GC mutation admission, the two-Collection GC property, checker calibration, cleanup/restart, transaction ownership, queue properties, Collection and transaction examples, optimistic payload/composition, outcomes, publication, and settlement boundaries. The GC properties retain all declared reach minima and both their fixed and random campaigns.

Calibration deliberately replaces transactions.delete(id, true) with ordinary transactions.delete(id). The existing optimistic-outcome owner rejects it at rows visible when isPersisted settled: a surviving peer disappears when its preceding settled entry is removed. Eight finite histories and both generated campaigns fail the intended comparison, for ten failures total. The fixed campaign shrinks to seed 851203, path 0:0:0:0:0:0, with absent initial rows, rollback, a non-optimistic first request and a surviving peer. This is an assertion rejection, not a setup error or timeout. The ownership owner passes its eight tests under that mutant and is not credited with this kill. The source is restored byte-for-byte before the final 608-test GREEN run.

Native Chromium passes 3/3 histories on the final source. Real Electric/Postgres passes 12/12 paths on the final built package, including both sibling settlement orders and a genuine SQL-open survivor beyond the first-release deadline. Each path reaches its declared provider-confirmation, renewal and usable-idle cuts. The previous cross-host limits remain: these are separate native-host and real-provider witnesses, not one combined schedule. DB source/test TypeScript, private-member cache, focused ESLint, Prettier and whitespace checks pass.

Commands and local logs are under /private/tmp/pr-2258-size. The full commands for the unchanged receiving drivers remain in the round-three report. No new oracle coverage or universal bug-class closure is claimed by this refactor.

Shipped bytes#

pnpm size:compare --packages db --files uses the normal minified production build, compression settings and app scenarios. The whole PR compares against merge-base 45c33292b on origin/main, rather than against an older PR version. Negative values mean fewer shipped bytes.

The DB bundle and typical app meet the net-negative target in both formats. Compression is not monotonic: the small increases in three feature scenarios are retained explicitly above. They prevent a claim that every application bundle becomes smaller. No names, comments, formatting, compression settings, scenario definitions or minifier options were changed to obtain these results. An attempted removal of the cleanup outcome wrapper reduced raw source but increased Brotli weight, so the existing wrapper and error handling remain.

Production source growth is separate from test/documentation weight. This refactor adds no tests and changes no test expectations. Its documentation records the size result and binds the final production sources; test and contract growth for the broader PR remains in the earlier reports.

Local install recovery#

After the main merge, the registry returns HTTP 403 for baseline-browser-mapping@2.11.32. Version 2.11.31 is absent from the registry; 2.11.30 is the nearest earlier installable release. A local-only workspace override and resulting lockfile changes remain uncommitted. The temporary main baseline uses the identical override. Existing overrides are preserved. This dependency is outside the changed production modules, and the workaround is not included in the pull request.