Effect startup delta follow-up#

Current implementation: unordered Effects track keys only and preserve graph update notifications. The key-only review repair supersedes the earlier retained-value/comparison designs below. Earlier counts, fingerprints, commands and design claims apply only to their named subjects.

Baseline restoration#

This follow-up restores the accepted Effect metadata owner from db38b4e009146af53e46aeeef3206269f5968af8, without editing its model, grammar, driver, recorder, comparisons, or reach. Its SHA256 is 4b738962e30f2b0f55b1894bab1f76334da8d66e6001773bbe24599408bb6130. The shared source-metadata model and cleanup support already exist on main. The owner also rejoins the package's explicit test:oracles command.

The baseline is main bea84cd4554b7bd49df50cb0c569a34d8b2cd915. Production was unchanged in baseline restoration commit 9b03e639f769502dccccff7656ba85ab4ca124e8. That commit establishes a separately reviewable instrument before the repair. No execution result was claimed at restoration. The maintainer subsequently authorized targeted oracle tests, while retaining the prohibition on full builds. The dated evidence below supersedes the pending execution status.

Law and preserved evidence#

Effect callbacks deliver the query's full before/after delta. An update requires an occurrence that previously entered; a startup-only intermediate change cannot invent such an occurrence. Explicitly selected source metadata, previous values, event multiplicity, and active event-key intervals are part of the comparison. Authority: public EffectConfig and DeltaEvent, the Effect guide, VirtualRowProps, and the source settlement law. Batch partition is unconstrained; the check allows empty batches and opaque event keys.

The reviewed pre-release owner has 200 finite histories across structured and functional projections, single sources, LEFT joins, and both populated-union orders. It varies independent optimistic outcomes, confirmations, settlement order, remote followups, peer exit/reentry, and surviving-row next use. It checks both batch and named callbacks at every declared action cut, continuing after a mismatch so startup failure cannot erase later evidence.

The preserved review records 152 passes and 64 LEFT-join startup REDs, all 5,512 declared lane/channel/cut cells reached, and 2,692 of 2,756 checkpoint comparisons passing. The failing checkpoint emits an update for a never-entered unmatched occurrence. Five hostile callback overlays add intended comparison failures; legal key renaming, fresh reentry keys and empty partitions retain the baseline. These are historical results, not fresh results on this branch.

The complete applicable ORC-001 through ORC-016 evaluation, fingerprints, calibrations, omission attacks and execution commands remain in chunk 16 of the accepted campaign record. The release deferral retains the broader P09–P12 and type-receiver scope. This restoration changes none of those verdicts.

Baseline repair plan (completed below)#

Before production changes, run the restored owner on unchanged main and retain the actual callback mismatch and all per-cell reach. Then run the same owner after the repair and with the original defect restored as a hostile mutant. Preserve neighboring single-source/union behavior, both callback channels, previous values, empty partitions, and later optimistic settlement/next use.

TANSTACK_DB_ORACLE_STATISTICS=1 pnpm --filter @tanstack/db exec vitest run \
  tests/query/virtual-row-fields-effect-oracle.test.ts --coverage.enabled=false

Limits remain eager controlled core sources, scalar projections, one mutation per source, and the named paths. This supplies no real-provider/framework, arbitrary scheduling, ordered repair, same-source overlap, or callback-input sanitization evidence. No GREEN, production repair, or bug-class closure is claimed by this baseline restoration.

Executed repair evidence (2026-10-29)#

Review identity: baseline restoration 9b03e639f769502dccccff7656ba85ab4ca124e8, with production still identical to main bea84cd4554b7bd49df50cb0c569a34d8b2cd915. The repaired packages/db/src/query/effect.ts has SHA256 b9ea784afe07e156507079224885c73bd21e8633116c3344af809c09971b0879. The owner retains the restoration fingerprint above. No model, history, recorder, comparison, reach requirement, or run budget changed after RED. This record accompanies the production repair in PR #2301.

Law, diagnosis, and repair#

An Effect update describes a changed row that stays in the result. For a preloaded LEFT join with a matched row and an unmatched anchor, startup owes only the two entering occurrences. It owes no update for an intermediate unmatched occurrence that never entered. The public Effect guide's delta sections and DeltaEvent.previousValue establish this judgment independently of the graph output.

The original classifier treated any insert plus delete as an unordered update. Startup emits and retracts an intermediate unmatched row within one graph run. Its inserted and deleted scalar projections are equal. The callback nevertheless received that row as an update, followed by the two legitimate enters. The update's previous and current values were identical, with all right fields undefined. The independent before/after model rejected the extra event and the absence of its prior enter in both callback channels.

The initial repair applied the existing deepEquals update condition to unordered Effects. That design used no retained unordered result rows and checked only scalar projections. The later membership repair retained values, and the final key-only repair removes both unordered value retention and comparison. These historical designs are not current implementation claims.

The earlier Effect examples lacked this exact startup delta comparison. The restored owner checks the complete emitted delta and active key intervals, rather than inferring correctness from final rows or selected event counts. No missing-input hypothesis remains for this witness: both source Collections finish preload before Effect creation, and the correct matched and anchor enters arrive in the same startup ledger as the extra update.

Results#

The pre-edit inventory comprises structured and functional projections through actual core createEffect, with onBatch and named handlers. Single sources, LEFT joins and both populated-union orders receive the same source model. There is no real-provider or framework claim requiring another receiving host. The unchanged-production neighboring run preceded the production edit.

The 200 native histories comprise 8 single-source, 64 LEFT-join and 128 union histories. The other 16 tests calibrate the shared judgment. Both structured and functional LEFT-join startup cuts fail in RED. Every later cut continues and passes, so startup failure does not hide settlement or next-use evidence. The original mutant restores the entire production file byte-for-byte to the baseline and reproduces the same startup assertion failures. The second mutant requires publishedRows for updates, wrongly dropping all unordered updates. It fails the intended changed-value comparisons in all 200 native histories. Both outcomes are assertion failures at the shared comparison, not setup errors or timeouts. Neither mutant remains in the source.

Neighboring suites pass 1,690 tests both before and after the production edit: effect.test.ts (74), effect-disposal-oracle.test.ts (12), and query/virtual-row-fields-oracle.test.ts (1,604). These retain skipInitial, previous values, object-valued Set updates, source failure, ordered Effects, disposal and root metadata observations within their existing bounds. The baseline remote DB-1 job also fails the same 64 startup cases: run 40930211146.

All local tests use Node 24.19.0 and installed Vitest 3.2.4 on macOS arm64. The manifest requests Vitest 4.1.12. No dependencies or lockfiles changed. Full builds remain unrun by maintainer instruction.

Commands run from packages/db using the worktree's installed binary:

TANSTACK_DB_ORACLE_STATISTICS=1 ../../node_modules/.bin/vitest run \
  --coverage.enabled=false tests/query/virtual-row-fields-effect-oracle.test.ts

../../node_modules/.bin/vitest run --coverage.enabled=false \
  tests/effect.test.ts tests/effect-disposal-oracle.test.ts \
  tests/query/virtual-row-fields-oracle.test.ts

TANSTACK_DB_ORACLE_STATISTICS=1 ../../node_modules/.bin/vitest run \
  --coverage.enabled=false tests/query/virtual-row-fields-effect-oracle.test.ts \
  tests/effect.test.ts tests/effect-disposal-oracle.test.ts \
  tests/query/virtual-row-fields-oracle.test.ts

Each executed command also uses default and JSON reporters. Full logs and JSON reports remain in /private/tmp/effect-startup-delta/ under baseline-red, baseline-neighbors, repaired-green, repaired-neighbors, original-mutant, overcorrection-mutant, and final-green. The finite owner reports every named reach cell, so the table's count 1 applies to each lane/channel/cut separately. It does not aggregate away a missing path. No random seed applies to this exhaustive finite grammar. Running the whole owner reconstructs every witness.

Self-review against the oracle rubric#

The accepted chunk-16 audit remains applicable to the unchanged owner. This follow-up adds executed evidence on the restored baseline and repaired source. The review here is a self-review, not a fresh independent review.

Closure and remaining limits#

P09 is repaired within this contract × history × production-path × observation boundary: exact scalar deltas for eager preloaded core Effects, the 200 stated histories, both projection forms and both callback channels, at every declared action cut. No known counterexample remains inside those exercised cells. This does not establish arbitrary join multiplicity or scheduling, changing join keys, same-source overlap, ordered repair, real providers/frameworks, or callback-input sanitization. P10, P11/P12 and the other deferred runtime, retention and public-type findings remain open under their existing owners. No broader metadata or general delta-classifier closure follows.

Subsequent prep-pr review#

Review of a7ea76cddf7ab2fe03f49d84c6d596dd598fedba found a newly exposed factorial object-Set comparison path. The historical Set-work repair records its independent model, pre-edit RED, receiving paths, GREEN and mutants. The scalar startup owner's native model and histories still pass all 216 cases. Its separate opaque-key calibration repair is recorded below; startup and work evidence alone would not close that test-integrity gap.

Key-interval checker calibration#

CodeRabbit review 5703351893 on a7ea76cddf7ab2fe03f49d84c6d596dd598fedba identified a compound-fault calibration. The key-collision and key-change controls changed only batch keys. Named callbacks retained their original keys, so cross-channel disagreement rejected the ledger even without the interval judgment. The checker still enforced continuity; these controls did not isolate it. ORC-006 requires sensitivity at the intended judgment.

The maintainer accepted this repair. Both controls now put the same bad key in the matching named entry. events() constructs named entries directly from the batch, preserving their order. No production behavior, independent model, native history, assertion, exclusion, run budget or reach requirement changed.

Before repair, deleting the interval judgment leaves both controls passing: the checker mutant survives. After coherent mutation, that deletion causes exactly two expected-to-throw assertion failures. Removing only the key-stability assertion from the full repaired owner gives 214 passes and those same two failures, with zero skipped tests. Cross-channel agreement therefore cannot substitute for continuity enforcement.

At that earlier head, the repaired owner passed all 216 tests. Combined equality, Collection/Effect receiving, proxy-revert and startup checks passed 619/619, including the later-removed Set work. All 5,512 startup reach cells still count 1; Set-work reach is unchanged. Targeted TypeScript, ESLint, oracle lint and formatting pass. Tests use at most two workers; no full build ran. Other applicable rubric outcomes above remain unchanged.

This is a test-integrity repair with a checker mutant, not a new production RED. These update faults calibrate continuity for a live occurrence. They do not independently mutate the distinct-key condition on a fresh enter, or establish arbitrary graph, provider, framework or rich named-callback coverage.

The restoration and execution fingerprints above remain historical evidence. The current owner intentionally changes only key-fault setup and its explanation. Its SHA256 is 56de9e93892c6a4772f71846e25fc9e3700359b2cadbe3d6d9b359d7b9bcd46f. Review identity is parent a7ea76cddf7ab2fe03f49d84c6d596dd598fedba plus that fingerprint and the Set-work fingerprints. The combined command is recorded in the Set-work review. Raw checker variants, reports and logs remain under /private/tmp/effect-startup-delta/prep-pr/key-calibration/accepted-cr1/.

Membership and coalesced source writes#

The high review of d97c1aaf412661e8d5d3e0d528fe5c4667f69d65 exposed a remaining membership distinction. An absent key can receive insert/delete/insert in one flush and must enter. A present key receiving delete/insert/delete must exit. Fresh callable and private-field projections also exposed a startup update for an unmatched occurrence that never entered. Presence cannot be inferred from payload equality or gross insert/delete counts.

The existing owner adds 24 finite histories: scalar/callable/private payload, one source/LEFT join, initially absent/present and separate/coalesced writes. Independent source snapshots determine the whole expected delta at initial, sequence, absent, reentry and next-use cuts. Both callback channels compare event multiplicity, previous readable values and active opaque-key intervals. All 120 new cells reach once. The original 200 histories and 5,512 reach cells stay. The rich recorder preserves payload kind and readable result, without a claim about exact re-evaluated instance identity.

Before production edits, the expanded owner on unchanged d97c1aaf4 failed 14 of these histories at delta/interval assertions. The complete recorded run passed 631 tests and failed 16, including two separate Set stack probes. The membership repair passes all 240 owner tests. Restoring the original classifier fails 14 histories; inferring prior presence from net count alone fails all 24. Public sync writes inside a real optimistic transaction's scheduler scope reach the reported three-write histories. No production information is missing at these observation cuts.

Every Effect now retains its preceding coherent result, including a skipped initial baseline. Prior membership plus net multiplicity decides membership at the next cut. Only a continued row compares values to decide an update. An Effect without update/batch handlers refreshes retained values without that comparison or an update dispatch, preserving its later exit's value. Public EffectConfig, DeltaEvent and the Effect guide supply the authority.

The Effect production and owner files are unchanged from the verified subject 375e9b26ca93b19cc266dd1ccdfcd10e97e820fc. Their SHA256 values are:

6835de1fc044efec14f80ce7aa546451c9c1ce6134a1c5198779ee5ebd45c8b9  packages/db/src/query/effect.ts
469415454de90a75db84caaa5b6893ae8fb54a1f85536872070f9a3c54ded1a6  packages/db/tests/query/virtual-row-fields-effect-oracle.test.ts

The exact-head high-review record preserves all eight original findings, experiments, hostile variants, grammar controls, rubric outcomes and reviewer assessment. That historical record's 647-test result includes the now-removed Set work. It is not the current count. Current claims remain bounded to the declared eager controlled core histories. Arbitrary provider/framework schedules, changing join keys/multiplicity and exact rich-payload instance identity remain outside this instrument.

Maintainer scope revision, 2026-10-30#

The maintainer explicitly removed Set performance work from this PR because no application performance problem had been reported. The old Set matcher returns byte-for-byte, together with the existing utils.1–utils.40 owner and registry. New Set models, receivers, properties 41/42, scale/work guards and the dedicated Set report are removed. Existing laws, comparisons, reach and budgets remain. The Set work's own new guards retire only with this explicit scope decision. Their known synthetic limitations remain recorded under the primary deep-equality owner in the coverage map; resumption requires a maintainer request or a real workload. Historical generators and full traces remain in commit 0b47ee825 and /private/tmp/effect-startup-delta/high-review/.

The high-review ledger now has three Effect repairs (1, 2, 4), one accepted-design semantic finding with a release-note correction (3), three explicitly deferred Set findings (5–7), and one refuted coverage allegation (8). The original factorial Set work is deferred with those findings. None is relabeled refuted or closed by removing its test. The Effect membership repair, checker calibration, callback comparison bypass and equal-projection suppression remain intact.

Validation after Set removal#

The six retained suites pass 581/581, with zero failures or skipped tests. They cover the existing change-event equality owner, ordinary utils tests, proxy revert, ordinary Effects, Effect disposal and the Effect metadata owner. All 240 metadata-owner tests pass. The original 5,512 reach cells and 120 added membership cells each reach count 1. Targeted TypeScript, ESLint, oracle lint, formatting and whitespace checks pass.

The executed local configuration uses the installed dependencies, jsdom, tests/test-setup.ts, a source alias for @tanstack/db-ivm, no coverage, and at most two workers:

node node_modules/vitest/vitest.mjs run \
  --config /private/tmp/effect-startup-delta/remove-set/focused.config.mjs \
  --reporter=default --reporter=json \
  --outputFile=/private/tmp/effect-startup-delta/remove-set/focused.json

The config, log and JSON report remain under that directory. The source alias changes dependency resolution only. It does not replace the production driver or independent comparison. The normal repository selectors name the same six test files. Every test:oracles path exists after the removed receiver leaves the script.

The maintainer separately authorized pnpm size:compare, including its normal package builds. Against merge base bea84cd4554b7bd49df50cb0c569a34d8b2cd915, the typical app remains 96,732 bytes gzip (0-byte change). The Effects app changes from 99,303 to 99,340 bytes gzip (+37 bytes). All other app scenarios remain unchanged. The DB package adds 20 bytes gzip. The log is /private/tmp/effect-startup-delta/remove-set/size-compare.log.

These results supersede the earlier combined 619/647 counts for the current scope. Those counts remain historical evidence for their stated subjects. The Effect source and owner fingerprints above still identify the verified implementation. Set production, its former utils.1–utils.40 owner and its registry are byte-identical to the merge base. No full repository test or build ran beyond the separately authorized size comparison's normal package builds.

Review of 2b3dd8c2d: key-only unordered results#

The ten-item review targets 2b3dd8c2d0a0f98a858dc1d08182f6673c9a7836. It supplied no executed tests or builds. This evaluation is a self-review. The maintainer explicitly chose: "Preserve existing notifications; use keys." That decision supersedes the earlier general equal-projection suppression policy recorded above. No oracle expectation changed to match production. The new model ran RED before production changed.

Law, experiment and inference#

EffectConfig, DeltaEvent and the live-query guide promise complete delta values at each callback cut. Handler subsets do not change an exit's current value or an update's pre-batch value. React promises current callbacks without remounting. The maintainer's decision preserves unordered graph notifications, including equal projections, without adding value comparison or retention.

The primary owner adds a shared companion, effect-callback-oracle.ts, and an actual React receiving driver. Sixty finite histories on each path cross:

• Scalar, keyless opaque class, keyed private-field class, prototype method over non-enumerable state, object Set and equal scalar projection.
• All callbacks, enter only, exit only, late update and late batch callbacks.
• Initial delivery and skipInitial.

Authored scalar source versions determine expected public readable payloads. Classes and Sets are constructed only in the actual functional projection, so source-Collection comparison cannot mask or create the reported output cost. The model never imports production equality or its delta classifier. Both paths compare initial delivery, an unobserved update, callback replacement, another update, exit, reentry, next use and disposal. Source updates while absent are excluded. Removing any profile loses its callback premise; removing any payload loses its specific comparison boundary. All combinations run, and every one of 480 declared cells on each receiving path reaches once. Original 200/24 histories and their 5,512/120 cells remain unchanged.

The recorder reads actual public payloads, including private-state accessors, and retains event multiplicity and previous values. Synthesized Set getters count reads outside the recorder. The restored unordered path traverses no payload leaves, for any callback profile. This bounds added traversal in these histories, not Set runtime or arbitrary graphs. The original metadata histories still check opaque-key intervals and both channels independently.

All ordinary runs and full mutant runs preserve per-path reach. The diagnostic old-test selection is not a normal GREEN claim. Initial React module resolution failed before collection; explicit aliases to the already-installed React 19.2.4 fixed that setup before the valid pre-production RED. No setup failure counts as a product or mutant result.

The final repair retains only keys for unordered Effects. The first deletion in a batch supplies its previous/exiting value. Ordered Effects still retain coherent values across held graph runs and use their existing comparison. observesUpdates and its creation-time snapshot are removed. Dynamic callback lookup remains live. The redundant value fallbacks are removed; deleteValue is now necessary for unordered delivery rather than dead state.

A deterministic cardinality probe at 8 and 64 rows confirms that reviewed unordered Effects added one Map entry per result, holding the same row objects as callbacks. It did not clone the rows. Disposal cleared that Map. The repaired path retains no output values in a Map. No total-memory multiplier was measured.

Lossless disposition#

Accounting: 10 = 9 fixed-now + 1 duplicate. Each technical qualification above survives its action verdict. No new item is deferred. Previously removed Set engineering work retains the maintainer's explicit scope decision and its primary equality owner; resumption requires a maintainer request or an actual workload. Removing the new unordered comparator caller resolves the new interaction without reopening that optimization project.

Reviewer assessment: lean hire for static defect discovery, with verification required before approval. Callback profiles, live replacement and React wrappers expose important missing cases. The memory multiplier, blanket private class claim, unconditional link-loss claim and field-deletion suggestion needed qualification. The proposed membership-only direction was useful. Severity and closure require executed evidence.

Oracle review and remaining limits#

This is bounded enforcement of unordered membership, notification compatibility and callback-visible values for the declared eager source histories and actual React commits. No known violation remains within these cells. It does not prove arbitrary scheduling, provider delivery, changing join multiplicity, other frameworks, ordered opaque-value semantics or exact rich-instance identity. Other comparator callers retain their previously documented domains and limits.

The final run selects utils oracle, utils, proxy revert, Effect, Effect disposal, Effect metadata, React Effect callbacks and ordinary React Effect tests. It uses jsdom, existing test setup, source aliases for DB/IVM, installed React 19.2.4, Vitest 3.2.4, no coverage, and at most two workers. Standard selectors are:

pnpm --filter @tanstack/db exec vitest run --coverage.enabled=false --maxWorkers=2 \
  tests/utils-oracle.property.test.ts tests/utils.test.ts \
  tests/proxy-revert-oracle.property.test.ts tests/effect.test.ts \
  tests/effect-disposal-oracle.test.ts tests/query/virtual-row-fields-effect-oracle.test.ts
pnpm --filter @tanstack/react-db exec vitest run --coverage.enabled=false --maxWorkers=2 \
  tests/effect-callback-oracle.test.tsx tests/useLiveQueryEffect.test.tsx

Local execution uses the same selectors in the source-alias config /private/tmp/effect-startup-delta/review-2b3dd8c/final.config.mjs. Reports, commands, original/mutant snapshots and the append-only raw review ledger remain in that directory. These local files are supplementary to the versioned histories, recipe and comparison. Targeted TypeScript and ESLint pass. The authorized pnpm size:compare reports typical app 96,732→96,732 bytes gzip, Effects app 99,303→99,309 (+6), and DB bundle 109,667→109,668 (+1), against merge base bea84cd45. Other app scenarios are unchanged. No full test/build ran beyond the size script's normal package builds.

Exact subject: reviewed parent plus the following final source fingerprints. Documentation-only changes do not alter these executable subjects.

354bbbcf9683c91047c61da1ea2323ed89623628eb6059dacfbb79bd56b85ce4  packages/db/src/query/effect.ts
47e862b092e643309bd804acab7c61c036ed7a345c873ca480f63b807aa4f24d  packages/db/tests/query/virtual-row-fields-effect-oracle.test.ts
ca9e1336e7ceeb243a8e32a49ad2c3c881deaa38334540e85f848122cbaf2d38  packages/db/tests/query/effect-callback-oracle.ts
9ed573ae14bf9bd87058c9136ef1de57f33422da235f6253a9c97f90599cf98e  packages/react-db/tests/effect-callback-oracle.test.tsx

Deferred Set replay inputs#

These inputs preserve the earlier synthetic findings without reinstating the Set rewrite, test campaigns or work budgets. No application incident or new latency guarantee follows. The existing utils-oracle.property.test.ts owner remains responsible for change-event equality; the maintainer controls whether to resume the excluded engineering work.

The factorial-work witness uses distinct object members. For size n, compare new Set(Array.from({length:n}, () => ({value:0}))) with a Set of n-1 such objects followed by {value:1}. The equality judgment is false. Instrument an enumerable value getter to count reads. The earlier direct n=4 experiment observed 128 reads, against its proposed bound of 32. The eight-member Effect receiver observed 219,200 reads against 1,296 while still returning the correct equality result. Those engineering bounds are historical and retired.

The stack witness uses n=2**13 words of 13 coordinates. Let g(j)=j^(j>>1). For right word j, coordinate k is a distinct instance of class First or Second according to bit k of g(j), with one enumerable value:0 field. Two different classes differ; a plain {value:0} can match either class. Left word i<n-1 follows right word i, replacing the coordinate that differs from right word i+1 with a plain record. The last left word copies right word zero. Each outer Set member is its word array. The complete assignment is i -> (i+1)%n, so equality is true. To create the false neighbor, change every coordinate value in the last right word to 1, leaving an isolated unmatched word. The earlier recursive matcher overflowed on both witnesses after setup. This is a runtime failure, not an equality-assertion rejection. Smaller generated versions used 2–6 bits and optional one-place right rotation. A recorded greedy mutant replay used utils.42, seed -75265719, shrink path 1; that retired selector alone does not run on the current registry. The structural recipe above supplies the inputs independently of the removed generator.

Historical reports at 0b47ee825aaa0c36b199a3c42652cab8d9672ec9 retain broader experiments. Their links are supplementary; the current source tree now holds these essential distinguishing inputs, observations and limits.