fix(db): classify Effect deltas from result membership

.changeset/fix-effect-left-join-startup.md

@@ -0,0 +1,5 @@
+---
+'@tanstack/db': patch
+---
+
+Classify Effect deltas from preceding result membership so LEFT-join startup and coalesced source writes emit the correct enter, update, or exit callback, including functional projections with callable or opaque values. Unordered Effects track only result keys and preserve their existing update notifications without comparing or retaining projected values.

docs/contributing/oracle-coverage.md

@@ -1889,6 +1889,19 @@ The [round 7 ledger](oracle-reviews/utils-unit-retirement-round-7.md) retains
 the retired examples. Caller-modified built-ins and Collection publication
 are outside this direct-return owner.
 
+Set performance work evaluated in PR #2301 is excluded by the maintainer's
+2026-10-30 decision: "we've never had this perf problem ... let's remove it."
+The production matcher and the existing `utils.1`–`utils.40` campaigns and
+budgets remain unchanged. The synthetic duplicate-valued object-Set family
+exposes factorial assignment work, and an 8,192-member flat assignment chain
+exposes recursive stack depth. No application incident was established.
+The [versioned replay inputs](oracle-reviews/effect-startup-delta.md#deferred-set-replay-inputs)
+retain the descriptors, observations and limits. The primary owner remains
+`packages/db/tests/utils-oracle.property.test.ts`; resume optimization when the
+maintainer requests it or an application workload requires it. The new
+`utils.41`/`utils.42` work guards and dedicated receivers leave with that work.
+This records a scope decision, not closure of those performance limitations.
+
 The [Duration query equality oracle](https://github.com/TanStack/db/blob/main/packages/db/tests/duration-query-equality-oracle.test.ts)
 uses ten public fields to predict direct BasicIndex/BTreeIndex exact lookups and
 public `eq` snapshots before and after indexing. The `in` evaluator owner also
@@ -5759,9 +5772,10 @@ Held mutations, independent confirmations/outcomes, both settlement orders,
 remote updates, peer exit/reentry and next-use reach 5,512 lane/channel/cut cells
 exactly once. No implicit Effect-root metadata or output Collection ID is claimed.
 
-Open P09: all 64 LEFT-join histories emit a startup update for a never-entered
-unmatched occurrence after its load finished. Those initial comparisons remain
-RED on unchanged `858b174a6`; histories continue and all later comparisons pass.
+Original P09 evidence: all 64 LEFT-join histories emitted a startup update for
+a never-entered unmatched occurrence after its load finished. Those initial
+comparisons were RED on unchanged `858b174a6`; histories continued and all later
+comparisons passed.
 The other 136 native histories and 16 controls pass. Independent hostile
 overlays reject dispatch, previous-value, duplication, colliding-key and extra
 undefined defects; lawful key renaming, fresh reentry and empty partitions retain
@@ -5773,8 +5787,62 @@ in the RC chunk record for exact fingerprints, evidence and review.
 Limits: eager controlled core sources, scalar rows, one mutation per source,
 and these named paths. No real provider/framework, arbitrary scheduling,
 ordered repair, same-source overlap, callback-input sanitization or exact
-internal graph-run callback-count guarantee follows. P09 requires production
-repair after the full oracle campaign; this entry makes no closure claim.
+internal graph-run callback-count guarantee follows.
+
+Historical initial P09 repair evidence for PR #2301:
+The exact owner and `test:oracles` entry were restored before production edits.
+On baseline `9b03e639f`, the fresh RED reproduced 64 startup failures with every
+reach cell at count 1. Repaired production passes all 216 owner tests and all
+2,756 checkpoint comparisons. Restoring the original classifier reproduces the
+same 64 assertion failures; suppressing all unordered updates fails all 200
+native histories. Each of the 5,512 reach cells remains at count 1 in both
+mutants and GREEN. All 1,690 neighboring tests pass before and after the edit.
+See [the Effect startup follow-up](oracle-reviews/effect-startup-delta.md) for
+commands, fingerprints and limits. Arbitrary scheduling, join multiplicity,
+ordered repair, same-source overlap and callback-input sanitization remain
+outside this bounded closure. Other deferred findings remain open.
+
+Subsequent checker calibration keeps both callback channels coherent when
+injecting a bad update key. The `key-collision` and `key-change` controls now
+reject deletion of the interval-continuity judgment itself: the full checker
+mutant passes 214 controls/histories and fails those two controls. Repaired
+production passed all 216 owner tests and the combined 619 targeted tests at
+that historical subject, including subsequently removed Set work.
+Native histories and all 5,512 reach cells are unchanged. See the follow-up's
+[calibration evidence](oracle-reviews/effect-startup-delta.md#key-interval-checker-calibration)
+for the new owner fingerprint and the limits of the continuity claim.
+
+The subsequent high-review repair extends the same owner with 24 finite
+histories crossing scalar/callable/private-field payloads, one source/LEFT join,
+initially absent/present membership, and separate/coalesced source transactions.
+The independent source snapshots still determine the full delta. Both callback
+channels preserve event multiplicity, previous readable payloads, and active key
+intervals across initial, sequence, absent, reentry and next-use cuts. All 120
+new reach cells have count 1. On unchanged `d97c1aaf4`, 14 histories fail;
+retaining preceding result membership repairs them without relying on payload
+equality. The original classifier and a net-count-only classifier fail these
+comparisons. This finite extension does not claim exact rich-payload instance
+identity, arbitrary nondeterministic projections, changing join keys, provider
+schedules or framework behavior. The earlier scalar and ordered-repair limits
+remain. See the [Effect membership evidence](oracle-reviews/effect-startup-delta.md#membership-and-coalesced-source-writes).
+
+Current key-only repair: the metadata owner has **300 tests**, including 60
+callback-profile histories from `query/effect-callback-oracle.ts`. The actual
+React driver `packages/react-db/tests/effect-callback-oracle.test.tsx` runs the
+same 60 histories, model and comparison. Six payload representations cross
+five callback profiles and `skipInitial`. Source-version updates, handler
+replacement, exit, reentry, next use and disposal reach 480 cells on each path,
+each at count 1. The original 5,512 cells and 120 membership cells remain.
+On unchanged `2b3dd8c2d`, 56 core and 40 React histories fail. The repair passes
+all **708 targeted tests**. The stale-value mutant survives the original 314
+tests and fails the extended histories at exit; gross-count membership fails
+79 comparisons. Unordered Effects now retain keys only and never compare
+projected values. Ordered repairs retain their preceding coherent values.
+The maintainer explicitly chose preservation of unordered update notifications,
+including equal projections, on 2026-10-30. The new histories cover eager core
+sources and actual React hook commits; other providers/frameworks, arbitrary
+schedules and ordered repair semantics retain their existing limits. See the
+[key-only review evidence](oracle-reviews/effect-startup-delta.md#review-of-2b3dd8c2d-key-only-unordered-results).
 
 ### RC bound schema mutation receivers (2026-10-28)
 

docs/contributing/oracle-reviews/2026-10-29-pr-2171-deferred-oracles.md

@@ -249,3 +249,14 @@ settlement and after it. The original oracle remains active and unchanged;
 there is no skip, narrowed comparison, retry-away, GREEN, or bug-class closure
 claim. Its CI RED remains an open baseline defect under this explicit release
 scope deferral.
+
+## P09 follow-up (2026-10-29)
+
+[PR #2301](https://github.com/TanStack/db/pull/2301) restores the exact accepted
+Effect owner and repairs its LEFT-join startup delta. Fresh unchanged-production
+RED precedes the production edit. The owner passes all 216 tests after repair,
+with every one of its 5,512 reach cells at count 1. Original and overcorrection
+mutants fail the shared comparisons. The
+[versioned follow-up](effect-startup-delta.md) records commands, fingerprints,
+neighboring guarantees and the bounded closure. This updates P09 only. All
+other deferrals and limits in this release record remain open.