fix(angular-db): align version support and query reactivity

.changeset/angular-query-snapshots.md

@@ -0,0 +1,5 @@
+---
+'@tanstack/angular-db': major
+---
+
+Require Angular 20 or later. Keep live-query result signals coherent with the current query before effects run, and release observer subscriptions through effect cleanup. Track the observer's row-state identity separately so unchanged row state does not rebuild the adapter's defensive Map copy.

docs/contributing/oracle-coverage.md

@@ -2864,6 +2864,28 @@ Angular's JIT test environment. It does not exercise Angular's AOT compiler,
 older Angular runtimes, or every scheduler cut. The same driver checks that an
 explicitly disabled query reports `disabled` and ready before effects run.
 
+The [Angular snapshot conformance oracle](https://github.com/TanStack/db/blob/main/packages/angular-db/tests/snapshot-conformance-oracle.test.ts)
+extends the framework owner to synchronous reads before effects run (maintainer
+decision 2026-11-01, stated in the shared conformance contract). Generated
+histories cross callback, params, config, and precreated Collection inputs;
+list and `findOne` results; opt-in pooling through params; populated, empty,
+and disabled selections; and source writes before attachment and while mounted.
+An independent fixture-row filter judges selected rows. A refinement relation
+also checks keyed state, data shape, Collection rows, status, readiness, and
+health meaning at each read. Initial reads, input replacements before effects,
+and settled source updates use the same checker. The original copied-signal
+adapter fails at the first read; omitting post-subscribe invalidation fails
+after a source write between an early read and attachment. Declared reach
+measures those checkpoints, including direct Collection, single-result, and
+pooled receiving paths. The oracle does not compare pooled view health-node
+identities or Collection IDs with their lazily materialized public Collection:
+`getPublicCollection` represents these as different objects; it compares health
+meaning instead. This does not retire the shared suite's within-query health
+identity checks. Limits: synchronous eager sources in Angular's JIT TestBed;
+no browser rendering, SSR, asynchronous provider schedules, or new synchronous
+read promise for another framework. Existing shared conformance retains
+ownership of their settled behavior and teardown laws.
+
 The query identity owner generates both equal and unequal primitive string
 pairs. Its fixed pairs also check public `groupBy` rows after initial
 publication. End-to-end subset-demand reuse is still unproven: the demand-key
@@ -7473,3 +7495,20 @@ requests while joined work defers a reentrant pass, then check the real
 subscription receiving path. This remains separate from PR #2346's observer
 metadata-sharing and discarded-publication scope. The test, grammar and run
 budget remain unchanged; a green later random campaign does not close it.
+
+## GC direct-start reach, 2026-11-04
+
+The existing `collection-lifecycle.gc-retention-boundaries` owner now begins
+each normal campaign with ten generated direct-start histories, then runs its
+unchanged 250 mixed histories. This removes the known sampling gap where no
+history starts sync through direct insertion. The model, comparisons, all
+declared reach states, and both mixed run budgets remain intact. Explicit path
+replays of the mixed property retain its original generator and skip the prefix.
+The separately registered
+`collection-lifecycle.gc-retention-boundaries.direct-start` property replays
+only the prefix, with its own reach recorder. Both identities use the same
+model, driver, and comparisons. Seed 706
+reproduces the old gap and rejects omission of the prefix; the repaired fixed
+and random campaigns reached the state 15 and 17 times. Existing controlled
+idle-host and native-browser limits remain in force. See the
+[CI repair evidence](oracle-reviews/2026-11-04-gc-startup-reach.md).

docs/contributing/oracle-reviews/2026-11-01-angular-snapshot.md

@@ -0,0 +1,144 @@
+# Angular synchronous snapshot evidence
+
+This self-review covers the Angular changes on `angular-fixes`, based on
+`95dbf663877cc0040ae8a502d34ee2448391379c`. The branch also includes the
+effect-cleanup refactor and breaking Angular 20 minimum. Tested source SHA-256:
+`94a4dddccbb3048fee3c0660319951159a530cdb899188b3519b3a467f1e0f70`.
+Test SHA-256:
+`730c8aa77c20435c7cf7fadff011e027c942bbbf1431e24538e63c02ee93266d`.
+
+- **Law:** synchronous Angular reads describe the current query, including
+  before subscription effects run. Rows, keyed state, Collection, status, and
+  health must agree. The maintainer authorized this change on 2026-11-01;
+  `packages/db/tests/conformance/contract.ts` records its Angular-only cut.
+- **Reach:** fixed and random campaigns each ran 36 histories: 16 shared
+  regressions plus 20 generated histories. In fixed/random order, initial reads
+  and pre-attachment source writes reached 36/36 histories each; direct
+  Collections 9/7; findOne 18/17; pooled reads 2/2; enabled replacements 22/26;
+  disabled reads 15/16; re-enabling 12/13; settled updates 36/36.
+- **RED:** from `packages/angular-db`, run
+  `../../node_modules/.bin/vitest --run tests/snapshot-conformance-oracle.test.ts`
+  with the adapter from the base commit. Both campaigns fail by row assertion
+  before attachment: `data()` is empty while `collection()` contains rows.
+  Fixed seed `1101`, shrink path `0:0:0:0` reduces to callback/list/unpooled,
+  initial value 0, selections `[0, 0]`.
+  Direct replay with `TANSTACK_DB_ORACLE_PROPERTY=angular.current-snapshot`,
+  `TANSTACK_DB_ORACLE_SEED=1101`, and `TANSTACK_DB_ORACLE_PATH=0:0:0:0`
+  executes only that history and reproduces the row mismatch on the base adapter.
+- **GREEN:** the same command passes with the snapshot adapter. The whole
+  package passes 97 tests with one existing TODO. `tsc --noEmit -p
+  packages/angular-db/tsconfig.json`, changed-file lint, ESLint for changed
+  TypeScript, and formatting checks pass.
+- **Mutant:** deleting the post-subscribe `invalidate()` preserves the early
+  cached value after a source write before attachment. Both campaigns fail the
+  row comparison after effect setup; this is an assertion failure after the
+  driver's bounded wait, not the test runner's timeout. Seed `1101`, path
+  `0:0:2:1:0` reduces to the same two-selection history.
+- **Limits:** three synchronous eager fixture rows, finite equality groups,
+  one ordered query, no joins/includes, and Angular JIT TestBed. No SSR,
+  browser-rendering, asynchronous-provider, or cross-framework synchronous
+  claim. Existing shared conformance remains the owner for settled behavior,
+  errors, required inputs, and teardown. Pooled view and materialized public
+  Collection IDs/health-node identities differ by design; the new oracle
+  compares health meaning across them, preserving existing identity checks.
+
+The size comparison against the base commit reports 1023 to 881 gzip bytes
+(-142 bytes), and -134 brotli bytes for the Angular package. This is the
+reactivity patch's production bundle, excluding dependency bundles and tests.
+
+## Oracle rubric assessment
+
+| Requirement | Evidence or limit |
+| --- | --- |
+| ORC-001 | Shared contract, maintainer authority, and limits above and beside the oracle. |
+| ORC-002 | Fixture-row filtering supplies membership; public Collection observations supply the snapshot-coherence relation. No adapter state or query evaluator computes expected membership. |
+| ORC-003 | Opening contract, pure row model, legal histories, real Angular driver, and shared comparison remain distinguishable in the test. |
+| ORC-004 | Regressions reconstruct populated swap, disable, empty re-enable, and return. Removing forms, shape, pooling, or the pre-attachment write loses a named path. Groups are 0/1/empty-2/null; initial values are -2 through 2; histories have 2 through 6 selections. Updates after destruction are outside the legal grammar. |
+| ORC-005 | Public injectLiveQuery and real environment-injector effects; reads before effects, after attachment, and after mounted writes. Source rows already exist at the synchronous cut. |
+| ORC-006 | Original copied-signal adapter and missing-invalidation mutant fail row assertions. |
+| ORC-007 | Fixed seed 1101 and unseeded campaigns share options, regressions, grammar, checker, and budget. Shared oracle replay variables select the property and shrink path. |
+| ORC-008 | No adapter state machine is mirrored: membership is stateless filtering of fixture rows. |
+| ORC-009 | Uses glossary terms. Health meaning explicitly omits identity across pooled view/materialization, not within one query. |
+| ORC-010 | ScenarioLifetime preserves primary and cleanup failures and ends the injector before source cleanup. |
+| ORC-011 | No named shared semantic fault requiring another formulation; numeric equality on three fixture rows is independent of adapter lifecycle code. |
+| ORC-012 | This record describes the tested patch and bounded evidence, not universal correctness. |
+| ORC-013 | Populated versus empty/disabled selections and a write between read/attach distinguish stale-data and missed-invalidation designs. |
+| ORC-014 | The receiving boundary is Angular's actual JIT scheduler. External-provider and browser claims remain outside scope. |
+| ORC-015 | createOracleReach checks all nine declared receiving states; counts above. |
+| ORC-016 | Fixed regressions and both campaigns share the same model, driver, and comparison; no separate expected answers. |
+
+## Stability, work, and the separate observer PR
+
+The diagnostic `snapshot-performance.test.ts` instruments real Angular computed
+executions and core snapshot reads without replacing snapshot behavior. It
+tracks container identities across warmup and measurement phases, so reading an
+existing container through another snapshot is not counted as a new allocation.
+For 1000 synchronous rows, unchanged reads reuse the same core snapshot, data,
+and Map and execute no additional Angular computations. Ten row writes before
+reading build one full snapshot; reading after every write builds ten.
+
+This branch includes an intermediate `rowState` computed. Angular's defensive
+Map copy follows the observer's row-state reference rather than every full
+snapshot. Against the current main-branch observer, metadata snapshots still
+supply new row containers, so health reports still rerun row consumers. Keeping
+those containers stable belongs to the independent `observer-snapshot-caching`
+branch. This PR does not require that change or promise metadata-only identity
+that the current core cannot supply. Both branches together were exercised by
+the generated Angular metadata receiving checks before the split; those
+integration results do not stand in for either branch's standalone evidence.
+
+The defensive copy remains because the public Angular state type is mutable;
+this PR does not expose the observer's cached Map as writable state. Scalar
+field computations may reevaluate after a notification while unchanged
+readiness/Collection consumers and the readiness effect skip execution. These
+are finite work counts, not heap bytes, elapsed runtime, or browser rendering.
+
+Validation uses Node 24.8.0 from `.nvmrc` and pnpm 11.1.0 from `package.json`.
+The changeset declares a major Angular package change because raising the
+minimum supported framework version is breaking. No observer production code,
+observer cache changeset, or core metadata-identity requirement belongs to this
+branch.
+
+## CI replay registration follow-up, 2026-11-04
+
+CI exposed a missing entry in `oracle-replay-manifest.ts` for the registered
+`angular.current-snapshot` property. The oracle itself passed in the framework
+job, but the central replay-owner check correctly rejected an unowned name.
+The fix maps it to `angular-db/tests/snapshot-conformance-oracle.test.ts`.
+
+Before the fix, `vitest run tests/oracle-replay.test.ts -t 'names an owner for
+every registered property' --coverage.enabled=false` failed with
+`['angular.current-snapshot']` instead of an empty unowned-property list. After
+the fix, the complete replay suite passed all 41 tests. The guarded replay from
+`packages/angular-db` also passed with property `angular.current-snapshot`, seed
+1101 and path `0`: one completed matching witness and one execution, with the
+fixed campaign skipped. The replay guard observed the actual Angular owner,
+rather than merely accepting a manifest entry. The existing oracle model,
+grammar, assertions and run budgets did not change.
+
+Node 24.8.0 was used. This follow-up changes replay discovery only and adds no
+new Angular behavior or observation checkpoint.
+
+## Main integration, 2026-11-04
+
+The merge combines Angular branch `c098d577d241042bc247ac41a770b85e50c0b48f`
+with main `8feab6ea3eaae4456f738da39fa93508ddd5cc01`. Main added ordered-window
+reuse and shared result-shape/window-history receiving tests after the original
+PR base. The resolution retains main's `resolveLiveQueryValue` window ownership
+and deferred move effect, and the PR's synchronous observer snapshot reads and
+effect-owned subscription cleanup. The conformance-driver comment now names
+both its pre-effect shape reads and its effect cleanup accurately.
+
+All replay property/owner entries and both sets of coverage-map additions are
+retained. No oracle model, assertion, history grammar, run budget, or expected
+result was changed to resolve these conflicts. The merge commit first adding
+this section binds the evaluated combined source.
+
+Validation on Node 24.8.0: the full Angular suite passed 129 tests with one
+existing TODO, including window history, result shape, and the PR's current-query
+snapshot campaigns. Angular typecheck, build, focused ESLint and `lint:changed`
+passed. The replay manifest's owner, file mapping and command-location checks
+passed (three selected tests), and guarded `angular.current-snapshot` replay
+1101/path `0` executed one matching witness. The GC seed-706 campaign passed and
+reached `write-started-sync` ten times. Existing SSR/browser and controlled-host
+limits remain unchanged; this resolution adds no client/SSR provider code.