Authored by Neo Claude Opus 4.7 (Claude Code). Session c2912891-b459-4a03-b2af-154d5e264df1.

Resolves #10946

Restores symmetric polluter-side cleanup in test/playwright/unit/ai/daemons/DreamService.spec.mjs so the DreamServiceGoldenPath sibling spec receives a clean singleton state instead of two leaked mutations from the synthesizeGoldenPath mock test + the afterAll clear-strategy _initPromise retention.

Evidence: L1 (static spec audit + empirical pair-test + full-bucket × 5 isolation runs at --retries 0) → L1 required (test-isolation regression with deterministic empirical reproduction). No residuals.

What changed

Two polluter-side cleanup gaps in DreamService.spec.mjs, both surfaced via the #10946 bisection protocol (Step 1: isolated spec 5/5 PASS → Step 2: full ai/daemons/ bucket 1 FAIL → Step 3: minimal pair-test reproduction → Step 4: code-read).

Gap 1 — monkey-patch asymmetry (lines 859, 911-921): The synthesizeGoldenPath should mathematically select and inject Golden Path test monkey-patches GraphService.getContextFrontier = () => ({ nodes: [], edges: [] }) but its companion restore block at the end of the test does NOT include a matching restore. The sibling patches (originalLinkNodes, originalGenerate, originalAppendFile, originalPrepare) are all captured + restored symmetrically; getContextFrontier was the asymmetry. Subsequent specs (notably DreamServiceGoldenPath.spec) then receive { nodes: [], edges: [] } from the leaked patch instead of the real method's topology → topology.frontier becomes undefined → assertion at DreamServiceGoldenPath.spec.mjs:86 fires with Cannot read properties of undefined (reading 'id') (matches the original ticket body's failure-mode anchor).

Fix: capture originalGetContextFrontier symmetrically to originalLinkNodes, restore in cleanup block.

Gap 2 — afterAll clear-vs-destroy strategy (lines 152-153): afterAll uses cleanupGraphService(..., 'clear') which only clears nodes/edges/vicinityLoadedNodes/storage but does NOT reset SystemLifecycleService._initPromise. Sibling specs that check if (!_initPromise) initAsync() else ready() then take the else ready() branch — never honoring their own aiConfig.storagePaths.graph. Real getContextFrontier() then returns null (no frontier node ever upserted in the wrong-DB state).

Fix: switch to cleanupGraphService(..., 'destroy') — already-supported by TestLifecycleHelper — which resets _initPromise, destroys db, and deletes the SQLite file + WAL + SHM. Sibling specs then take the initAsync() branch and run with their own config.

Both fixes are polluter-side per feedback_symmetric_spec_cleanup.md discipline. No consumer-side changes; no skip-guards added; no spec-level settle-patterns introduced.

V-B-A trail (substrate hygiene)

The original ticket body framed the race-surface as Promise-unsettlement between synthesizeGoldenPath() and getContextFrontier(). V-B-A revealed all relevant methods (linkNodes, upsertNode, getContextFrontier, Database.getAdjacentNodes) are SYNCHRONOUS (better-sqlite3 + in-memory collections). The race is NOT async-Promise-class but workers:1 sibling-spec singleton-pollution — captured in prior ticket comment (V-B-A 2026-05-10). This PR concludes the investigation lane with the empirically-traced fix.

Test Evidence

Empirical reproduction + fix verification, all with --retries 0 (overriding CI=true's default retries=2 to detect first-attempt-only flake):

• Step 1 baseline — isolated DreamServiceGoldenPath.spec.mjs × 5 runs: 5/5 PASS (4-6s each) — confirms sibling-pollution hypothesis
• Step 2 bisection — full test/playwright/unit/ai/daemons/ bucket: 97/98 PASS, 1 FAIL (the deterministic flake) — confirms within-bucket pollution
• Step 3 minimal-pair pre-fix — DreamService.spec → DreamServiceGoldenPath.spec: 14/15 PASS, 1 FAIL — minimal reproduction
• Step 4 minimal-pair post-fix — DreamService.spec → DreamServiceGoldenPath.spec: 15/15 PASS — fix verified
• Step 5 AC4 verification — full test/playwright/unit/ai/daemons/ bucket × 5 iterations post-fix: 98/98 PASS each run, 5/5 runs PASS — deterministic stability achieved

Commands used:

CI=true npm run test-unit -- --retries 0 test/playwright/unit/ai/daemons/DreamService.spec.mjs test/playwright/unit/ai/daemons/DreamServiceGoldenPath.spec.mjs
CI=true npm run test-unit -- --retries 0 test/playwright/unit/ai/daemons/

Acceptance Criteria Closure

• AC1 (empirical reproduction) — Step 2 above: full bucket reproduces 1/98 deterministic fail at --retries 0
• AC2 (identify race surface) — getContextFrontier monkey-patch leak + _initPromise retention in afterAll. Both empirically confirmed via minimal-pair bisection.
• AC3 (choose fix path + rationale) — Polluter-side over consumer-side (matches feedback_symmetric_spec_cleanup.md symmetric cleanup discipline; doesn't add defensive code to every consumer)
• AC4 (5x deterministic runs) — Step 5 above: 5/5 runs PASS, 98/98 tests per run
• AC5 (skip-guard removal) — None to remove. The test.skip(!!process.env.NEO_TEST_SKIP_CI, ...) guard at DreamServiceGoldenPath.spec.mjs:78 was already removed in PR #10940; the comment text was left as historical context. Could optionally clean up the stale skip-comment in a follow-up; non-blocking.

Commits

• 72f4f15c9 — fix(test): restore singleton state in DreamService.spec afterAll + getContextFrontier patch (#10946)