G5#3: PermissionService.spec AGENT:* singleton cross-pollution under workers:1

Context

Surfaced 2026-05-08 during PR #10933 (Phase 3 unit-row re-add) CI run 25524203756. Originally classified as G5#3 in the #10924 G5 triage matrix. Triage hypothesis was "auto-resolve post-G4-merge"; falsified empirically by the unit-row first activation in CI.

Skip-guarded in PR #10933 commit 8e6c3fd78 per the established pattern; this ticket tracks the proper investigation + fix.

The Problem

Test test/playwright/unit/ai/mcp/server/memory-core/services/PermissionService.spec.mjs:165 flakes with:

expect(node.type).toBe('BroadcastSentinel');
  Expected: "BroadcastSentinel"
  Received: undefined

The test pre-seeds AGENT:* as a BroadcastSentinel via GraphService.upsertNode, then asserts the type survives grantPermission. Empirically node.type is undefined — meaning either:

• A sibling spec re-upserted AGENT:* with a different shape (no type field), OR
• The singleton GraphService was re-initialized between the seed and the read, dropping the in-memory state

workers:1 substrate amplifies cross-pollution. Per grep, 4 specs touch AGENT:*:

• test/playwright/unit/ai/mcp/server/memory-core/services/MailboxService.spec.mjs
• test/playwright/unit/ai/mcp/server/memory-core/services/WriteSideInvariant.spec.mjs
• test/playwright/unit/ai/mcp/server/memory-core/services/GraphService.spec.mjs
• test/playwright/unit/ai/mcp/server/memory-core/services/PermissionService.spec.mjs (this one)

Whichever sibling runs first and writes AGENT:* without a type field clobbers the BroadcastSentinel shape.

The Architectural Reality

• test/playwright/unit/ai/mcp/server/memory-core/services/PermissionService.spec.mjs:165-181 — the flaky test
• ai/mcp/server/memory-core/services/GraphService.mjs#upsertNode — last-writer-wins on the id key; updates fields per the upsert payload
• 4 spec-files all write AGENT:* to the same singleton GraphService
• workers:1 serializes execution → cumulative state pollution

The Fix (TBD via investigation)

Two candidate paths:

1. Spec-level: each spec uses a UNIQUE AGENT:* analog (e.g., AGENT:*permission-test, AGENT:*mailbox-test) so writes don't collide on the same node ID
2. Symmetric beforeEach+afterEach: each spec resets AGENT:* to a known-clean shape at the start AND end of each test (per feedback_symmetric_spec_cleanup pattern from earlier session)

Investigation needed to determine which sibling clobbers first + which fix shape minimizes blast radius.

Acceptance Criteria

• (AC1) Empirically reproduce the flake locally with WORKERS=1
• (AC2) Identify which sibling spec(s) write AGENT:* without preserving type
• (AC3) Implement chosen fix path (per-spec namespaced ID OR symmetric reset)
• (AC4) Remove the NEO_TEST_SKIP_CI guard added in PR #10933 commit 8e6c3fd78
• (AC5) Verify PermissionService.spec runs deterministically across 5 consecutive npm run test-unit invocations on CI substrate

Out of Scope

• Refactoring GraphService.upsertNode to require explicit type field (architectural change beyond spec-isolation scope)
• Migrating GraphService to per-test instances

Avoided Traps

• Increasing retries: 2 → 5: papers over without addressing root cause
• Skip-guard as permanent solution: applied as immediate ship-the-PR move + tracked here for proper fix

Related

• Surfacing CI run: 25524203756
• Triage origin: #10924 G5 row (G5#3)
• Triage correction: #10924 comment 4401547656
• Sibling state-pollution patterns: G5#2 (KBRecorderService — singleton-data), #10934 (FileSystemIngestor — singleton SQLite-close), #10935 (TransportService residual)
• Substrate config: test/playwright/playwright.config.unit.mjs workers: 1 in CI
• Skip-guard commit: 8e6c3fd78 on PR #10933

Origin Session ID: 7e897a0b-33ce-4d6c-b1a9-a1ff93e4e571

Retrieval Hint: query_raw_memories(query="PermissionService AGENT:* singleton cross-pollution workers 1 flake G5#3 #10924 PR 10933")