Add regression coverage for Memory_DatabaseService#exportGraph newline emission

Context

PR #10131 landed a 2-char correctness fix in ai/mcp/server/memory-core/services/DatabaseService.mjs — Memory_DatabaseService#exportGraph was writing JSONL record separators as the literal two characters \n (backslash + n) instead of \n (newline). Every historical graph-backup-*.jsonl artifact became single-line malformed JSONL unrestorable via the current import path.

The fix is live and verified empirically post-merge: the first npm run ai:backup run produced graph-backup-2026-04-20T18-59-41.643Z.jsonl with wc -l = 369 (one line per record, matching direct-SQL graph count).

But the fix has zero CI regression coverage. If a future refactor reverts to the escaped form, no test catches it. My self-review on #10131 scored EXECUTION_QUALITY at 78 primarily because of this gap.

The Problem

#exportGraph has no Playwright coverage. Sibling specs skip it:

• DatabaseService.backupPath.spec.mjs — only memories + summaries
• backup.mjs orchestrator spec — asserts graph/ subfolder exists but not populated (GraphService.db not wired under unitTestMode)

Adding coverage requires either a minimal .db.storage.db mock (fake prepare()/iterate()) or a real better-sqlite3 in-memory DB seeded with Nodes + Edges.

The Architectural Reality

• ai/mcp/server/memory-core/services/DatabaseService.mjs — #exportGraph at lines ~139-210 (newline emission at lines 187 + 201)
• ai/mcp/server/memory-core/services/GraphService.mjs — singleton whose .db.storage.db needs stubbing
• New spec: test/playwright/unit/ai/mcp/server/memory-core/services/DatabaseService.exportGraph.spec.mjs

Pattern to mirror: DatabaseService.backupPath.spec.mjs stubs Memory_StorageRouter.get{Memory,Summary}Collection; new spec stubs Memory_GraphService.db.storage.db analogously.

The Fix

1. Add spec, stub Memory_GraphService.db.storage.db with fake prepare(sql).get() + prepare(sql).iterate()
2. Seed with ≥2 nodes + ≥2 edges
3. Call Memory_DatabaseService.manageDatabaseBackup({action:'export', include:['graph'], backupPath: tmpDir})
4. Assert:
   • wc -l equals seeded record count (newline emission)
   • Each line parses as valid JSON independently (separator, not concatenation)
   • Record content matches seed data
5. Apply test.describe.configure({mode: 'serial'}) per the sibling-spec convention
6. Sanity check: revert the fix locally, confirm spec fails

Acceptance Criteria

• New spec DatabaseService.exportGraph.spec.mjs under test/playwright/unit/ai/mcp/server/memory-core/services/
• Asserts line count + per-line JSON parseability + record content fidelity (≥2 nodes + ≥2 edges)
• test.describe.configure({mode: 'serial'}) applied
• Spec fails when \n is reverted to \\n at either emission site — validated locally as part of PR
• Spec passes under default parallel workers

Out of Scope

• Full SQLite round-trip testing (heavier setup)
• #importGraph regression coverage — separate concern
• Generic parallel-worker race fix — one-off per-spec config is the established convention

Avoided Traps

• Rejected: rely on empirical evidence. The first live ai:backup confirmed the fix today. But empirical evidence ≠ CI regression protection. If someone reverts in a refactor, no automated signal catches it. This is the rationalization I flagged in #10131's self-review.

Related

• Follow-up on #10131 — where the \n fix landed
• Sibling spec: DatabaseService.backupPath.spec.mjs
• Self-review: PR #10131 comment (2026-04-20) flagged this as the main EXECUTION_QUALITY gap

Origin Session ID: 5a521819-dc75-4549-888e-fcea818d0401