End-to-end test and validate memory system

[geier]

Context

After deploying Graphiti (#240), building the watcher (#241), configuring MCP (#242), and optionally setting up the Kubeflow service (#243) and custom entity types (#244), the full system needs end-to-end validation to ensure everything works together reliably.

Goal

Validate the complete memory pipeline: session → watcher → Graphiti ingestion → MCP retrieval, across the key scenarios.

Test Scenarios

1. Basic Ingestion Flow

• Start a new OpenCode session
• Have a conversation that includes: an architectural decision, a code preference, an error + fix
• End the session
• Verify the watcher detects session completion and ingests the transcript
• Verify Graphiti extracts entities and facts from the transcript
• Query Graphiti to confirm knowledge is searchable

2. Cross-Session Retrieval

• Start a new session in the same project
• Use search_memory_facts to recall the decision from the previous session
• Use search_nodes to find the error solution
• Verify results are relevant and correctly attributed

3. Multi-Project Scoping

• Run sessions in two different project directories
• Verify each project's knowledge is scoped to its group_id
• Verify cross-project queries work when specifying multiple group_ids

4. Multi-Machine Simulation

• Run the watcher from two different directories pointing to the same Graphiti instance
• Verify both sessions are ingested without conflicts
• Verify knowledge from both is searchable

5. Temporal Knowledge

• In session 1: "We're using React for the frontend"
• In session 2 (later): "We switched from React to SolidJS"
• Verify Graphiti correctly invalidates the old fact and tracks the change
• Verify search_memory_facts returns the current state (SolidJS), not the old one

6. Watcher Resilience

• Kill the watcher while a session is active
• Restart the watcher
• Verify it catches up on the missed session
• Verify no duplicate ingestion occurs

7. Deduplication

• Complete a session, verify it's ingested
• Restart the watcher, verify it does NOT re-ingest the same session
• Add a new message to the session, verify the delta is ingested

Acceptance Criteria

• All 7 test scenarios pass
• Memory retrieval returns relevant results (not random/unrelated)
• Temporal facts are correctly tracked and invalidated
• No duplicate ingestion on watcher restart
• No data loss on watcher crash/restart
• System works across multiple projects and (simulated) machines
• Document any issues found and file follow-up issues

[geier]

Marking as blocked for now: end-to-end validation depends on prerequisite implementation issues (#240, #241, #242, and optionally #243/#244).