End-to-end orchestrator demo

Runs every orchestrator surface against in-memory stores and the MockLlmProvider so the full v0 flow is visible top to bottom. The DB-backed equivalent will share these interfaces and exit points.

1. 1. Launch

   ok

   launchInvestigation passes the refusal classifier, transitions draft → active, dispatches the opening message.

   {
     "investigationId": "inv-1",
     "status": "active",
     "dispatched": 1
   }

2. 2. Inbound participant message

   ok

   handleInboundMessage appends the turn, primes the debounce machine, audits interview.turn.inbound.

   {
     "turnId": "turn-cited"
   }

3. 3. Process debounced batch

   ok

   processDebouncedBatch waits for the 60s window, gates working hours + attention budget, calls the LLM, dispatches on Teams, persists the outbound turn.

   {
     "outboundActivityId": "mock-activity-2",
     "promptTokens": 76,
     "completionTokens": 32
   }

4. 4. Run synthesis

   ok

   runSynthesis transitions active → synthesis, calls LlmProvider.synthesizeFindings, persists FindingRow entries with citations.

   {
     "findingCount": 1,
     "tokens": 530
   }

5. 5. Produce report draft

   ok

   produceReportDraft transitions synthesis → report_draft.

   {
     "investigationId": "inv-1"
   }

6. 6. Finalize report

   ok

   finalizeReport runs the anonymity scan and citation scan; both must pass to transition report_draft → closed.

   {
     "publishedAt": "2026-05-11T16:06:10.000Z"
   }