feat(0.12.0): production trace-sink — close the data-leak

[drewstone]

Summary

Every production chat session has been emitting zero replayable trace data. Eval runs capture everything; production captures nothing. RL training, research analyses, and the self-improvement loop all run on synthetic personas only. This primitive turns every real user conversation into data the downstream channels (Prime Intellect, GEPA, research, canaries, analyst loop) can consume.

API

`createProductionTraceSink(opts)` returns:

• `traceStore` — the in-memory store the agent's `TraceEmitter` writes to during a chat session (built on agent-eval's existing `InMemoryTraceStore`; no reinvention)
• `onRunComplete` — `RunCompleteHook` the agent registers; on `endRun` composes a canonical `ProductionRunRecord`, persists to a durable store, and POSTs the run as OTLP to a configured collector (Langfuse, etc.)
• `recordFeedback(input)` — appends a `FeedbackLabel` to the run's `FeedbackTrajectory`; creates the trajectory anchored to `runId` on first feedback

Per-agent wiring (~10 lines)

```ts
const sink = createProductionTraceSink({
projectId: 'tax-agent',
otlp: { endpoint: env.LANGFUSE_OTEL_ENDPOINT, authHeader: env.LANGFUSE_OTEL_AUTH },
runRecordStore: drizzleRunRecordStore(db),
feedbackStore: drizzleFeedbackStore(db),
})

const emitter = new TraceEmitter(sink.traceStore, {
onRunComplete: [sink.onRunComplete],
})
await emitter.startRun({ scenarioId: sessionId, projectId: 'tax-agent', layer: 'app-runtime' })
// ... existing chat flow ...
await emitter.endRun({ pass, score })
```

CF Worker semantics: `ctx.waitUntil` the hook from the chat handler.

Fail-loud where it matters

• runRecordStore failures → logged, not thrown
• OTLP POST failures (network / non-2xx) → logged, not thrown
• feedbackStore failures → `null` returned, logged

Test plan

• `pnpm test` — 144/144 pass (13 new under `tests/production-trace-sink.test.ts`)
• `pnpm typecheck`
• Bumps to 0.12.0

Next steps (separate PRs per agent)

Each of tax/legal/gtm/creative wires this into `packages/api-worker/src/services/agent-runtime/chat.ts` + adds the matching `runRecordStore`/`feedbackStore` Drizzle adapters + Langfuse env vars.