Update Processor Service documentation: complete Phase 1 and detail architecture, metrics, and integration tests

.claude/agent-memory/ts-node-backend-engineer/project_processor.md

@@ -1,23 +1,52 @@
 ---
 name: Processor Service
-description: processor service scaffold, conventions, and known test patterns learned during Phase 1 tasks 1.1–1.4
+description: processor service: Phase 1 complete (all 11 tasks), key patterns, conventions, and quirks
 type: project
 ---
 
-Phase 1 tasks 1.1–1.4 landed. Key facts for tasks 1.5–1.11:
+Phase 1 complete. All 11 tasks landed. The throughput pipeline is done: consumer + writer + metrics + integration test + Docker + CI.
 
 **Architecture divergence from tcp-ingestion:**
-- ESLint `import/no-restricted-paths` zone: `src/core/` cannot import `src/domain/` (not `src/adapters/` like tcp-ingestion). `src/domain/` is preemptive for Phase 2.
-- `INSTANCE_ID` defaults to `'processor-1'` (fixed string), not a random UUID prefix. Rationale: deterministic names are better for container environments.
-- `REDIS_CONSUMER_NAME` defaults to `INSTANCE_ID` at runtime (after schema parse), not as a zod schema default — because it depends on `INSTANCE_ID` value.
+- ESLint `import/no-restricted-paths` zone: `src/core/` cannot import `src/domain/` (preemptive for Phase 2).
+- `INSTANCE_ID` defaults to `'processor-1'` (fixed string, not random UUID). Deterministic for container environments.
+- `REDIS_CONSUMER_NAME` defaults to `INSTANCE_ID` at runtime (after schema parse), not as a zod schema default.
+- `connectRedis` lives in `src/core/consumer.ts` (not `src/db/`). Rationale: only the consumer needs a Redis connection in Phase 1.
+
+**Buffer.toJSON() gotcha in JSON.stringify replacers:**
+- `Buffer.isBuffer(value)` is false inside a `JSON.stringify` replacer for nested Buffer properties because `Buffer.prototype.toJSON()` fires before the replacer.
+- Must handle both `instanceof Uint8Array` (direct reference) AND the `{ type: 'Buffer', data: number[] }` shape. See `serializeAttributes` in `src/core/writer.ts`.
+
+**writer.ts — RETURNING duplicate detection:**
+- `ON CONFLICT DO NOTHING` does not return conflicting rows. Duplicates detected via set-diff on (device_id, ts) between input and RETURNING rows.
+
+**LRU Map trick (state.ts):**
+- Plain `Map` with `delete` + `set` on every update. `keys().next().value` is always the oldest. O(1), no external library.
+
+**Metrics (task 1.9):**
+- `src/observability/metrics.ts` exports: `createMetrics()`, `startMetricsServer()`, `createPostgresHealthCheck()`, `createConsumerLagSampler()`.
+- `createPostgresHealthCheck`: cached `SELECT 1`, 5 s TTL, background interval, sync `isReady()` accessor for `/readyz`.
+- `createConsumerLagSampler`: uses `XINFO GROUPS` → `lag` field (Redis 7.2+); falls back to `XLEN` if absent.
+- `processor_device_state_size` is a gauge updated via `metrics.observe()` in the sink (not in state.ts — avoids coupling state to metrics).
+- prom-client counters with label dims do NOT emit `{label} 0` at init. Test accordingly (check label-less counters for zero baseline).
+
+**Integration test (task 1.10):**
+- `test/pipeline.integration.test.ts`: 4 scenarios. Skip-on-no-Docker pattern: try container start in `beforeAll`, set `dockerAvailable = false` on error, each `it` early-returns.
+- TimescaleDB image: `timescale/timescaledb:latest-pg16` (not stock postgres).
+- Consumer uses a separate `connectRedis` connection; test XADD uses a separate `redisClient`. Mirrors production topology.
+- Test 4 (retry): stops pgContainer then restarts it. `pgContainer.restart()` returns a new `StartedTestContainer`. Must reassign `pgContainer` for the `afterAll` cleanup to reference the right instance.
 
 **Test patterns:**
-- `connectWithRetry` tests: use `maxAttempts=1` for the "calls process.exit(1)" case to avoid fake-timer unhandled rejection noise. The noise occurs when a backoff `setTimeout` resolves after the outer promise has already thrown via process.exit mock.
-- Migration runner tests: `isMigrationApplied` calls `pool.query` (pool-level, not through client). Handlers on client mock do NOT affect EXISTS checks.
-- `vi.mock('node:fs/promises')` pattern used in migrate.test.ts to isolate filesystem reads.
+- `connectWithRetry` tests: use `maxAttempts=1` for process.exit test to avoid fake-timer noise.
+- Consumer tests: sink calls `void consumerRef?.stop()` to exit the loop.
+- Inline `import()` type annotations forbidden by ESLint (`@typescript-eslint/consistent-type-imports`). Always use top-level `import type`.
+
+**Dockerfile:**
+- `EXPOSE 9090` only (no TCP listener — Processor has none).
+- `HEALTHCHECK` uses `${METRICS_PORT:-9090}` default.
+- Label `org.opencontainers.image.source` not included (not in tcp-ingestion original either).
 
 **Why:**
-Phase 1 throughput pipeline. Tasks 1.5–1.8 add consumer/writer/state on top of this foundation.
+Phase 1 is the throughput pipeline + operational baseline. Phase 2 (domain logic) and Phase 3 (hardening) build on top.
 
 **How to apply:**
-When implementing 1.5 (Redis consumer), 1.6 (device state), 1.7 (position writer), 1.8 (main wiring) — all imports must use `.js` extension, no path aliases, strict ESM. The `connectWithRetry` pattern from pool.ts should be mirrored for ioredis in consumer.ts.
+Phase 2 adds domain logic to `src/domain/` — no changes to `src/core/`. Phase 3 adds graceful shutdown polish, XAUTOCLAIM, and state rehydration.