- Add prometheus-client dependency
- Create core metrics infrastructure (src/eventkit/metrics.py)
  - Custom registry
  - Version info metric
  - Uptime gauge (auto-updated every 10s)
  - Component health gauges
  - Metrics HTTP server on port 9090

- Add metrics modules for all components:
  - API metrics (requests, duration)
  - Event processing metrics (received, processed, failed)
  - Storage metrics (files, bytes, operations, duration)
  - Queue metrics (enqueued, dequeued, depth)
  - Ring buffer metrics (written, published, size)
  - Warehouse loader metrics (discovered, loaded, pending)

- Instrument key components:
  - API middleware tracks all HTTP requests
  - Processor tracks event flow (received → processed → failed)
  - GCSEventStore tracks storage operations

- Config: EVENTKIT_METRICS_ENABLED and EVENTKIT_METRICS_PORT
- Tests: 238 passing (6 new metrics tests)

Design principles:
- Counter-focused (prefer counters over gauges)
- Low cardinality labels (no unbounded values)
- Naming: eventkit_{verb_noun}_{unit}_{suffix}
- Separate metrics server (port 9090) isolates monitoring traffic

Next: Instrument queue/ring buffer, add documentation

- Ring buffer metrics:
  - Track writes, publishes (success/failure), marked published
  - Track total size and unpublished count (gauges)
  - Auto-update size metrics on write/mark operations

- Ring buffer publisher:
  - Track successful/failed publish attempts

- Async queue metrics:
  - Track enqueue/dequeue operations
  - Track processing success/failure
  - Track queue depth per worker (gauge)

All 238 tests passing. Metrics are now wired throughout the pipeline:
API → Queue → RingBuffer → Processor → Storage

Metrics server runs on separate port (9090) from API (8000).
E2E test now verifies:
- Prometheus format
- API request metrics
- Event processing metrics
- Storage metrics

Document all Prometheus metrics:
- API layer (requests, latency)
- Event processing (received, processed, failed)
- Storage (bytes, files written)
- Queue/ring buffer (depth, enqueue, dequeue)
- Warehouse loader (files processed, errors)
- System (uptime, health, version)

Includes:
- Metrics server configuration
- Example Grafana queries
- Design principles
- All available metrics with labels

Track PubSub-specific operations:
- Ring buffer enqueue (before Pub/Sub publish)
- Pub/Sub publish success/failure
- Messages received from subscription
- Ack/nack with reasons (success, decode_error, processing_error, no_loop)
- Queue depth per worker

Labels match AsyncQueue for consistency:
- queue_mode="pubsub" or "pubsub_published"
- result="ack_success", "nack_processing_error", etc.

All queue modes now fully instrumented for production observability.

The ring buffer integration test relied on Firestore which was removed.
Coverage for ring buffer → queue → storage flow is now provided by:
- Unit tests for individual components
- E2e test script (scripts/e2e_test.sh) with GCS emulator