streaming-platform 0.1.0

First release.

• A compose stack, one container per node: Apache Kafka 4.3.1 (three KRaft nodes,
  RF 3, min.insync.replicas=2), Confluent Schema Registry 8.0, Flink 1.20.2 (one
  JobManager, two TaskManagers, RocksDB, checkpoints every 5 s to SeaweedFS S3), ClickHouse
  25.8, Iceberg on SeaweedFS through the Iceberg REST catalog, Prometheus and Grafana.
  make up, make demo, make down.
• The contract: ClickstreamEvent v1 in Avro under FULL_TRANSITIVE, a compatible
  v2, and three breaking variants that the registry must refuse. Producers never
  auto-register.
• streamplat, the Python tooling: a seeded generator of storefront traffic for the
  fictional retailer Lumen & Loom, with rate control, late, out-of-order, duplicate
  and four kinds of poison events and a ledger of each; the gatekeeper, which routes
  raw Avro to events.valid or events.dlq in one Kafka transaction per batch; the
  freshness probe exporter; a reconciler that counts a run at every layer against
  its ledger; and the contract check.
• Flink SQL jobs: clean (deduplicate on event_id; exactly-once to Kafka and
  Iceberg) and windows (10 s tumbling counts, 30 s-gap sessions with funnel stage,
  late events to a side output).
• ClickHouse: Kafka engine tables reading read_committed, materialized views into
  ReplacingMergeTree tables for events, windows, sessions, late events and the
  dead-letter queue.
• A freshness SLO (99% of probes queryable within 15 s) with multi-window burn-rate
  alerts, and a provisioned Grafana dashboard.
• scripts/measure.py and MEASUREMENTS.md: load levels, a broker kill, a
  TaskManager kill, a stalled job against the alert, schema rejection, a v2
  rollout, ClickHouse query latency.
• CI: ruff, mypy, pytest; compose and alert-rule validation; the contract test
  against a live registry; an end-to-end run of the full stack.