Adaptor packages: Flink and time-series databases/services

[microprediction]

Expose skaters (laplace) as streaming operators/UDFs so the one-step predictive distribution is available where the data already lives.

Targets (roughly in order): Apache Flink; Kafka Streams; InfluxDB tasks; TimescaleDB/Postgres; QuestDB; ClickHouse; kdb; Prometheus/VictoriaMetrics; AWS Timestream; Azure Data Explorer.

Shape of each adaptor

• one-step predictive Dist per tick, surfaced as mean / interval / logpdf columns (anomaly z from the parade where useful)
• skater state is pure data by design, so operator checkpoint/restore is free
• zero-dependency core stays zero-dependency; each adaptor is its own thin package