Outbox processor — asyncio batch engine with backpressure

[slusset]

Parent Epic

Part of #16 — [Epic] Databricks EHR Writeback — Delta Outbox + Canonical Prediction Model

Summary

Build the async batch processor that polls the Delta outbox, dispatches predictions to EHR endpoints with bounded concurrency, and commits results back.

Dependencies

• Depends on: Canonical ClinicalPrediction model + outbox schema (sub-issue 1 of [Epic] Databricks EHR Writeback — Delta Outbox + Canonical Prediction Model #16)

Tasks

• Asyncio dispatcher with asyncio.Semaphore for configurable concurrency (default 10-20)
• Batch claim → process → commit loop: claim N records, dispatch concurrently, MERGE results back
• Retry logic: increment retry_count on failure, route to dead letter after configurable max retries
• Graceful shutdown: finish in-flight requests before exiting
• Integration with existing dead_letter.py for dead letter routing
• Configurable batch size, concurrency, retry limits, lock timeout
• New module: src/ehr_writeback/outbox/processor.py

Acceptance Criteria

• Processor respects concurrency semaphore (never exceeds N concurrent calls)
• Failed records are retried up to max, then dead-lettered
• Batch results are committed atomically via Delta MERGE
• Stale locks from crashed processors are reclaimed on next batch
• Unit tests with mock EHR adapter