feat: per-sink delivery, journal forwarding, and operational API

Persistence:
- one delivery state per sink, with claim, crash recovery, and index replacement
- retention for delivered and dead records
- LMDB writes moved off the event loop

Pipeline:
- one attempt per dispatch, with the next try scheduled by next_retry_at
- backpressure through max_queue_size
- reload re-reads YAML; SIGTERM drains and SIGHUP reloads

Plugins:
- journald source, enricher and filter processors, Loki sink
- built-ins register when entry points are missing

API and CLI:
- health, metrics, per-sink queue, and dead-letter replay
- startup prints the pipeline on one line

Tests:
- dispatch, plugins, metrics, CLI, and persistence coverage

tests/conftest.py

@@ -34,6 +34,7 @@ def sample_record() -> StoredRecord:
         record_type="test",
         body={"message": "hello"},
         metadata={"env": "test"},
+        sinks=["out"],
     )