src/core/consumer.ts — XREADGROUP loop with consumer-group resumption,
ensureConsumerGroup (BUSYGROUP-tolerant), decodeBatch (CodecError → log
+ skip + leave pending; never speculative ACK), partial-ACK semantics,
connectRedis (mirroring tcp-ingestion's retry pattern), clean stop.
src/core/state.ts — LRU Map<device_id, DeviceState> using delete+set
bump trick (no third-party LRU dep); last_seen = max(prev, ts) so
out-of-order replays don't regress the high-water mark; evictedTotal()
counter.
src/core/writer.ts — multi-row INSERT ON CONFLICT (device_id, ts) DO
NOTHING with RETURNING. Duplicate detection by set-difference between
input and RETURNING rows (xmax=0 doesn't work for skipped-conflict
rows, only returned ones — confirmed in the task spec's own Note).
Sequential chunking to WRITE_BATCH_SIZE; bigint→string and Buffer→base64
attribute serialization that handles Buffer.toJSON shape.
src/main.ts — full pipeline: pool → migrate → redis → state → writer →
sink → consumer → graceful-shutdown stub. Sink ordering is
state.update BEFORE writer.write per spec rationale (state stays
consistent with what's been seen even if not yet persisted; redelivery
is idempotent on state). Metrics is still the trace-logging shim from
tcp-ingestion's pre-1.10 pattern; real prom-client lands in task 1.9.
Verification: typecheck, lint clean; 112 unit tests passing across 7
test files (+39 from this batch).
