feat: streaming merger for lattice-compatible distributed aggregation

[cigrainger]

Summary

• Dux.Remote.StreamingMerger — folds worker IPC results incrementally using lattice merge operations instead of batching all results for a single SQL merge query
• Coordinator auto-selection — checks streaming_compatible? from PipelineSplitter; uses streaming for SUM/COUNT/MIN/MAX pipelines, falls back to batch for non-lattice aggregates (MEDIAN, etc.)
• Telemetry — emits [:dux, :distributed, :streaming_merge] events after each worker result is folded, with progress metadata (workers_complete, workers_total)
• Worker failure tracking — record_failure/1 counts failed workers without crashing the merge
• Unordered fan-out — streaming path uses ordered: false in Task.async_stream for lower latency

Benchmarks

Metric	Value
Streaming MIN+MAX (2 workers)	3.29ms
Streaming SUM+COUNT grouped (2 workers)	9.41ms
Distributed filter+agg (2 workers, streaming)	4.09ms

Test plan

• Happy path: fold SUM, fold MIN/MAX, fold grouped, progress tracking, failure recording
• Sad path: empty IPC, no folds → empty Dux, all workers fail, mixed lattice+non-lattice, filter-only pipeline
• Adversarial: 100 groups, repeated folding (5x), nil group keys, AVG rewrite end-to-end
• Wicked: streaming vs batch idempotence for MIN/MAX, many groups × many workers, fold order commutativity
• Property tests: SUM commutativity (3 partitions, any order), MIN/MAX commutativity
• Integration: SUM+COUNT, MIN+MAX, grouped SUM, 3-worker SUM, penguins dataset
• Peer tests: 5 cross-node streaming tests (SUM+COUNT, MIN+MAX, grouped, AVG rewrite across 3 peers, penguins)
• Telemetry: streaming_merge events emitted per worker
• 568 total tests (38 properties, 487 tests, 43 doctests), 0 failures, 0 credo issues

🤖 Generated with Claude Code