Summary

• New merge::streaming::streaming_merge_sorted_parquet_files: async N-input → M-output merge consuming inputs as Box<dyn ColumnPageStream> and writing each output column-by-column via StreamingParquetWriter (single row group per output).
• Inputs drained per-RG via PR-6a's StreamDecoder, then concatenated per input. Per-RG decode memory is bounded; per-input concat matches the existing engine's input shape for the merge planner.
• Reuses the existing permutation / KV metadata / sorting columns / MC-3 sort-order helpers (super::writer) and union schema / output-optimisation helpers (super::schema) — now pub(super). PR-7 folds the non-streaming path away.
• Multi-RG output at metric_name boundaries lands in PR-6c.

Why the win is on the output side

The standard ArrowWriter materialises a full row group worth of column-chunk buffers before serialising. PR-2's StreamingParquetWriter flushes one column chunk at a time, so output peak memory is bounded by the largest single column chunk plus bookkeeping (page index, bloom filters), not by the row group total.

The input side stays at "drain whole input into a RecordBatch per input" because per-RG streaming through the merge driver requires synchronising RG boundaries across inputs, which only works once prefix=1 multi-RG inputs are the dominant compaction case. PR-6b is correct without it; the input-side streaming optimisation is a follow-up.

Stack

main
└── gtt/parquet-streaming-base (= main ∪ PR-2 #6384 ∪ PR-4 #6386)
    └── gtt/column-page-stream-trait (PR-5a #6406)
        ├── gtt/legacy-input-adapter (PR-5  #6408)
        └── gtt/parquet-page-decoder  (PR-6a #6407)
            └── gtt/streaming-merge-engine-merger ← PR-6b (this PR)
                └── gtt/streaming-merge-engine-multi-rg (PR-6c — next)

Test plan

• test_two_inputs_simple_merge — total row count + ascending sorted_series
• test_output_is_single_row_group — PR-6b's contract: every output is exactly 1 RG
• test_total_rows_preserved — MC-1 invariant
• test_sort_schema_mismatch_rejected — input metadata validation
• test_window_start_mismatch_rejected — same
• test_output_has_page_index_metadata — column index + offset index present (query pruning)
• test_kv_metadata_propagated_to_output — qh.sort_fields, qh.window_start, qh.window_duration carry through; qh.num_merge_ops incremented
• test_all_empty_inputs_no_output
• test_one_empty_input_among_nonempty
• test_output_drainable_by_stream_decoder — round-trip output → StreamDecoder → row count

CI gates locally green: clippy --workspace --all-features --tests with -Dwarnings, nightly fmt --check, cargo doc --no-deps, cargo machete, license headers, log format, typos. 411/411 crate tests pass.

🤖 Generated with Claude Code