feat: page-stream → RecordBatch decoder (PR-6a)

[g-talbot]

Summary

• Bridges PR-4's ColumnPageStream (raw compressed pages) to arrow's standard ParquetRecordBatchReaderBuilder (decoded arrays). PR-6b's merge engine drains each input row-group through this.
• Adds header_bytes: Bytes to Page and a new StreamDecoder whose next_rg() yields one RecordBatch per input row group, idempotent at EOF.
• Per-call memory cost: one row group's reconstructed column-chunk bytes (~tens of MB); the next call drops the previous buffer.
• Pure plumbing — no merge logic.

How byte-exact reconstruction works

Each Page now carries the original Thrift-compact bytes for its header (header_bytes) alongside the parsed PageHeader and the raw compressed body (bytes). Concatenating header_bytes ++ bytes for every page in a column chunk reproduces the on-disk byte layout. The decoder allocates one buffer per RG sized to max(col_byte_range_end), places each column chunk at its original offset, and asks ParquetRecordBatchReaderBuilder::new_with_metadata (built on the original ParquetMetaData) to read just that RG via with_row_groups([rg_idx]).

We could in principle re-encode parsed PageHeader structs (Thrift-compact is deterministic for a given struct), but "deterministic in practice" is not a contract — encoder-version drift inside the compactor would silently corrupt outputs. Carrying the original header bytes sidesteps the problem.

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)
        └── gtt/parquet-page-decoder ← PR-6a (this PR)
            └── gtt/streaming-merge-engine-merger ← PR-6b (single-RG merge — next)
                └── gtt/streaming-merge-engine-multi-rg ← PR-6c (multi-RG-at-metric_name)

Test plan

• 9 new tests in page_decoder.rs: single/multi-RG drains, multi-page columns, dict columns, null preservation, codec roundtrip (uncompressed/snappy/zstd), idempotent EOF, byte-exact reconstruction proof, I/O failure → PageDecodeError::PageStream
• All 14 existing streaming_reader.rs tests still pass with the header_bytes field addition (401/401 crate-wide)
• cargo clippy --workspace --all-features --tests with -Dwarnings
• cargo +nightly fmt --all -- --check
• cargo doc --no-deps -p quickwit-parquet-engine with -Dwarnings
• cargo machete
• bash quickwit/scripts/check_license_headers.sh
• typos clean on changed files

🤖 Generated with Claude Code