[JingsongLi]

Purpose

Move distributed input bucket computation and POSTPONE input statistics/inference into Rust Core so PyPaimon/Ray can use the same implementation as local writers without creating file writers for preclustering.

Changes

• Add schema-based BucketRouter::new(&TableSchema), with_bucket_plan and input-order Arrow route APIs for fixed and POSTPONE buckets. Reuse existing bucket functions and write-array normalization; payload columns are not needed for routing.
• Expose pure PostponeBucketPlanner with input_statistics, plan and current_plan. Worker statistics can be concatenated and duplicate partitions are aggregated before inference; known layouts are preserved and invalid statistics leave the plan unchanged.
• Keep buffered automatic-writer ownership separate and reuse the same planner/router. Resolve write metadata through the existing write builder, preserving REST write authorization.
• Make PostponeBucketPlan Arrow codecs depend on TableSchema instead of a live Table. Add typed Python binding classes and SDK docs; no new dependencies.
• Match current Java's NULL BinaryRow layout for noncompact DECIMAL/TIMESTAMP/LTZ, including reserved variable slots and offsets. Cover exact Java bytes/hashes, nullable-key writer restarts, and exact NULL partition filtering. Existing format partition reconstruction uses the same typed-null helper to preserve its filtering behavior.

Java semantics

Reference fixtures were generated from current Apache Paimon Java InternalRowSerializer, RowHelper and BinaryRowWriter on paimon master 7befc15f18d81bc406679e14b48aa92cfe863335. Tests include precision boundaries, a following variable-length string, fixed/mod/Hive buckets, nullable/sliced FixedSizeBinary partitions, option precedence, explicit-plan replacement, overflow/cap behavior, and normalized PK event inputs.

Validation

• Core focused suites: POSTPONE, BinaryRow, BucketRouter, partition filtering and partition reconstruction.
• nullable_primary_key_test integration suite.
• Registered Python SDK write_routing API test.
• Joint PyPaimon execution against the registered current SDK: 116 passed, including all 8 new real Native routing/planning cases and existing writer/authorization regressions; 36 Native plans exercised. Used a temporary embedded test executable, without installing or replacing a Python extension.
• Rust 1.98.0 Clippy and formatting checks.

Paired PyPaimon PR apache/paimon#10770 uses these APIs for Ray preclustering and POSTPONE statistics/planning. PyPaimon public APIs remain unchanged.

CI follow-up

• Migrate the C binding bucket-plan import to table.schema() after the Core Arrow codec signature change. This fixes the workspace MSRV and Windows build type error.
• Re-encode logical partition keys with Java's typed NULL layout. Source Arrow partitions and planned split partitions now share the same identity, so NULL DECIMAL/TIMESTAMP/LTZ upserts update the existing row instead of appending a duplicate.
• Replace the old untyped-NULL bucket expectation with independently generated current-Java hashes/buckets, and cover two consecutive upserts while preserving the original row ID.
• Validation: Core library suite (fulltext,internal-api) 4629 passed, 7 ignored; Rust 1.94.0 workspace/all-targets/all-features check; Rust 1.98.0 workspace Clippy with fulltext,vortex and warnings denied; formatting and diff checks.

[leaves12138]

LGTM