[JingsongLi]

Purpose

Align Native write validation, binary partition names and data-file creation timestamps with current Java semantics for existing Python read/write/update operations.

Companion Python/Java PR: apache/paimon#10660.

Brief change log

• Resolve configured or explicit row kinds and apply RowKindFilter before inspecting nested payload values. Keep schema/layout validation and Java's row-kind-specific top-level NOT NULL checks before filtering.
• Preserve actual NULLs in write-only PyArrow C Data transport, including sliced batches with nonnullable root fields. Core owns the logical row-kind-specific NULL checks; input metadata and nested fields remain intact.
• Preserve LargeUtf8 rowkind inputs, automatic/precomputed bucket writes, and the generated system-column schema. Ignored ARRAY/ROW/MAP child NULLs no longer fail a valid batch; surviving invalid children still fail.
• Use stable UTF-8 BINARY/VARBINARY partition values in both legacy-name modes, coordinated with the Java fix. Preserve other legacy types.
• Persist local wall-clock TIMESTAMP(3) creation fields, as Java's normal DataFileMeta writer factories do.
• Resolve local creation fields using system ZoneRules when generating REST partition statistics, and reuse that conversion for partition listing. Preserve exact milliseconds and Java DST gap/overlap behavior.
• Preserve epoch zero and negative creation times in partition statistics; merge multiple file times using Java's maximum regardless of input order.
• Assign append/PK creation times after auxiliary file indexes and managed Blob references finish.
• Finish postpone files in arrival order and assign creation timestamps at result/close time, matching Java's replay ordering. Remove the background-close and file-open timestamp machinery.
• Keep the existing REST authorization implementation and Python bindings. The proposed custom query-authorizer API, callable bridge and related tests have been removed, along with the earlier general Python FileIO bridge.

Tests

• filtered_write_validation_test: 4 passed, including 24 positive layout/kind/bucket combinations and schema/nullability/INPUT-changelog negative regressions.
• New filtered-write regressions first reproduce the failure on the old implementation. Cover ARRAY/ROW/MAP, configured/explicit row kinds, LargeUtf8, automatic/precomputed buckets, all-filtered inputs, schema failures, retained invalid payloads, top-level NULL rules and INPUT changelogs.
• Rust core library: 4307 passed, 6 ignored; creation_time_write_test: 5 passed; postpone_native_write_test: 5 passed.
• Actual append, PK, input-changelog, postpone and dedicated-Blob writes, commit, manifest roundtrip and applicable reads in Asia/Shanghai and America/New_York.
• Regressions first fail with the old premature index-output timestamps and epoch-zero/negative statistics.
• Exact creation-time wire values, epoch zero/negative partition maxima in both file orders, Java DST gap/overlap goldens and postpone close-time/rolling-order regressions. The independent clock oracle honors explicit TZ overrides on Windows.
• Existing Python-binding Rust unit suite, including Blob callback behavior and real PyArrow C Data write imports: 101 passed. Review follow-up also verifies ordinary STRING and LargeString rowkind inputs against DELETE, UPDATE_BEFORE, INSERT and NULL-key rules through the real WriteState path (8 combinations).
• Rust 1.98.0 core and binding all-target Clippy with -D warnings, rustfmt and diff checks.
• Companion Python and Java regressions. Local Python tests did not rebuild or install the Native extension.

API and Format

No Python binding or query-authorization API changes remain. Data-file creation timestamps retain their serialized TIMESTAMP(3) representation; wall-clock/epoch interpretation now matches Java. Binary legacy names use the existing UTF-8 cast instead of allocation identity text.

[leaves12138]

LGTM