[JingsongLi]

Purpose

Align existing PyPaimon Native write validation, binary partition naming and data-file creation-time semantics with current Java, and remove unnecessary Native fallback caused by unused context fields.

Depends on apache/paimon-rust#1172 for the corresponding Native behavior fixes. Native CI continues building Rust main; this PR remains draft until the Rust fixes land.

Brief change log

• Let Rust own Native payload normalization after row-kind filtering. The Python facade checks compatible logical schemas without inspecting ignored nested values.
• Apply Java's top-level NOT NULL checks before filtering on Python Arrow/InternalRow paths, including the deduplicate DELETE exception for non-key/non-sequence fields.
• In the Python writer, filter first and then convert LargeString inputs and retain strict nested-schema validation for surviving values; preserve schema-only writer behavior and existing Arrow transport metadata. Native passes STRING offset layouts to Core without rebuilding payload arrays in Python.
• Keep Pandas DELETE payloads available until Java-equivalent row-kind NULL checks and filtering. Only filtering PK writers use a nullable top-level transport schema; logical NOT NULL declarations, nested fields, metadata and non-filtering writer checks are preserved.
• Add actual ARRAY/ROW/MAP write/read regressions for ignored DELETE/UPDATE_BEFORE, automatic/precomputed buckets and invalid retained values, plus batch/stream NULL-check and thin-facade regressions.
• Fix Java BINARY/VARBINARY legacy partition names: byte[].toString() used allocation identity and produced different paths for the same contents. Use the existing UTF-8 cast in both naming modes, preserving other legacy types. Align Python and remove obsolete Native binary-partition exclusions.
• Resolve data-file creation fields in the system time zone for REST statistics, preserving exact milliseconds and Java DST gap/overlap behavior.
• Preserve epoch zero and negative partition creation times; verify multiple-file maxima in both input orders through real REST commits.
• Take INPUT changelog creation times after the file write, independently of precomputed row statistics; verify batch/stream writes with an advancing deterministic clock.
• Preserve integer millisecond precision when Avro exposes serialized local TIMESTAMP fields as UTC datetimes.
• Match Java writer factories' TIMESTAMP(3) precision for append, PK/input changelog, postpone, dedicated Blob and Parquet vector files. Explicitly supplied factory timestamps remain unchanged.
• Update the existing PK manifest timestamp assertion to resolve local wall fields once and retain exact milliseconds.
• Keep Native commit and vector-search eligibility when unused hadoop_conf/prefer_io_loader/fallback_io_loader context fields are present.
• Remove the proposed custom query_auth integration, factory/session state plumbing and dedicated tests. Existing standard REST authorization, eligibility checks and custom-hook Python fallback remain in place. The earlier general Python FileIO bridge is also removed.

Tests

• Review follow-up: real LargeString Arrow write/commit/read regressions and Pandas row-kind/nullability regressions first fail on the old code, then pass. Final focused suite: 56 passed, 44 Native skipped; adjacent writer/schema suites: 207 passed, 169 skipped, 6 subtests passed. Invalid retained INSERTs, PK/sequence NULLs and ordinary append/PK Pandas constructor checks remain covered.

• Before the final Pandas-only transport adjustment, full related write/schema/Python 3.6 regressions: 437 passed, 392 skipped, 33 subtests passed. The final changed Python sources pass Flake8, Python 3.6 syntax parsing and diff checks.

• Additional Native NULL-transport regressions: 39 passed, 24 Native skipped locally; real PyArrow C Data and WriteState checks pass in the companion Rust binding unit executable (101 tests).

• New ignored nested-value write/read matrix: 15 passed, 15 Native skipped locally.

• Related write/schema/Python 3.6 regressions: 418 passed, 349 skipped, 33 subtests passed.

• Independent review caught and closed LargeUtf8, InternalRow kind, retained nested-schema and schema-only writer edge cases.

• New related writer regressions: 135 passed, 8 subtests; commit/creation-time regressions: 58 passed, 3 Native skipped, 10 subtests (one unrelated optional Lance case deselected).

• Final creation-time regression rerun (both file orders): 22 passed, 6 Native skipped.

• Previously verified related Python regressions: 247 passed, 234 Native skipped, 44 subtests passed. Restored Native-plan regression suite: 68 passed, 44 subtests passed.

• Standard REST authorization and custom-hook fallback, commit, Native vector context, binary partitions, real REST creation-time statistics, exact Avro timestamps and Pure writer metadata roundtrips.

• Java InternalRowPartitionComputerTest: 5 passed; Java sources are unchanged since the verified Checkstyle/Spotless run.

• All changed Python files pass repository Flake8 and Python 3.6 syntax parsing; diff checks pass.

• Companion Rust core, writer and Python-binding Rust unit suites plus pinned Clippy and rustfmt.

• Local Python tests did not rebuild or install the Native extension. Native cases remain enabled for CI with the Rust behavior fixes.

[leaves12138]

LGTM