[JingsongLi]

Purpose

Complete the remaining verified Native write/update nullability gaps for existing PyPaimon APIs, using current Java semantics.

Depends on apache/paimon-rust#1179. Native CI continues building Rust main; keep this PR draft until the Core fixes land.

Changes

• Match Java's separate nullable KV value type for non-cross-partition deduplicate tables. Valid DELETE records can omit non-key/non-sequence NOT NULL values without changing the logical schema.
• Apply logical row-kind-specific NULL checks before optional filtering on Arrow and InternalRow writes. Preserve strict INSERT/UPDATE and other-engine/cross-partition rules.
• Use the physical value schema for Arrow/Pandas transport and Parquet system-field output, preserving keys, sequence fields, nested constraints and metadata.
• Validate actual selected row-ID update output against its declared NOT NULL fields, after matching/deduplication/merging, and restore physical root field flags without casting hidden children.
• Apply the same target constraint to dedicated BLOB output, which bypasses the ordinary Arrow writer. Nullable Blob values and NULL children of non-NULL containers remain valid.
• Add batch/stream Native regressions for upsert/MERGE source NULLs that Core must discard, backed by the companion Rust transport fix. Python Native APIs remain thin wrappers.

Verification

• Review follow-up: chunked MERGE source regression covers batch/stream execution and NULL chunks in either order, NULL keys, unmatched NULL payloads, and dictionary values that decode to NULL. Final source-nullability suite: 25 passed, 25 Native skipped locally; Flake8 and Python 3.6 syntax pass. The companion Rust binding suite executes the real Core MERGE path without installing a Python extension.

• New retained-delete and adjacent writer regressions: 237 passed, 121 skipped locally. Actual commits/readback, sequence ordering, physical schema, required fields and existing writer APIs remain covered.

• New update-source and target-nullability matrix: 21 passed, 21 Native skipped. Covers Arrow/GenericRow upserts, MERGE, unused source columns, discarded duplicates, selected invalid NULLs, nullable targets, dedicated Blob and Blob collection children.

• Final combined update/upsert/MERGE/chunked writer suite: 281 passed, 21 Native skipped, 44 subtests passed. Includes restoration of NOT NULL physical root fields and the existing RecordBatch columns protocol.

• Final changed Python files pass repository Flake8, Python 3.6 syntax parsing and diff checks.

• Java's existing deduplicate-DELETE/sequence reference tests: 2 passed.

• Companion Rust retained-delete and filtered-write suites, 104 Python-binding Rust tests, 4319 core library tests (7 ignored), Rust 1.98.0 all-target Clippy and rustfmt.

• Reproductions first fail on the old implementation, then pass. Local Python tests did not rebuild or install the Native extension; Native cases remain enabled for CI after the Rust fix lands.

No new dependency, public API, extension-version fallback or implicit commit-message deletion is added.

[leaves12138]

LGTM