[JingsongLi]

Purpose

Continue aligning Python and Rust execution APIs while preserving public PyPaimon contracts. This follows the merged #10712.

Depends on apache/paimon-rust#1188 for captured append writer configuration and the private row transport query. Native CI continues to build Rust main; this PR remains draft until that dependency is merged.

Changes

• Replace unbound executor mixin calls with explicit bound Python/Native scan and read executors. Both backends use the same actual operations, while public hooks continue to observe live configuration.
• Bind Python writer, row-ID/predicate update, delete, upsert, MERGE and shard executors to their public contexts. Prepared upsert/MERGE requests use matching execution entry points on either backend. No generic proxies, compatibility aliases or cross-backend imports are added.
• Give writer executors one explicit prepare_commit(identifier) boundary. Public batch facades accept no arguments and retain the one-time guard; public stream facades require a signed 64-bit identifier. Native facades enforce the same signatures before preparation, including after an explicit switch to Python.
• Align append reconfiguration with current Java: each partition/bucket captures its columns and Blob consumer, retaining them across prepares. Future buckets use the new configuration. Check every target layout before staging any group.
• Keep Native row conversion thin: ask Core for destination columns using routing metadata only, retain the captured order and avoid opening unused Blob inputs. NOT NULL/event validation precedes this query. Core owns routing, schema history and physical preflight.
• Delete the acceptance package and its tests, and remove its test/distribution configuration.
• Remove the Python benchmark package, its ACT/RoboMIND tests and documentation, the benchmark-only act extra, pytest input option and packaged experiment resource.
• Move deletion-vector bitmap codecs, file loading and row filtering to execution/python/deletionvectors, keeping shared DeletionFile / DeletionVectorMeta metadata in their domain modules. Update all callers and verify distribution contents.
• Separate array cast execution and executable-cast probes into execution/python/casting. Put shared Arrow read schemas and logical cast rules in schema, shared numeric/calendar/metadata formatters in utils, and private Python commit partition-spec parsing in its executor. Native read/commit adapters do not import Python cast implementations.
• Complete the remaining corresponding backend placement: move primary-key interval grouping, Blob View lookup, physical MAP/VARIANT codecs, dynamic-bucket HASH publication and MERGE condition/action execution into execution/python. Keep shared metadata/schema/value helpers, SQL parsing/validation, callable input adapters and cross-backend bucket input routing in their domain modules. Local and Ray MERGE use the same Python action factories.
• Remove the remaining benchmark package path references from the packaging checks. Add fresh-process tests that import every Native adapter with Python execution and SDK initialization blocked, and restore a real HASH index through the shared input assigner with Python execution blocked.
• Audit common against actual backend callers. Move external-file placement strategies into Python write execution and MAP BLOB physical key serializers into the Python backend. Split authorization JSON Arrow filtering/masking/CAST execution from shared field-reference extraction. Preserve public authorization factory signatures and shared path-factory configuration/existing-file lookup; load Python implementations only on explicit execution. URI/FileIO contracts, catalog-owned caches, generic integer codecs and shared predicate/output helpers remain outside the backends.
• Document execution contracts and writer lifecycle. Update existing regression expectations for Java-style reconfiguration and propagation of prefetch errors.

Verification

• Common/execution follow-up: 215 storage/Blob regressions passed, 117 Native isolation/protocol regressions passed, and 354 authorization/temporal regressions passed (168 Native cases skipped because the local SDK was not rebuilt). All 9 packaging tests pass and 22,363 current tests collect. Fresh-process checks block both execution backends during public authorization validation and block Python execution during Native existing-file path resolution with external placement configured.
• Independent cross-review confirms both whole-module relocations and 35 authorization execution functions preserve their AST bodies, while shared date-transform field dependencies remain unchanged.
• Final placement review: two independent cross-reviews found no further actionable issues. AST comparisons confirm moved functions/classes preserve their bodies; internal imports and distribution paths are updated.
• Read/codecs: 290 tests passed (177 subtests; 4 local-SDK-dependent Native cases deselected). Blob View integration: 11 passed.
• Boundary/Blob regressions: 207 passed; Native import/read/write/update protocol isolation: 117 passed. Local MERGE regressions: 48 passed, 16 optional cases skipped; two actual Ray MERGE regressions passed.
• All 22,362 current Python tests collect successfully; all 9 sdist/wheel packaging tests pass. The source tree has no remaining benchmark package path references.
• Cast relocation: 559 focused tests passed, 100 Native/optional cases skipped; 9 packaging tests passed. Fresh-process Native schema and commit-path adapters reject Pure imports, and all 22,357 tests collect successfully. Moved cast, calendar, schema and partition function bodies retain their prior semantics.
• Benchmark removal: all 22,357 remaining Python tests collect successfully; the 9 packaging tests verify wheel/sdist contents and updated extras. Documentation sidebar syntax validation passes.
• Deletion-vector relocation: 331 passed, 58 skipped; actual wheel/sdist contents test passes.
• Focused read/write/update execution, lifecycle, public-hook, prepare-identifier, backend import-isolation and reconfiguration tests.
• Native transport regressions use mock SDK handles; actual destination validation, routing, prepare/readback and callback GC are covered by the companion Rust tests.
• 9 actual sdist/wheel packaging tests pass.
• Full package Flake8, changed-file Python 3.6 grammar check, license headers and git diff --check pass.
• Independent API/boundary reviews confirm existing public method arguments/defaults, class paths, hooks and backend isolation.

The local installed SDK predates this change. It was not rebuilt or installed, and no Native version probing/fallback or CI dependency override was added.

[leaves12138]

LGTM