[refactor](be) Encapsulate cached constant columns - #66614

Request changes: the production refactor is sound, but the test-contract migration is incomplete and normal BE-UT configurations cannot pass as submitted.

Critical checkpoint conclusions:

• Goal and proof: Centralizing the cached single-row data column removes caller-side physical-shape checks and preserves production behavior, but the changed/new tests do not prove a passing cross-build contract; two blocking test failures are called out inline.
• Scope and clarity: The change is conceptually focused. The accessor migration covers all repository consumers; the related VMatchPredicate removal was verified as a no-op for valid plans because MATCH retains a nonconstant slot child.
• Concurrency: The wrapper is immutable after construction and existing shared pointers are populated during expression open. No new thread, lock, atomic, dependency-order, or deadlock behavior was introduced; valid scan constant-false paths still signal readiness as before.
• Lifecycle, memory, and nullability: The wrapper owns the unwrapped ColumnPtr; FunctionContext clones share that ownership, so cached StringRef values remain live. Nullable and array/map/struct row-zero semantics and COW ownership are preserved, with no new untracked allocation or reference cycle.
• Configuration and compatibility: No configuration, FE-BE variable, serialization, storage-format, or rolling-upgrade contract changes. The internal source API migration is complete across the repository.
• Parallel paths and special checks: File, mock, OLAP, generic scan, scalar-function, collection-IN, constant rebroadcast, and ANN paths were traced. The one-row check is a valid production invariant, but its direct test/mock callers and build-mode-specific failure behavior were not fully migrated.
• Test coverage and results: Existing scan-normalization cases now fail during setup, and the new unconditional death test fails under NDEBUG. The visible macOS BE check is compile-only with MAKE_TEST=OFF, so it does not cover either issue. No local builds or tests were run because this review runner prohibits builds.
• Observability: No new runtime path requires logs, metrics, or tracing.
• Transactions, persistence, and writes: Not applicable; the PR changes neither transaction state nor data-write/persistence behavior.
• Performance: Constant unwrapping remains no-copy and is performed once at wrapper construction; no hot-path regression or redundant allocation was found.
• Additional user focus: none was provided. The complete PR was reviewed without further narrowing.

Review completion: converged in Round 2 with two inline findings and no unresolved candidate.

Review opinion: no new blocking findings at 82d95715c456989a9987e5b383c96bf9ee31d509. The two earlier inline concerns are addressed at this head, and the change looks ready from a code-review perspective.

Critical checkpoint conclusions:

• Goal and proof: The change consistently stores cached constants as owned, single-row physical data columns, unwraps ColumnConst without copying, and recreates ColumnConst only when broadcasting. Production get_const_col() executes for one row and validates size and declared type before caching. The added wrapper tests cover direct data, const unwrapping, and build-mode-specific invariant failure; the surviving scan tests cover both const-wrapped and direct one-row Boolean constants.
• Scope and clarity: The refactor is focused. Repository-wide searches found no stale public-field access or missed ColumnPtrWrapper producer/consumer, and the mechanical accessor changes preserve the existing row-zero semantics.
• Data correctness, nullability, and lifecycle: ColumnConst delegates Field, StringRef, null, and complex-value reads to nested row zero, so eager unwrapping preserves scalar, nullable, array/map/struct, JSON, regex, date/time, IN-set, and ANN behavior. The wrapper owns the nested column, and cloned FunctionContexts retain shared ownership; retained buffers remain live with no COW mutation, dangling reference, or ownership cycle.
• Concurrency: Constant wrappers are populated at fragment open before thread-local cloning and are read-only afterward. Scan false/NULL normalization still sets EOS and wakes the scan dependency before scanner preparation. No new thread, lock, atomic, lock-order, blocking, or deadlock behavior is introduced.
• Error handling: Null and one-row wrapper preconditions fail loudly. Production expression size/type mismatches are returned as Status before wrapper construction, and existing status/exception boundaries remain intact.
• Parallel paths and special checks: File, OLAP, Mock, and shared scan normalization paths, function-state consumers, constant rebroadcast, MATCH index/slow execution, and ANN TopN/range extraction were traced. Valid MATCH retains a slot child, so the removed root constant-cache call was a no-op; child constants and analyzer state are still initialized for original and cloned contexts. ANN keeps an owning nested Float32 pointer and preserves outer-null, empty, inner-null, array-shape, element-type, and dimension checks.
• Configuration and compatibility: No configuration, serialization, storage-format, FE-BE variable, external symbol, transaction, persistence, data-write, or rolling-upgrade contract changes apply.
• Test coverage and results: The two earlier test-contract findings are fixed at this head. Live BE UT, COMPILE, Clang Formatter, and relevant regression checks are passing. No local build or test was run because this review runner explicitly prohibits builds.
• Observability and performance: No new runtime path needs logs or metrics. Normalization is no-copy, removes repeated const-shape dispatch, and adds no hot-loop allocation or obvious CPU/memory regression.
• Additional user focus: No additional focus was provided; the complete PR was reviewed without narrowing.

Review completion: converged in Round 1. The main risk scan, two normal full-review partitions, and a separate risk-focused review completed; all returned no new valuable findings, every risk item has a concrete conclusion, and the final changed-file/unresolved-candidate sweep is clean. No new inline comments are needed.