Request changes for one P1: the ORC revision makes libdeflate mandatory on ARM64, but this branch still links it only on x86. The inline finding identifies the missing companion dependency change. No existing blocking inline findings were present.

Reviewed all 17 changed paths and all three commits included by the ORC revision, with independent JSON/WAL, SCHEMATA/ORC-integration and risk-focused coverage. The other intended fixes look consistent with their caller contracts under static inspection.

Critical checkpoints:

• Goal and proof: names-only SCHEMATA avoids database-object enumeration; JSON/WAL preserve logical rows with empty projections; the intended ORC seek fix handles filtered stripe suffixes. Added tests target each behavior; they were inspected, not executed.
• Scope and clarity: local changes are focused; the dependency pin includes two additional decompression changes. The all-architecture libdeflate change needs the companion linkage fix reported inline.
• Concurrency: no new threads, shared mutable state or locking. Reader state remains scanner-owned; catalog name enumeration uses the existing cache/snapshot synchronization.
• Lock cost and order: no new critical sections, heavy work under locks or lock-order changes identified.
• Lifecycle and static initialization: logical row state resets per batch; malformed rollback restores it with columns. WAL captures row count before ownership moves. ORC stripe transitions reset readers/indexes/followers; decompression scratch retains its owning pool. No new ownership cycle or cross-translation-unit initializer dependency.
• Configuration: no configuration items added or dynamic-update contract changed.
• Compatibility: optional Thrift field 8 defaults to true, preserving legacy ID-returning requests; old FEs can ignore it. No storage/WAL format change. ARM64 link compatibility is the reported defect.
• Thrift/protobuf identity: compared the touched definition on master and branches 4.2, 4.1, 4.0 and 3.1. Master/4.2 have the identical optional field 8; older branches have no conflicting field. Reserved field 7 remains intact; no required field or enum change.
• Parallel paths: eager/lazy ORC and SARG on/off are covered by added tests; the dependency affects both reader generations. JSON roots/paths/outer arrays and WAL empty/subset/reordered/full projections were traced through table materialization.
• Conditional checks: JSON path absence, root selection, malformed-null handling and ORC stripe-end/range/EOF conditions preserve their documented boundaries; no additional defect verified.
• Test coverage: inspected FE name/privilege/pattern/default-wire tests; JSON malformed/empty/root/path/batch/partition tests; WAL multiple-block/projection/EOF/replay tests; ORC filtered-suffix, lazy/range/backward-seek and decompression tests. ARM64 linkage needs coverage after fixing the finding.
• Expected results: WAL golden rows, null counts, sums and predicate count match input; ORC expected IDs/payloads match the generator and static fixture stripe metadata. Query comparisons are ordered. Generation/execution of expected artifacts was not repeated.
• Observability: existing row, parse/materialization/filter, ORC seek/cache and I/O counters remain on the affected paths; no additional logging/metrics defect identified.
• Transactions and persistence: no EditLog, visible-version, publish/delete-bitmap or WAL-byte-format change. Replay transaction handling and failover contracts are unchanged by the row-count capture.
• Data writes, atomicity and crash behavior: production edits change enumeration/read accounting, not commit/write atomicity; no new crash-cleanup, leak or deadlock path identified in the changed lifecycle.
• FE-BE propagation: only SCHEMATA opts out of database IDs; its consumer does not read IDs. Other senders retain the default behavior; no further transmitted state introduced.
• Performance and memory: name enumeration avoids per-database object loading; row counting adds constant state. Lazy ORC compressed-input scratch still grows before split-chunk copies and is reused. No additional verified allocation or complexity regression.
• Other issues: the complete dependency comparison exposed the ARM64 link defect; no second actionable issue survived independent rechecking.

No additional user focus was supplied. All four initial risk questions have explicit dispositions in the ledger; there are no unresolved candidates or static coverage gaps. Review converged in round 2. Validation limits: static review only; no builds, tests, benchmarks or external SQL were executed, and submodules were left uninitialized.

Review converged after 2 round(s).

Static review of PR #70319 at head 47c2235 found no new actionable P0-P3 findings across all changed paths. The previously reported P1 on contrib/apache-orc (comment 4488893681) is resolved at this head: be/cmake/thirdparty.cmake now adds deflate on every architecture, so it is not reposted.

Critical checkpoints: JSON v2 now tracks accepted logical rows independently of empty file-column blocks; its array/root/path, malformed rollback, batch, EOF, filter, and TableReader paths remain aligned. WAL captures rectangular source-block rows before projection moves ownership, then passes that count through the existing materialized-filter pipeline. The FE name-only request uses optional Thrift field 8 with default true, so old or unset callers retain database IDs; only SCHEMATA opts out, and its non-ID response lists stay aligned with privileges, patterns, display-name formatting, and empty-catalog handling. The ORC revision advances from a filtered stripe suffix to the next selected stripe and the added tests cover initialized/fresh seeks, split ranges, lazy reads, min/max combinations, both columns, EOF, and backward seeks.

No new threads, locks, configuration items, storage/WAL wire formats, persistence paths, or ownership cycles were introduced. Added unit and regression tests were inspected for deterministic assertions and cleanup; they were not executed. This was a static review only: no builds, tests, benchmarks, or external SQL were run, and repository submodules remained uninitialized. No additional user-provided review focus was supplied.

Review converged after 1 round(s).