Reviewed the complete four-file diff at 1cb6c79. Two P2 findings need correction: normal grouped changelog loads never report their attached changelog rowsets, and the new Java import violates the mandatory FE Checkstyle gate. Both are posted inline. There were no existing inline findings to carry forward and no additional user-provided focus points.

The review converged in 2 rounds: all three independent reviewers returned NO_NEW_VALUABLE_FINDINGS in the latest round, every changed file and assigned mechanism was covered, and all candidates and initial risks were verified, deduplicated or dismissed with source evidence.

Critical checkpoint conclusions:

1. Goal and proof: separate base/changelog load counters are the intended behavior. The FE accumulator works for directly supplied values, but the BE loop sees base transaction rowsets while the changelog resides in attach_changelog; therefore ordinary loads do not populate the new counter. The added test bypasses this producer and does not prove the complete behavior.
2. Scope and reuse: the four-file change is focused and reuses the existing tablet delta report and statistics serialization.
3. Concurrency and locking: the publish parent owns the report map and joins its workers; existing changelog-first tablet locks are unchanged. All accesses and lazy initialization of the new map take the TableStatsMeta monitor. No added IO, journal waits, or lock-order cycle was found. A repair must read the final attachment count after workers finish because MoW publication can append changelog rows.
4. Lifecycle and initialization: TabletTxnInfo retains attachment ownership through publication and recovery. The new map belongs to the statistics record, survives ordinary analyze updates, and is discarded with that record. No new static initialization or reference cycle is introduced.
5. Configuration: no configuration is added.
6. Compatibility: the optional Gson field is initialized lazily when absent, following existing concurrent-map serialization patterns. The existing report field is reused; no new protocol shape or toolchain/library dependency is introduced. Older producers do not supply changelog history.
7. Thrift/protobuf IDs: no definition is changed, so no field-ID or required-field compatibility change applies.
8. Parallel paths: normal, retry, asynchronous and cloud reporting paths were traced. Existing replica averaging/missing-report approximations remain. Cloud/legacy events still report base table scalars and do not populate the new index counter; this is a feature limitation, with no separate current consumer failure established.
9. Conditional logic: comments explain why changelog rows differ from base rows, but the BE condition tests the wrong transaction representation (first inline finding). Zero additions intentionally leave the map unchanged. Missing-tablet/checkpoint concerns were checked against actual index rebuild and replay registration, without another reachable regression.
10. Test coverage: the new deterministic JUnit case covers additive map behavior, zero additions and base isolation. It adds no publish/report integration, routing, replay or serialization coverage. Existing attachment recovery tests were inspected. No builds, tests, Checkstyle execution, benchmarks or runtime validation were performed under the static-only contract.
11. Expected results: the added direct-map assertions match their inputs; no regression output artifact is changed.
12. Observability: existing tablet/transaction identifiers and report logging remain. No separate logging or metrics defect was substantiated.
13. Persistence and failover: transaction delta maps already persist with VISIBLE state and use the same statistics application on replay; image/CREATE/TRUNCATE paths register changelog tablet metadata. The new map participates in statistics JSON/image persistence. No additional replay divergence was verified.
14. Writes and atomicity: the patch changes statistics reporting and estimates, without changing rowset visibility, version selection, delete bitmap publication or committed data writes. No new data-loss or memory-safety issue was identified by source inspection.
15. FE/BE passing: the existing callback and aggregation chain accepts tablet-keyed counts, but the attached changelog count is missing at its producer (first finding). No other new variable needs propagation.
16. Performance: the new FE classification adds one inverted-index lookup per reported tablet on changelog tables and bounded synchronized map work. No additional current-head performance regression was substantiated.
17. Other issues: the import-order finding is supported by error-severity CustomImportOrder, the Maven validate binding and the FE Checkstyle workflow. A suspected stream-wrapper null dereference was dismissed because the sole current consumer cannot reach that fallback with the wrapper's empty local partition map.

Coverage limits: this is a static review. The lifetime accumulator alone is not an exact retained-row or per-stream unconsumed-row count after TTL, statistics recreation or upgrade. The stream-estimation consumer described in dependent PR #69850 is outside this diff, so no end-to-end stream-plan improvement or failure is claimed here. There are no unresolved candidates or unread assigned source paths.

Review converged after 2 round(s).

Reviewed the complete four-file diff at exact PR head 5bb280e against base 374d87d. The intended goal is to carry the rows written to the attached changelog index into FE statistics so changelog scans can use a useful row estimate. The accumulator and its direct unit test cover the storage side, but two end-to-end paths still lose that information.

Critical checkpoint conclusions:

1. Goal and proof: the FE accumulator correctly keeps changelog rows separate from base rows for a direct tablet-record event, but a partial BE publish retry drops already-published tablets from the final report, and cloud commits never populate the accumulator. The added unit test does not exercise either producer path.
2. Scope and reuse: the changes are focused and reuse the existing tablet delta report, transaction publication, and TableStatsMeta persistence mechanisms.
3. Concurrency and locking: BE publication tasks run concurrently and are joined by token->wait() before the rowset count is read; existing changelog-first tablet lock order is unchanged. TableStatsMeta additions and reads are synchronized. The retry lifecycle still loses a successful tablet's report.
4. Lifecycle: TxnManager removes a successful tablet's txn entry during publish. The worker clears its delta map before retrying, so a later successful retry cannot rediscover that tablet's attached changelog rowset.
5. Configuration: no configuration is added.
6. Compatibility: the new Gson field is optional and lazily initialized, with no thrift/protobuf or wire-shape change.
7. Parallel paths: local BE tablet-record reporting is updated, but the cloud commit path still reports only table-level base-index rows. Asynchronous tablet-stat refresh may eventually fill the gap, but the immediate post-load estimate remains zero.
8. Conditional logic and data correctness: changelog rows are intentionally excluded from base-table updatedRows and partition base-index accounting. The two findings are statistics undercounting issues; rowset visibility and delete-bitmap publication are otherwise unchanged.
9. Persistence and failover: the existing transaction delta map is persisted with visible transaction state and replayed through AnalysisManager; the new per-index map participates in TableStatsMeta JSON persistence. No separate replay or locking defect was verified.
10. Testing and observability: the direct JUnit test covers additive accumulation, zero additions, and base-index isolation. No publish-path, retry, cloud, replay, or serialization integration test was added. Under the static-only contract, no build, test, Checkstyle, benchmark, or runtime validation was run. Existing identifiers and report logging remain sufficient for these paths.
11. Performance: the new local classification adds one inverted-index lookup per reported tablet and bounded synchronized map work; no additional current-head performance regression was substantiated.

Existing inline threads 4422920699 and 4422920705 were verified as fixed on this head and are not reposted. There is no additional user-provided focus.

The review reached the round limit after the required final sweep. All changed files and main producer/consumer paths were statically covered and the two candidates below were independently verified, but the latest delegated sweeps were interrupted by the temporary model-capacity error, so the contract's all-subagent NO_NEW_VALUABLE_FINDINGS convergence condition was not evidenced.

Review did not converge after 3 rounds. This automated review cannot pass the review check.

The final static sweep covered all four changed files and verified both findings, but the required latest-round delegated NO_NEW_VALUABLE_FINDINGS evidence was unavailable because both round-2 child sweeps hit the temporary model-capacity error. No unresolved code candidate remains; convergence is false solely because the delegated convergence evidence could not be obtained before the three-round cap.

Reviewed the complete four-file diff at exact PR head fecb190 against base 374d87d.

The PR's goal is to carry rows written to attached changelog rowsets into the FE per-index statistics accumulator. The grouped-load producer fix is present, and the direct accumulator test is correct, but three lifecycle paths still make the new count incorrect:

1. TableStatsMeta(OlapTable) creates a zero changelog baseline. After DROP STATS (or another record recreation), existing changelog rowsets are still readable while the physical changelog index count is unknown. StatsCalculator then falls back to getRowCountWithDeltaRows, whose new branch returns only post-creation indexUpdatedRows, underestimating the visible changelog until a later tablet report.
2. Partial TRUNCATE PARTITION replaces changelog tablets but retains the table statistics record. The asynchronous dropStats path invalidates column/partition statistics without clearing indexUpdatedRows, so the branch continues returning rows from the removed partition.
3. A load whose partition is force-dropped before visibility can still carry old tablet delta rows. DatabaseTransactionMgr removes the missing partition from commit info, but still forwards the transaction delta map; after the old tablet metadata is removed, isChangelogTablet returns false and this code routes those changelog rows into the base updatedRows counter permanently.

Critical checkpoint conclusions:

• Goal and proof: the normal attached-rowset path now reports the final row count after token->wait(), and FE routes direct tablet records by index. The new unit test covers additive accumulation and base-index isolation, but it does not cover record recreation, partition replacement, missing metadata, or the publish/report integration.
• Scope and reuse: the four-file change is focused and reuses the existing tablet delta map and TableStatsMeta JSON persistence. The three findings above are in the new counter's lifecycle and consumer boundary.
• Concurrency and locking: BE publication workers are joined before the attached rowset metadata is read; existing changelog-first tablet lock order is unchanged. The FE counter methods synchronize on TableStatsMeta, and no new deadlock or blocking I/O was found. Existing publish-retry aggregation remains a known P2 thread and is not duplicated here.
• Lifecycle and initialization: the new accumulator has no partial-truncate reset and no pre-existing-row baseline, and the missing-metadata branch silently chooses the wrong index. No static initialization or ownership cycle was introduced.
• Configuration, Thrift/protobuf, ABI, and protocol compatibility: no configuration, generated definition, field ID, wire shape, or toolchain change is present.
• Parallel paths: local tablet-record reporting is updated. Cloud table-level reporting still omits changelog rows, and publish retries can lose successful siblings; both are already reported by existing inline P2 comments 4424460818 and 4424460813 and are not reposted. The grouped-load omission and import-order comments are fixed on this head.
• Error handling and data correctness: rowset visibility, version selection, delete bitmaps, and committed data writes are unchanged. The new failures affect statistics/cardinality accounting; unknown tablet metadata must not be silently treated as a base tablet.
• Persistence/failover: the optional Gson field participates in normal serialization, but stale or zero lifecycle state is what causes the findings above. Transaction delta maps are still applied through the existing visible/replay path.
• Testing and observability: only the added deterministic JUnit behavior is represented in the diff; no publish-path, replay, truncate, serialization, or cloud integration test was added. Under the static-only contract, I did not run builds, tests, Checkstyle, benchmarks, or runtime validation. Existing transaction/tablet identifiers and logging are sufficient for the reviewed paths.
• Performance: the new classification performs one inverted-index lookup per reported tablet and bounded synchronized map work; no separate current-head performance regression was substantiated.
• User focus: no additional user-provided focus points were supplied.

The review converged after 2 rounds. All three assigned round-2 rechecks returned NO_NEW_VALUABLE_FINDINGS; all changed files and assigned mechanisms were covered, and every candidate was verified, deduplicated, or dismissed. The review remains static-only, so runtime ordering of asynchronous reports is a stated coverage limit. Existing P2 threads 4424460813 and 4424460818 remain applicable to this head but are not duplicated. No existing P0/P1 inline finding applies, so existing_blocking_comment_ids is empty.

Review converged after 2 round(s).

Static review of PR 69896 at head fecb190 converged after three bounded rounds. Four new P2 findings are inline: the synthetic changelog wrapper returns a zero row estimate before the new TableStatsMeta fallback; user-injected base stats bypass the changelog accumulator; TTL compaction leaves the persisted accumulator stale; and per-partition replica allocations are divided by the table default. Existing review context was independently revalidated: stale/unknown tablet routing (r4425241009), partial-truncate reset (r4425241002), recreated-statistics baseline (r4425240994), and the cloud limitation (r4424460818) still apply and were not reposted; the retry concern (r4424460813) is shared pre-existing base-delta behavior; the checkstyle import issue (r4422920705) is fixed. No additional user focus was supplied. BE publication, FE routing/locking, lifecycle, serialization, and test coverage were inspected; no new issue was found beyond the inline set. This was a static-only review: builds and tests were not executed. All anchors were verified against the authoritative diff, and all final reviewers returned NO_NEW_VALUABLE_FINDINGS.

Review converged after 3 round(s).

Reviewed exact head 9fb10c4c0e147d4b9b079b4e57961ab39d03bdaf. One new P2 finding is inline: unconditional companion rows change the semantics of a base-only report consumed by older FEs during supported BE-before-FE upgrades. No existing P0/P1 finding remains to carry forward.

The reviewed flow is grouped load -> attached changelog rowset -> publish workers/MoW append -> token wait -> tablet delta report -> transaction visibility/replay -> partition MaterializedIndex -> public row-count lookup.

Critical checkpoint conclusions:

1. Goal and evidence: the same-head shared-nothing path supplies final changelog counts to partition estimates. The added FE test's assertions cover both public getters; full transport/P/Q integration is not tested here.
2. Scope and reuse: the production change is focused and reuses existing reporting and partition index state.
3. Concurrency and locking: master visibility and follower/checkpoint transaction replay hold table write locks, excluding TabletStatMgr's read-locked count replacement. The coordinator reads rowset counts after worker completion. No new lock order or heavy locked operation was found.
4. Lifecycle/static initialization: shared ownership retains attached rowsets through reporting; DROP STATS preserves index counts and partition replacement discards the affected baseline. Checkpoint loading/replay rebuilds tablet metadata. No new static initializer, cycle or leak was found.
5. Configuration: none added.
6. Compatibility: existing storage fields remain usable, but reused report semantics are incompatible with older FE readers; see the new P2.
7. Thrift/protobuf IDs: no definitions changed, so cross-branch field-ID comparison is not applicable.
8. Parallel paths: cloud commits and async/cross-attempt publication still depend on later tablet statistics for omitted changelog deltas. Existing threads already cover cloud and retries; they are not reposted. Wrapper delegation and injected-stat priority remain consumer work discussed in #69850.
9. Conditions: attachment-null, missing-tablet and removed-partition branches have concrete cases and nearby rationale; partition-specific replica allocation is used for changelog rows.
10. Test coverage: inspected positive and negative FE assertions cover mixed one-/three-replica partitions, missing metadata, DROP STATS, baseline replacement and partial TRUNCATE with late reports. BE emitted-map/MoW integration, mixed-version and end-to-end consumer tests remain absent.
11. Test results: no result artifacts changed. This was static inspection only: no builds, tests, benchmarks or runtime validation were executed.
12. Observability: existing publish and delta-report logs provide transaction/tablet/count context; no additional critical gap was found.
13. Transaction/persistence/failover: visible transaction state carries the delta map and replays through the same path; MaterializedIndex.rowCount already persists. No new journal record is required. Counts remain approximate across independent statistics samples and failover.
14. Data writes/atomicity/crashes: rowset visibility, versions and delete-bitmap transitions are unchanged; the new state is an estimate. Known failed-attempt omissions and stale/double-counted baselines do not establish an exact count or upper bound.
15. FE/BE propagation: same-head transport reaches the correct index and keeps changelog rows out of base/partition update counters; mixed-version propagation is the inline defect.
16. Performance/memory: additional work is linear in reported tablets with one metadata lookup and a shared pointer per companion; no material new retention or performance defect was found.
17. Further sweep: all three changed files, upstream/downstream paths, candidate dispositions and risk items were rechecked. No other verified new finding or unresolved static candidate remains.

Existing review context: attached-rowset reporting and import ordering are fixed. The revised partition-index design addresses the reported replica-allocation, missing-tablet, DROP STATS, TRUNCATE and lifetime-TTL accumulator issues. Retry/cloud omissions remain acknowledged limitations; wrapper/injected-stat concerns remain consumer follow-ups. All existing findings are P2, so existing_blocking_comment_ids is empty. There was no additional user-provided focus.

Static review converged in 2 rounds. All assigned reviewers completed the final pass without additional valuable findings. Coverage limits are the unexecuted integration/upgrade behavior and absent tests described above; all changed paths and static risk questions were covered.

Review converged after 2 round(s).