Requesting changes for four substantiated issues, represented by three inline anchors (the two config issues share the new definition).

1. Candidate pushdown can make the synthetic Roaring iterator the norm source and change BM25 score filtering/top-k results.
2. The mutable ratio accepts unsafe values that reach undefined floating-to-integer conversion, including a transient runtime-publication window.
3. The one-time gate misses candidate sets that become selective after earlier index conjuncts.
4. The 0.1 default cannot activate for the PR's stated 22.4%-selective short-key-only workload, so the claimed default approximately 4.5x improvement does not occur.

Critical checkpoint conclusions:

• Goal and test proof: direct exact-phrase bitmap restriction works in the added unit cases, but score equivalence, production activation, config boundaries, and the advertised default workload are not proved and are contradicted by the findings above.
• Scope and compatibility: the nine-file change is focused. It introduces no storage format, FE/BE protocol, persistence, transaction, data-write, rolling-upgrade, or cloud-specific behavior.
• Concurrency, ownership, lifecycle, and errors: the borrowed candidate bitmap is segment-owned and used synchronously; query objects do not retain it, the DEFER cleanup covers exits, and reader exceptions/status propagation remain sound. No new lock-order, static-lifetime, or memory-accounting issue was found.
• Parallel paths and cache behavior: non-scoring exact/sloppy/ordered match semantics, Boolean/null/virtual candidate-universe algebra, iterator sentinels, and full-result cache hits are sound. Candidate-restricted misses do not escape into a reusable full-segment cache. Multi-term phrase-prefix shares the scoring defect; single-term/phrase-edge/SEARCH-v2/SNII paths do not consume the candidate.
• Configuration, performance, and observability: runtime observation exists, but value safety/publication and default calibration are unsound. The gate also misses later selectivity crossings. Existing inverted-index timers expose aggregate work, but there is no candidate-engagement/cardinality metric.
• Verification: three normal+risk rounds converged with no new valuable finding in the final capped round. No additional user focus was provided. Per the review prompt, no local build or test was run; current CI has formatter/style/license and FE/Cloud checks passing, while BE UT, compile, performance, and automated code-review checks remain pending.

LGTM

Automated review result: REQUEST_CHANGES

One new P2 performance issue is reported inline: publishing candidate_rows currently prevents cache warming for legacy fulltext query shapes that never consume the candidate and still compute a full-segment bitmap.

Existing-thread status:

• The scoring path now selects a non-candidate norm source for the regular multi-term phrase path; no distinct new scoring issue was posted.
• The finite/use-site ratio guards and the 0.3 default address the unsafe conversion and stated 22.4% motivating workload.
• The column-predicate-to-expression threshold crossing is now refreshed. Expression-to-expression conjuncts are still evaluated as a batch and intersected only afterward, so a selective earlier MATCH expression cannot shrink the candidate for a later phrase expression; this remains covered by the existing threshold-refresh thread rather than receiving a duplicate inline comment.

Critical checkpoint conclusions:

• Goal and data correctness: Candidate intersection preserves row and score semantics within the published candidate for exact, unordered-sloppy, ordered-sloppy, and multi-term phrase-prefix evaluation. Full cache hits remain safe because SegmentIterator intersects them with _row_bitmap; no partial data/null/score result was found escaping into a full-result cache.
• Scope and performance: The implementation is otherwise focused, but the unconditional cache-insert guard creates a default-on repeated-query regression for non-consuming fulltext query families, so changes are requested.
• Concurrency and lifecycle: Evaluation is synchronous and segment-local. The non-owning candidate pointer is not retained, _row_bitmap is stable while a query consumes it, and DEFER clears it on success and all Status exits. No lock-order, thread-context, or cross-TU static issue applies.
• Configuration and compatibility: The mutable ratio is observed dynamically and guarded before floating-to-integer conversion; non-positive values intentionally disable the feature. There is no FE/BE protocol, storage-format, transaction, persistence, or rolling-upgrade change. SNII/query_v2 paths ignore the new field and remain behaviorally compatible.
• Tests and hygiene: The added unit tests cover candidate restriction, full-cover equivalence, BM25 norm equivalence, column-to-expression threshold crossing, reset, and non-finite direct assignment. Coverage is still missing for non-consuming cache warming and the expression-to-expression boundary. The read-only BE build-hygiene gate passed; per the review environment, no build or test suite was run, and live BE UT/compile/performance checks were pending during review.
• Observability: Existing inverted-index timers and cache statistics are sufficient to diagnose these paths; no transaction/write observability applies.
• User focus: No additional user-provided focus was specified.

Review convergence completed in two rounds. All Round 2 normal and risk-focused agents returned NO_NEW_VALUABLE_FINDINGS, and every candidate is accepted, duplicate-fenced, or dismissed with code evidence.

Automated review result: REQUEST_CHANGES

Two distinct new issues are reported inline:

1. P1 correctness: a candidate-restricted phrase TRUE bitmap is combined with a full-segment NULL bitmap and consumed as a full-domain compound result. Nullable AND can short-circuit before a later FALSE child, so NOT (A AND B) can drop a row whose SQL result is TRUE; virtual projections can likewise materialize NULL instead of FALSE.
2. P2 performance/style: the new release-mode _norm_source invariant check runs once per scoring match inside the document loop even though the source is immutable after add().

Existing-thread status:

• Multi-term phrase-prefix scoring can still select a UnionTermIterator whose norm() throws; this remains covered by discussion_r3864368344 rather than receiving a duplicate comment.
• The finite use-site guard and 0.3 default address the original unsafe-conversion/default-workload portions of discussion_r3864368348.
• Post-column refresh is present, while remaining within-phase refresh opportunities stay fenced by discussion_r3864368353.
• Cache insertion is still suppressed for non-consuming fulltext shapes and remains covered by discussion_r3868272817.

Critical checkpoint conclusions:

• Goal and proof: the change implements selective candidate pushdown for legacy multi-term phrase and phrase-prefix evaluation. Added unit cases prove standalone exact restriction, candidate-lead score equivalence for real term postings, full-cover equivalence, post-column threshold crossing, non-finite use-site rejection, and reset. They do not prove the new nullable compound/domain case, cold-cache behavior, or union phrase-prefix scoring.
• Focus/minimality: the 12-path change is focused on one BE config, the SegmentIterator/context handshake, phrase DISI support, cache eligibility, and tests.
• Concurrency: evaluation is synchronous and segment-local. The candidate bitmap is not mutated while the adapter consumes it; no new thread, lock, atomic, or lock-order surface is introduced.
• Lifecycle/static state: candidate_rows borrows SegmentIterator-owned _row_bitmap, is valid for the synchronous reader call, and is cleared by DEFER on normal, downgrade, and error exits. No static-initialization, ownership-cycle, or dangling-pointer issue was found.
• Configuration: the mutable ratio is observed at each refresh; the finite and 0 < ratio <= 1 use-site guard makes multiplication/conversion safe, and non-positive values intentionally disable the feature.
• Compatibility: this is BE-internal and adds no storage-format, persisted-state, public-symbol, FE/BE protocol, cloud, or rolling-upgrade change.
• Parallel paths/conditions: single-term phrase/prefix, other legacy fulltext shapes, SEARCH query-v2, and SNII do not consume the candidate. Leapfrog alignment and exact/sloppy/ordered matching are sound on the active candidate domain. The distinct full-domain three-valued consumer failure is reported inline.
• Tests/results: the new test is discovered by the recursive BE storage-test glob; the read-only build-hygiene gate passed. Per the review prompt, no local build or test suite was run. Live compile, BE/FE/Cloud UT, P0, cloud_p0, vault_p0, performance, formatter, style, license, and secret checks are green; NonConcurrent Regression is failing, while External Regression and the code-review check are pending.
• Observability: existing inverted-index timers, cache statistics, filtered-row counters, scoring instrumentation, and downgrade logging are adequate for this internal optimization.
• Transactions/persistence/writes/transmission: no transaction, EditLog, data-write, crash-atomicity, or FE-to-BE variable-passing path is changed.
• Performance/other: the non-owning Roaring adapter avoids a bitmap copy. The repeated release check is the distinct new hot-path issue; cache and threshold performance concerns remain in their live threads. No further substantiated issue survived the final sweep.

No additional user focus was provided. Review convergence completed in two rounds: both Round 2 normal full-review agents and the dedicated risk agent returned NO_NEW_VALUABLE_FINDINGS, and every candidate is accepted, duplicate-fenced, or dismissed with code evidence.

Review result: no new inline findings on 6c6514b156b7480c0fc75e61384246bb2c45e4db, but this is not an approval. Two substantiated P1 edge cases on the current head remain covered by existing threads, so I did not duplicate them:

• A VirtualSlotRef-wrapped compound can bypass the root-only candidate suppression and reach the same candidate-domain/three-valued shortcut problem already described in discussion_r3869542660.
• Scoring multi-term phrase-prefix can select UnionTermIterator as _norm_source, whose norm() throws; this is the same norm-capable-source problem already described in discussion_r3864368344.

Critical checkpoint conclusions:

• Goal, scope, and proof: the PR adds a segment-candidate handshake so multi-term legacy phrase-family queries evaluate postings/positions only inside a sufficiently small surviving row set. The implementation is focused across config, IndexQueryContext, the phrase DISI adapter, reader cache policy, SegmentIterator gating, and targeted BE tests. Direct tests cover candidate restriction, full-coverage equivalence, score equivalence, cache consumption policy, cross-reader reset, threshold refresh, compound-root suppression, non-finite config, and teardown.
• Data correctness and parallel paths: simple MATCH results are consumed only within the same _row_bitmap domain; direct compound roots are evaluated without the candidate; virtual results are materialized only for selected row ids. FullText/StringType/BKD/SNII selection, exact/sloppy/ordered-sloppy/prefix routing, single-term non-consumption, cache hits, and copied contexts were traced. Apart from the two existing P1 cases above, no additional row, null, scoring, or cache corruption path was found.
• Concurrency and lifecycle: IndexQueryContext is segment-local and these reader/query calls are synchronous. The borrowed Roaring bitmap outlives each query object and is mutated only after query execution returns. No new lock ordering, cross-thread publication, circular ownership, or cross-translation-unit static-initialization issue is introduced.
• Error and memory handling: reader exceptions are converted to Status, expression evaluation converts thrown exceptions before the candidate pointer is restored, and the enclosing DEFER clears the handshake on every exit. The adapter is a small query-scoped object; no significant untracked allocation or ownership leak was found.
• Configuration: inverted_index_candidate_pushdown_ratio is mutable and read at each refresh, so runtime changes are observed without restart. The use-site finite/domain guard protects the multiply/cast even during invalid transient publication; non-positive values disable the path and values above one are rejected. The strict threshold and 0.3 default cover the cited 22.4% workload.
• Compatibility and persistence: no FE/BE protocol field, function symbol contract, storage format, transaction, EditLog, or data-write path changes. No mixed-version or crash-recovery compatibility work is required for this BE-local optimization.
• Performance and observability: the candidate participates only in the leapfrog approximation and never in positional matchers; full cached bitmaps remain reusable, while actually partial results stay out of cache. Existing inverted-index timers and cache counters cover the affected path. No separate hot-loop, redundant-copy, or observability defect was substantiated beyond the already-filed threshold/cache/check threads.
• Tests and gates: repository header/build hygiene passed all layering, budget, PCH, extern-template, and unity-skip checks. Current-head CI shows COMPILE, BE UT, and P0 Regression passing. The NonConcurrent Regression check is red, but its TeamCity log returned HTTP 401 here, so this review could not establish whether that failure is related. Per the review-only instruction, no local build or test run was attempted.
• User focus: no additional focus was supplied; the full PR was reviewed.

Round 1 converged with both normal full-review agents and the separate risk-focused agent returning NO_NEW_VALUABLE_FINDINGS. All suspicious points were either cleared with code evidence or deduplicated against the existing threads above.

Automated review conclusion: request changes for one P2 performance issue.

Reviewed exact head f82c48c054d6cdabee6d949c4be4beeb81848ae1 against base c800f2aa615e9f3ab445da5738c944644faeb986, using the authoritative bundle and existing inline threads as duplicate fences.

• Goal and focus: the patch is focused on restricting legacy multi-term phrase and phrase-prefix evaluation to selective scan candidates, and the local reader/iterator mechanism works. No additional user-provided focus was supplied.
• Correctness and parallel paths: direct candidate-domain evaluation, compound/virtual-root suppression, exact/sloppy/ordered/prefix matching, and exact-term norm sourcing are consistent. Single-term and other non-consuming legacy queries, query-v2 SEARCH, SNII, BKD, and ANN retain their full-domain behavior.
• Cache, lifecycle, concurrency, and errors: only a consuming legacy multi-term phrase marks a result partial, so those results stay out of the query cache while full-domain results remain cacheable. The candidate bitmap is scanner-owned and synchronously consumed, no reader retains it, and reset/error paths are covered by the established DEFER and Status boundaries. No new thread, lock, transaction, or static-lifecycle behavior is introduced.
• Configuration and compatibility: the mutable ratio is finite/range guarded before arithmetic, and its 0/1/exact-threshold behavior is a cost decision. There is no storage-format, protocol, persistence, data-write, or FE/BE compatibility change.
• Performance and observability: one production-reachable gap remains. The final refresh occurs before condition-derived ranges, per-segment delete bitmaps, and external scanner-split ranges are applied, so those late restrictions cannot engage the optimization even when they reduce a segment far below the 0.3 threshold. Existing timers remain sufficient for the path.
• Tests and hygiene: the changed tests cover direct restriction, score equivalence, phrase-prefix norm selection, partial/full cache policy, threshold refresh/reset, non-finite ratios, and nullable compounds, but not threshold crossing caused solely by late condition/delete/split pruning. This was a review-only run, so no build or test suite was executed; build-support/check-build-hygiene.sh passed on the pinned head.

Review completion: Round 2 converged with all normal and risk-focused reviewers returning NO_NEW_VALUABLE_FINDINGS. Every candidate is accepted, duplicate-fenced, or dismissed with code evidence; the single accepted P2 is attached inline.

Automated review conclusion: complete on head 4057b24. No new distinct valuable findings remain, so this is a summary-only review with zero inline comments. Existing inline threads were treated as hard duplicate fences and their affected current-head paths were rechecked.

Critical checkpoint conclusions:

• Goal and proof: the change makes multi-position phrase and phrase-prefix evaluation intersect the surviving SegmentIterator candidate while preserving real postings for positions and scoring. The changed real-index and SegmentIterator tests exercise restriction/equivalence, scoring norms, cache policy, threshold refresh, late delete/split/condition pruning, compound three-valued logic, config boundaries, and cleanup.
• Scope and clarity: the production change is focused on the candidate handshake, phrase leapfrog adapter, cacheability reply, pruning order, and the required dynamic threshold. The broader pruning reorder was traced through lazy initialization and preserves the final scan domain.
• Concurrency and lifecycle: no new concurrent execution or lock interaction is introduced. Each SegmentIterator owns the context and evaluates synchronously; the non-owning candidate pointer targets the address-stable row bitmap, is read only during a query, and is cleared by the enclosing DEFER on normal, Status-error, and exception exits. No static-initialization or ownership cycle issue was found.
• Configuration: the mutable ratio is read at each refresh, accepts non-positive values as disabled, validates finiteness and the upper bound, and repeats the full finite/domain guard before bounded floating-point arithmetic and conversion.
• Compatibility and parallel paths: there is no persisted format, RPC/protocol, symbol, transaction, write, or rolling-upgrade contract change. Multi-position phrase-family paths consume the candidate; single-term, ANY/ALL/equal, regexp/edge, range, BKD, direct, SEARCH query-v2, and SNII paths remain full-domain or use their established behavior. FullText and StringType cache paths reset and classify results per search.
• Conditions and error handling: threshold, compound-root suppression, VirtualSlotRef unwrapping, and cache-insertion conditions were checked against concrete failure scenarios. Existing catch-and-convert boundaries and Status propagation remain intact.
• Testing and hygiene: no compilation or unit suite was run by this reviewer because the review workflow prohibits builds. The changed tests and their registration were inspected, and the pure-text BE build-hygiene gate passed. Author-reported ASAN/build evidence in the live threads was considered context but was not independently rerun.
• Performance and observability: the candidate participates in the existing document-frequency ordering without entering positional matchers; cache hits remain usable and only actually partial leaf results skip insertion. Existing inverted-index timers and cache statistics remain available. No substantiated new performance or observability defect was found.
• Data correctness, persistence, and writes: the change is read-only and does not alter visibility versions, transactionality, persistence, or write atomicity. Delete bitmap and scanner-range restrictions are applied to the same final scan bitmap before index evaluation.
• Additional review: a suspected enclosing SEARCH DSL cache interaction was independently traced and dismissed because cacheable legacy SEARCH uses query-v2 objects that do not consume candidate_rows, while the SNII reader branch disables the enclosing DSL cache. No accepted or unresolved candidate remains.

Convergence: both complete-coverage reviewers and the separate risk-focused reviewer returned NO_NEW_VALUABLE_FINDINGS in Round 1, so the review converged without reaching the three-round cap.

User focus: review_focus.txt specifies no additional focus; the whole PR and its upstream/downstream interactions were reviewed normally.