Reviewed all four changed files at 68a2d36. No actionable new findings and no existing inline blockers. Review converged in 1 round: production, tests, and the separate norm-invariant reviewer all returned NO_NEW_VALUABLE_FINDINGS; all candidates and risk focuses are dispositioned.

The fix is supported by the complete scoring path:
MATCH_PHRASE_PREFIX -> prefix expansion -> PhraseQuery -> document intersection -> phrase frequency -> UnionTermIterator::norm -> BM25 score.
The scored document is nonterminal and matches the heap root. All prefix alternatives share the same reader and field; the pinned CLucene implementation obtains norms from that field's document-indexed array. Consuming phrase positions leaves the document and norm unchanged. Empty expansions retain a fallback child, and exhausted postings do not reach scoring.

Required checkpoint conclusions:

1. Goal: the delegation removes the reported data-dependent scoring exception. The new unit test and V2/V3 regressions target it; the r and gr cases make the union strictly lead by document frequency.
2. Scope: one production accessor and focused tests; no unrelated production change.
3. Concurrency and locking: query-local iterators execute synchronously under existing searcher ownership. No new shared mutation, thread, lock, heavy critical section, or lock-order issue.
4. Lifecycle and static initialization: existing owners retain children and reader through scoring. No new allocation, ownership cycle, static dependency, or cleanup obligation.
5. Configuration: no configuration item or dynamic-update behavior added.
6. Compatibility: signatures and serialized/index formats are unchanged; both V2 and V3 use the reviewed matching path.
7. Thrift/protobuf: no definitions changed; cross-release field-ID comparison is inapplicable.
8. Parallel paths: standalone prefix traversal never calls this norm accessor; the separate query_v2 union implementations already obtain a current-document child's norm. No omitted counterpart fix found.
9. Special conditions: existing nonempty, positioned-iterator prerequisites and the terminal-document scoring gate justify direct delegation. No new conditional is introduced.
10. Test coverage: single/multiple expanded terms, varied document lengths, multiple preceding phrase terms, both formats, and an unscored row baseline are covered. The mock always returns norm 1, so document-dependent mock norms would be optional strengthening; real regression scores already distinguish document lengths. Missing/exhausted posting cases were checked statically.
11. Expected results: every new row and rounded score follows from exact phrase positions and the branch's formulas: 8 documents, 20 tokens, average length 2.5, apple document frequency 6, red frequency 4, yielding 0.3544, 0.3008, and 0.9415. order_qt_ sorts comparison results, and LIMIT 10 includes all matches. Output generation provenance cannot be proved by static inspection.
12. Observability and errors: existing query timers and CLucene/Doris exception-to-Status boundaries cover the accessor; no new logging or metric is required. No failure is silently discarded.
13. Transactions and persistence: no EditLog, visibility-version, delete-bitmap, or failover behavior changes.
14. Data writes and crash safety: no production write, transaction boundary, or resource acquisition is added.
15. FE-BE variables: no new transmitted value or sending path.
16. Performance and memory: constant-time heap-root delegation; no new scan, copy, allocation, or memory-tracking obligation.
17. Other concerns: the branch's existing phrase-prefix IDF excludes expanded terms, as the PR notes; this change preserves that behavior. No new nullable-column, bitmap, or storage-visibility issue found.

User focus: no additional focus was supplied. Both initial risk focuses (norm validity and test/expected-score coverage) were resolved without an inline finding. No unresolved candidate or source-coverage gap remains.

Validation limits: static source review only. No builds, unit/regression tests, runtime queries, benchmarks, or sanitizers were run. Repository source was not modified and submodules were not initialized. The dependency norm implementation was inspected read-only at the checkout's pinned CLucene commit 08d45e541fef1e48dc4ab00d2d5fde9569b49ee3.

Review converged after 1 round(s).

Reviewed all five changed files at head 52df5f3. No verified new defect was found. There are no existing inline findings to retain. The review converged after one round: the production, test, and separate norm-risk reviewers all returned NO_NEW_VALUABLE_FINDINGS; all risk questions were independently checked and closed.

The complete path was traced: MATCH_PHRASE_PREFIX expansion -> same-field posting iterators -> union/phrase lead selection -> document alignment -> phrase-frequency counting -> norm -> BM25 collection. Position consumption preserves the current document, and the heap top is also the source of the union's document ID. The pinned CLucene dependency (08d45e541fef1e48dc4ab00d2d5fde9569b49ee3), inspected through read-only GitHub API requests, confirms that each child's norm comes from that document's field norm array.

1. Goal and evidence: the change removes the throwing norm implementation on union-led scored phrase prefixes. Added V2/V3 cases force union leadership for both one and multiple expansions and check scores and matching rows. These tests were inspected, not executed.
2. Scope and clarity: one constant-time delegation plus focused unit and regression changes; the comment explains document norm identity.
3. Concurrency and locks: no new concurrency or shared mutable state. Each synchronous query owns its posting iterators and heap; cached searcher ownership remains retained during search. No lock order or work under locks changes.
4. Lifecycle and initialization: shared child ownership keeps heap pointers valid. No owner cycle, cleanup obligation, global, or static initializer is introduced.
5. Configuration: no configuration item or dynamic update behavior changes.
6. Compatibility: no function signature, encoded norm representation, storage layout, or wire format changes; no rolling-upgrade adaptation is required.
7. Thrift/protobuf: no definitions are touched, so release-branch field/enum comparison is inapplicable.
8. Parallel paths: both V2/V3 MATCH_PHRASE_PREFIX cases reach this legacy implementation. Plain PrefixQuery does not read union norms; sloppy phrase paths reject multi-terms. The separate query_v2 UnionPostings already delegates to a child on the current document.
9. Special conditions: no new condition is added. Expansion has a nonempty fallback, and scoring reads norm only after a finite aligned document with positive phrase frequency. Empty/default/exhausted unions are not newly reachable through this call.
10. Test coverage: changed unit cases, registration, fixture setup, deterministic order_qt output, multiple expansions, three-term phrases, unequal document lengths, nonmatches, and the unscored control were checked. Existing unit cases also cover cross-document movement and exhaustion.
11. Expected results: every new row and score is consistent with the branch formulas. Eight documents contain 20 tokens (avgdl 2.5); apple/red document frequencies are 6/4. The resulting rounded scores are 0.3544, 0.3008, and 0.9415, matching the oracle. This branch's existing plain-term-only phrase IDF behavior remains unchanged.
12. Observability and errors: existing search timers and CLucene/Doris exception conversion cover this path; a scalar read adds no logging or metric requirement and discards no Status.
13. Transactions and persistence: no EditLog, failover, committed-data visibility, visible-version, rowset, or delete-bitmap behavior changes.
14. Writes and atomicity: no production data-write or transaction path changes; fixture inserts use existing mechanisms. The new read introduces no resource acquisition or crash cleanup.
15. FE/BE propagation: no new transmitted variable; the existing similarity flag is preserved through phrase-prefix expansion and child creation.
16. Performance and memory: the method is O(1) with no new allocation, copy, scan, or memory-accounting obligation.
17. Other correctness: document/position separation, shared-document ties, missing expansions, omitted norms, field identity, and reader boundaries were rechecked. No nullable-column or three-valued bitmap semantics change, and no additional issue was substantiated.

No additional user focus was supplied. All changed files and assigned risk scopes are covered; no defect candidate remains unresolved. Validation is static only: no builds, tests, benchmarks, or runtime checks were performed, and submodules were left uninitialized. Test breadth limits remain: mock norms are constant, multi-expansion matches have equal lengths, and the new suite does not separately exercise null/empty input or multiple scanners/segments. Runtime results and oracle-generation provenance were not independently validated; these limits do not establish an introduced defect.

Review converged after 1 round(s).

No new actionable defects found at head 365d0e9. Reviewed all five changed files and the complete caller-to-score path. Review converged in one round: both the complete-review and separate norm-risk checks returned NO_NEW_VALUABLE_FINDINGS; main independently verified their conclusions. There are no existing inline findings or outstanding candidates. No additional user focus was supplied.

The fix is consistent with the execution flow: MATCH_PHRASE_PREFIX expands its last token within one field, PhraseQuery intersects document iterators, exact matching consumes positions, and scoring reads the lead iterator's document norm. The heap top is a live child on that document. The pinned CLucene implementation refreshes its cached norm on next/skipTo and preserves it while reading positions, so same-document ties and exhausted siblings do not select a different document's norm.

Critical checkpoint conclusions:

1. Goal: the forwarding accessor removes the reported exception when a rare prefix union becomes the phrase lead. Added unit and V2/V3 SQL tests exercise the affected path; inspected only.
2. Scope and clarity: one production accessor change with focused unit/regression coverage.
3. Concurrency and locks: posting cursors belong to a local query and are consumed synchronously; no new shared mutable state, threads, locks, heavy locked work or lock order.
4. Lifecycle and static initialization: existing shared child ownership keeps heap pointers alive; no new allocation, ownership cycle, destruction dependency or static initialization.
5. Configuration: none added or changed; the regression sets an existing expression-pushdown variable.
6. Compatibility: no signature, on-disk index format, wire format or mixed-version protocol changes; V2 and V3 use the same norm behavior.
7. Thrift/protobuf: no definitions changed; release-branch ID comparison is inapplicable.
8. Parallel paths: checked non-leading unions, unscored phrases, single-token prefixes, ordinary/sloppy phrases and phrase-edge queries. SEARCH/query_v2 already reads a child norm for its current union document; no mirrored fix is needed. Existing single-token-prefix scoring behavior is outside this change.
9. Special conditions: nonempty expansion fallback, finite document intersection and positive phrase frequency establish the norm accessor's preconditions. Position exhaustion does not advance documents or invalidate the cached norm.
10. Test coverage: inspected child switching/ties in unit assertions and real-index V2/V3 cases with one/multiple expanded terms, multiple phrase terms, differing document lengths and an unscored control. Empty/exhausted and scoring-disabled paths were traced statically. Constant mock norms limit isolated unit sensitivity; the SQL regression checks real nonzero length-dependent scores.
11. Expected results: every new result row and score agrees with token positions and the existing BM25 formula: eight documents, 20 tokens, avgdl=2.5, apple document frequency 6 and red frequency 4 yield 0.3544, 0.3008 and 0.9415 as recorded. String casts reject zero scores through exact comparison; order_qt_ provides deterministic output. Existing omission of expanded-prefix IDF remains unchanged, consistent with the PR's stated scope.
12. Observability and errors: the existing query timers and exception-to-Status boundaries remain sufficient; no new distributed operation warrants additional logs or metrics.
13. Transactions and persistence: no EditLog, commit, replay, failover, visibility-version or delete-bitmap changes.
14. Data writes and atomicity: production code is query-only; no write/crash semantics changed.
15. FE/BE variable propagation: no new transmitted state or send paths.
16. Performance and memory: heap-top lookup and cached norm access are O(1), without new allocations, copies, scans or untracked ownership.
17. Further checks: rechecked heap ties, exhausted siblings, missing-prefix fallback, norm-loading flags, phrase-position consumption, multiple CLucene segments and score-test sensitivity. No reportable issue or unresolved question remains.

Coverage limits: static review only. No builds, tests, benchmarks, source edits or submodule initialization were performed. Runtime behavior and generated-output provenance were not independently validated.

Review converged after 1 round(s).