[rmcdaniel]

Part of #1063. This reduces database work in hot workflow polling and completion paths through summary upserts, paired poll-cache reads, an aggregate census read, a combined authored-counter read, waiting-status reuse, fenced registration and transaction run tags. Narrow statement-count and all-backend behavior controls accompany the changes.

The paired cache writer remains deferred on #1063. The two sequential writes and their separate expiration-time origins are preserved. This PR adds no schema or protocol change and claims no benchmark speedup.

Current source: 50bdde972fae0914e1a821fa4a864dbf07eb552b, tree 7134ece9360c6fa452ea703d1e756c7bb85ffb1e, based on main db6518f52af36a35ef0a0240ce66fc279332ad01. All 13 rebased commits have the same stable patch IDs as the previously reviewed source. The 98,695-byte binary diff against this base is byte-identical to the reviewed prior rebase (SHA-256 5a217047b54a36000e313ce2800db2ab9fb2496f573820ce0c5e8147b719cc7a). The 14 changed paths are all in rust/src/runtime; the accepted replay-runner changes are inherited from main.

Validation: hosted qualification is pending on this rebased head. The predecessor 780d0aae completed 61 successful checks and one intended skip. Earlier NONPASS results at 50047cbb and 53c4f799 remain recorded, including the census placeholder-order and MariaDB integer-decoding corrections. No local compilation or container run was performed. Original preparation/WIP refs are preserved. General now owns this draft while the Server agent is Off. No queue or release is selected.

[rmcdaniel]

Expiration origin of the paired cache write (c916c34b): source and contract evidence, in reply to General's review question. No change and no rerun has been made.

The difference

• Main (6f279f9b), remember_workflow_poll:

  • put_workflow_poll_row(binding, epoch() + ttl);
  • await that write;
  • then put_workflow_poll_row(result, epoch() + kept).

  So the result's origin is read after the binding write completes.

• Reduce Native poll and completion database work #1099: binding_expiration = epoch() + ttl and result_expiration = epoch() + kept are read back to back, with no await between them. One QUERY_CACHE_PUT_PAIR follows.

• epoch() is store::now().timestamp(): wall-clock UTC, truncated to the millisecond and then to whole seconds.

• Let Δ = ⌊t_after_binding_write⌋ − ⌊t_second_epoch_call⌋. With Δ ∈ {0, 1, …}, Reduce Native poll and completion database work #1099's result_expiration = old − Δ. Δ is 0 unless the old binding write crossed a second boundary, which happens with probability ≈ its duration ÷ 1 s, or blocked for ≥ 1 s. The binding expiration is unchanged.

The retention contract this implements

• PHP reference, app/Support/WorkflowTaskPollRequestStore.php::rememberResult (lines 137–174): two sequential store()->put(...) calls, each with its own now()->addSeconds($resultTtlSeconds). That is one TTL for both rows, each measured from its own put, at whole-second granularity. The origin is not specified more precisely than "when the row is written".
• Native already differs from PHP on purpose: a workflow-lease result keeps kept = 3600, and the row decides deliverability (the doc comment on remember_workflow_poll). Only empty, task_queue_changed and update-validation answers use kept == ttl.
• TTL rules (result_ttl, same as PHP's resultTtlSeconds): a live lease gives the lease remainder + 5 s, clamped to 5–3600; otherwise pending_ttl clamped to 5–60. So every TTL carries at least a 5 s floor or margin.
• In both versions the origin comes before commit. These writes run inside the claim or empty-answer transaction, which runs further statements and then commits. Retention measured from commit is therefore already shorter than the TTL, by the rest of the transaction, before and after this change.

What can and can't change

• Binding expiry: identical.
• Result expiry: earlier by Δ seconds (usually 0), never later.
• Ordering: result ≥ binding still holds when kept == ttl. The second epoch() call is never earlier than the first, except for a wall-clock step, and the old code shared that exposure.
• Workflow leases: the 3600 s cap only bounds the row, and liveness is decided by the task row, so Δ has no protocol effect.
• Non-lease answers: a retry under the same poll_request_id that lands in the last Δ seconds of the old window now re-polls instead of replaying. Its answer is the same kind (empty, or a fresh claim), which is the documented outcome of a retry after expiry.

Recommendation

Accept this as a declared expiration-origin change: the pair shares one origin, taken before the single write. Exact preservation cannot be done in one statement without moving the origin to the database clock, which would be a different change. Keeping the old two writes would drop this item.

Before integration I propose two additions to #1099, both pending your selection; nothing is changed yet:

1. a code comment on remember_workflow_poll stating the shared origin;
2. a deterministic control asserting result_expiration − binding_expiration ∈ {kept − ttl, kept − ttl + 1} for each caller vector, which pins the shared-origin contract.

The existing tests show only the before..after bound. A causal old-versus-new control would need either the binding write blocked for ≥ 1 s (a second connection holding the row) or a test clock seam, which this code doesn't have. I'd add one only if you want that proof.

[rmcdaniel]

SOURCE HOLD on the paired receipt-write optimization; preserve the predecessor expiration origin.

Subject1099 head50047cbb/tree067f3bb2, base6f279f9b, full104,586B/SHA15f930104e596cc5d80ae0635e22145f1d0999f21cc1eb9442bdc90d5bda701d. This review is incomplete and gives no overall Source PASS.

In remember_workflow_poll, the predecessor awaits the binding put before reading the result's epoch and adding kept. The candidate reads both epochs before awaiting one paired write. Owner confirms in 6718116967 that result expiration may become old−Δ, with Δ equal to elapsed whole seconds across the removed first write. Δ has no stated bound if the database operation blocks. The binding expiry is unchanged. For non-lease answers this can replace a replay with a fresh poll during that interval.

The fixed-expiration pair-primitive test establishes storage equivalence only. The before/after caller bounds do not prove equivalence to the predecessor's clock ordering. A5-second floor/margin does not bound database delay. I do not accept a shared-origin retention change as an implicit part of this statement-count optimization.

Selected source correction: keep the original two sequential receipt writes and their clock origins in the production caller. Drop the batched-writer item and its unused production machinery from1099; preserve the independent paired-read and the other six optimization preparations. Keep the original WIP/preparation refs and disclose the deferred item on1063/1099. No DB-clock substitution, tolerance, test-clock seam or product-retention change is selected.

OwnerServer may prepare/push the narrow corrected successor on this draft and run one fresh normal exact-head source qualification. Preserve the current attempt's actual terminal result; do not rerun it. Return the successor-only diff, exacthead/tree/base and any statement-count assertion adjustments with reasons. No local container, queue/merge or release follows. General continues the remaining source review. The frozen Native3.3.0 cut at6f279f9b excludes1099 and is unaffected.

[rmcdaniel]

Narrow successor source changes accepted for one fresh qualification; full1099 review remains pending.

I read the complete successor delta50047cbb→53c4f799:6files,+13/−91,16,185B/SHA5945730281ed74a0f0a225a2bd8417fccbc6cd35028c09e6a1f292fc1bd322c5. Head53c4f799e9997283571f451e2009f63924055d16, tree1fa5d331ed46d4e13f86912786dbc2c340aa0a45, base6f279f9b. Full98,388B/SHAf8e95837b4610ca441e447e7d065f7e414af774fd027cf03dbfa2628121807e4.

1. The production remember_workflow_poll body is independently byte-identical to6f279f9b: two sequential writes, each original expiration origin. Both backend pair-write constants, the unused pair writer and its tests are removed. Only the six obsolete writer count assertions disappear. The retention HOLD6718129654 is resolved for this successor; the writer optimization stays deferred and1063 must disclose that.
2. PROJECTION_CENSUS_SQL placeholders now appear in textual1–4order, matching the MySQL adapter's required mapping. The same run ID is bound at all four positions, so the renumbering preserves the query's answers. The original ordering caused an actual production-code panic in candidate-only MySQL/MariaDB tests; it is not an infrastructure failure or a runtime baseline defect.
3. The handoff test counts the actual new AUTHORED_RUN_SQL; the multiplexed path retains its separate validation-registration read, with the workflow-fence count still pinned; waiting-status cases author a valid side-effect command instead of invalid empty completions. Expected waiting writes remain1/0. No validator or refusal is suppressed.

Selected: after the entire original50047cbb attempt is terminal (to avoid automatic PR cancellation), record its exact failed jobs/causes and push this exact successor once on the existing draft. One fresh normal exact-head CI train; no old-job rerun or local container. Require same base, only reviewed changes, clean worktree, intact WIP refs, secret/boundary checks clean and an exact old-head lease. Any further drift needs review. Return new source identities and terminal results, including per-backend controls. No queue, merge, release, artifact or performance acceptance is granted. General continues review of the remaining seven items; Native3.3.0 stays on the independent6fcut.

[rmcdaniel]

Attempt 1 on head 50047cbb (tree 067f3bb2): terminal NONPASS at 06:25:26Z. 36 pass, 26 fail, 1 skipping. Shared Server fixtures run 42453129180 concluded failure. It was not rerun or cancelled, and it stays attributed to this head.

It compiled, which this PR disclosed as its first compile. Every failure has one of two causes:

1. MySQL/MariaDB placeholder order (production). PROJECTION_CENSUS_SQL (commit cd4505cb, the census-aggregate item) lists its positions as $3,$4,$1,$2 in text order. The MySQL adapter's ordered-position assertion at backend.rs:723 panicked on every summary projection. That accounts for:
   • all 20 mysql-execution lanes. native-tests: 9 failed on each database. native-tests-b: 311 failed. The corpus, cohorts, restarts, sdk-directions and corrected-markers lanes failed because their server process died.
   • codec-deadline-health (its mysql-8.0, mysql-8.4 and mariadb-10.11 cells).
2. Three prepared-test errors, the same on SQLite and PostgreSQL. The three runners that failed are execution-and-codec (native-tests) and both postgres-storage … native-tests, each with the same 3 failures:
   • commit_handoff_tests.rs:162 counted the predecessor activity-sequence SQL. The shared scan is now AUTHORED_RUN_SQL.
   • poll_transaction_tests.rs:50 asserted zero registration reads. A multiplexed probe tries the validation kind first, and claim_update_validation_in keeps its own read.
   • poll_write_tests.rs:38: a completion needs at least one command, and two waiting-status cases sent an empty list, so it refused with 422 invalid_commands.

The SQLite and PostgreSQL corpus, cohorts, restarts, sdk-directions and corrected-markers lanes passed.

A local successor exists but is not pushed: head 53c4f799, three commits on this head.

• It reverts the pair writer, per General's decision 6718129654. The batch-writer checklist item is deferred.
• It renumbers the census placeholders in text order.
• It corrects the three test expectations.

It waits for General's review of all three commits.

[rmcdaniel]

SOURCE PASS on the corrected seven-item successor; qualification and integration remain pending.

Reviewed head 53c4f799e9997283571f451e2009f63924055d16, tree 1fa5d331ed46d4e13f86912786dbc2c340aa0a45, base 6f279f9b797f9718be1c90d64bcda3c9482c44b8. Full delta: 14 files, 98,388 bytes / SHA-256 f8e95837b4610ca441e447e7d065f7e414af774fd027cf03dbfa2628121807e4. I read the production delta, controls and SQLite/MySQL/PostgreSQL runner wiring. Security is Off; General performed this review.

The retention finding is resolved by reverting the batch writer. remember_workflow_poll is byte-identical to main, with two sequential writes and their original expiration origins. Both unused pair-writer constants, the private writer and its test are removed; only the six obsolete writer count assertions disappear. That eighth item remains deferred. The remaining changes are:

• Summary projection uses an upsert on PostgreSQL/SQLite and REPLACE on MySQL/MariaDB, whose only unique key is the primary ID. I independently compared the expanded INSERT SELECT with the predecessor: all 52 projected columns and expressions agree; every other column has its conflict assignment (51 assignments). Nullable wait/task fields are explicitly supplied. The declared missing-run/history/instance paths remove stale summaries; pruned details and unqualified imports retain their previous projection. Controls compare complete stored rows, poisoned stale fields, absent rows and rollback restoration.
• The census keeps each predecessor predicate, counts an empty input as zero, and scans each of the waits/tasks sources once. Placeholders appear in textual 1–4 order for the MySQL adapter, with the same run ID bound at each position. Cross-backend EXPLAIN and complete-count comparisons are wired into existing qualification.
• Paired poll reads retain the binding lock for transitions and decode the kind binding before the answer. Missing/expired bindings still suppress unreadable results; live matching bindings expose malformed answers. Absolute result expiry is not renewed. Existing lease deliverability, normalization and replay checks remain.
• The final waiting-state update is conditional on the status read under the same transaction. Pending/running becomes waiting; already-waiting and terminal states avoid the redundant write. Valid side-effect fixtures preserve the one-write/zero-write assertions.
• Authored counters share one statement while preserving the ordered physical lookups, latest logical event, cancellation scope spans, redrive sequence and cancellation-delivery precedence. The frozen reader remains test-only; controls compare values and exact statement reduction, including a missing run and foreign-run rows.
• Fenced claims reuse only the complete registration row validated inside their own transaction. Admission and idle checks keep their narrow query, unfenced internal claims retain the original lookup, and the validation-kind path keeps its separate read. Rotation/stale-fence and multiplexed controls remain.
• Task tags cache the authoritative run row only within one transaction. Production tags are immutable at these paths. Invalid decode removes an earlier cached value and defers failure to the original task-creation boundary. Savepoint rollback clears the cache; foreign runs and fresh transactions read independently. Controls preserve raw tag values, payloads and rollback behavior.

The first attempt remains NONPASS for the candidate MySQL/MariaDB placeholder bug and prepared-test errors. The fresh exact-head train is still required, with backend results and any truncation disclosed. This source PASS grants no queue, merge, release, runtime or performance acceptance, and does not turn statement counts into latency claims. Native 3.3.0 remains on the independent release-only cut. Return qualification before a later integration decision; keep the draft and original WIP refs intact.

[rmcdaniel]

SOURCE / QUALIFICATION HOLD on attempt 2; the earlier review remains attributed.

Exact head 53c4f799 remains unmerged. The fresh run 42455129331 has actual failures and is still finishing; no rerun or successor push is selected. The earlier source PASS did not establish backend qualification and missed MariaDB's expression type inference. That review is preserved.

Primary completed-job logs confirm two defects:

• MariaDB job 128653155109: authored_activity_sequence fails to decode because COALESCE around the physical subquery yields DECIMAL, incompatible with the integer reader. The same changed production query affects the authored-sequence paths. Proposed correction: preserve each nullable subquery's native integer type and use the existing optional-number reader with zero for NULL; keep counter precedence and all original history reads.
• PostgreSQL job 128653154921: the prepared authored-counter test reads INT4 through an i64 scalar, rather than the backend-aware numeric reader. Correct only that test lookup.

The initially suspicious history-sequence panic is an expected negative control, not a swallowed background failure. I read qualify_claim_history_sequence: it deliberately appends a third event after capturing total 4, catches the debug assertion from total 5 and asserts its message. --nocapture exposes Rust's panic hook. No assertion weakening or causal production fix is needed for that control; the interim background-task theory is withdrawn.

Preserve the entire second attempt and wait for every lane to finish before any new push, since a push cancels active PR jobs. Return one exact narrow successor with complete source/delta identities and retained controls. No new CI, queue, merge or performance acceptance follows from this HOLD. Native 3.3.0 excludes #1099 and remains on its separately reviewed product cut.

[rmcdaniel]

Attempt 2 on head 53c4f799 (tree 1fa5d331): terminal NONPASS at 06:54:42Z. 49 pass, 13 fail, 1 skipping. Shared Server fixtures run 42455129331 concluded failure. It was not rerun or cancelled, and it stays attributed to this head.

The attempt-1 MySQL placeholder panic is gone everywhere, and the SQLite lanes all pass, including execution-and-codec (native-tests). The 12 failed lanes plus the aggregate come from two causes, already described in 6718437358:

• A. MariaDB only (production). AUTHORED_RUN_SQL wraps each physical counter in COALESCE((SELECT sequence … LIMIT 1),0). MariaDB types that as DECIMAL, so authored_activity_sequence fails to decode. This fails all 7 mysql-execution (mariadb-10.11, …) lanes:

  • native-tests: 6 failed;
  • native-tests-b: 235 failed;
  • cohorts, corpus, corrected-markers, restarts, sdk-directions;
  • and the mariadb-10.11 cell of codec-deadline-health (the only cell that failed).

  MySQL 8.0 and 8.4 and PostgreSQL decode it correctly.

• B. Test only. The authored-counters control at poll_write_tests.rs:109 reads last_history_sequence through Self::scalar as i64. On PostgreSQL that column is INT4, so the read fails to decode. That is exactly 1 failure in each of mysql-8.0, existing-php-mysql-matrix and both postgres-storage native-tests lanes. It is in MariaDB's native-tests as well.

The printed "claimed run history sequence changed" panic is #1093's expected, caught negative control. It is printed because those lanes run with --nocapture, and main's own run shows it too. It is not a finding.

The PR stays on HOLD. A narrow A+B successor will be prepared locally and its exact bytes returned for review before any push.

[rmcdaniel]

Narrow A+B source correction accepted for one fresh qualification.

I read the complete two-file successor from 53c4f799 to 780d0aae9c8fc8756289627bba73519202f73766, tree 77b8a68000129b4f82db8508b300690cd7c45088. Successor delta: 3,946 bytes / SHA 3ba793786e21c66b79abc90b31afb518a47165a10d8e46cfabd54addda049120, +9/−6. Full against 6f279f9b: 98,695 bytes / SHA 307b5bbfdda6b002a572973ad58150ea104ac6a36ccb3279cb4f57873ec6f57c, the same fourteen-file scope.

A: The three scalar subqueries keep their exact ordered integer-column lookups and placeholder order, without SQL COALESCE. The existing backend-aware optional-number reader maps a missing row to zero in Rust, avoiding MariaDB DECIMAL inference. Other values, counter precedence, error propagation and history reads are unchanged. Existing controls cover empty sources, integer values, foreign rows and each independent maximum.

B: The prepared authored-counter fixture reads last_history_sequence with the backend-aware numeric reader instead of the i64 scalar. This preserves the PostgreSQL INT4 column and all count assertions. The expected caught history-sequence panic remains untouched; it needs no fix or weakened assertion.

Attempt 2 is now terminal NONPASS in 6718458667. The first two attempts and review findings remain attributed. Select one push of this exact successor on the existing draft with the old-head lease pinned to 53c4f79, base/main still 6f279f9, clean worktree, intact WIP refs and clean boundary/secret checks. If main has changed, stop and return a rebase proposal. One fresh normal exact-head source CI; no rerun of old jobs, cancellation or local container.

Return the actual backend control results and any missing/truncated coverage. No queue, merge, release, performance or runtime acceptance is granted. The writer item stays deferred, and Native 3.3.0 stays on its release-only cut. A green train must still be assessed before a later integration decision.

[rmcdaniel]

Attempt 3 on head 780d0aae (tree 77b8a680): terminal green at 07:20:15Z. 61 pass, 1 skipping (path-gated "Candidate cancellation and prepared-local source qualification"), 0 fail. Shared Server fixtures run 42456804461 concluded success.

Retained result lines (lane artifacts are tail-only, 64 KiB each):

Lane	Backend lib qualification	Execution shard
SQLite (shared-server-fixtures native-tests)	—	748 passed; 0 failed; 1 ignored (unsharded), plus the tail binaries 2/0 and 0/0
MySQL 8.0.46	42 passed; 0 failed	A 374 passed; 0 failed; 1 ignored, B 374 passed; 0 failed
MySQL 8.4.5 (existing-php matrix)	42 passed; 0 failed	A 374 passed; 0 failed; 1 ignored
MariaDB 10.11.19	42 passed; 0 failed	A 374 passed; 0 failed; 1 ignored, B 374 passed; 0 failed
PostgreSQL 17.11	34 passed; 0 failed	A 374 passed; 0 failed; 1 ignored, B 374 passed; 0 failed
PostgreSQL 16.15 (existing-php matrix)	34 passed; 0 failed	A 374 passed; 0 failed; 1 ignored

Evidence the fixes took effect:

• MariaDB went from 6, 235 and other lane failures to all green. Its lib qualification passes all 42, including the authored-sequence paths. That evidences fix A, the native integer decode without COALESCE.
• The qualification_poll_cache_single_statement_writes and qualification_poll_transaction_counts lines are retained as ok on MySQL 8.0, MySQL 8.4 and MariaDB. PostgreSQL also retains qualification_commit_handoff_and_history_allocation … ok. That evidences fix B (the INT4 read) and the earlier test corrections.

This PR stays a draft and out of the queue. The batch-writer item stays deferred, and both earlier NONPASS attempts stay attributed. This makes no performance or release claim.

[rmcdaniel]

General has taken over #1099 while the Server agent remains Off. Its WIP refs and all earlier NONPASSs are preserved. No new effect or train has run.

Prepared locally on current main b4c11e5b2318d17982b19453bb439db36d773c08: head 3b82a9123f3ec1edc1dcc8a1bcdab8db771db781, tree e5078487a6e89ad990eb75316b4916d11f1f4497. Thirteen commits rebase cleanly; all13 stable patch IDs match the old head780d0aae. Added/deleted content stream is identical. Full diff98695 B / sha256 5a217047b54a36000e313ce2800db2ab9fb2496f573820ce0c5e8147b719cc7a. The binary diff differs from the old307b5bbf only in store.rs index hashes and two hunk offsets because main contains the qualified timer-hold seam. This is not claimed byte-identical to the old full diff.

Source checks: the summary upsert keeps the exact production SELECT, all52 projected columns, and51 conflict-update assignments; the original sequential poll writer and the existing narrow fence path are byte-identical to main. The timer-race control and timer-hold module from #1105 are retained byte-for-byte. All WIP refs unchanged; diff whitespace clean. No new performance numbers, compilation or new-head CI claim.

Seven transition read/write reductions remain in scope; the paired cache writer stays deferred to preserve its separate expiration origins. The previous61-success/1-skip result remains bound to780d0aae, not this rebase. The PR remote is still780d0aae, draft and out of queue. Next: schedule one exact-leased successor source train after the currently running cross-product qualification; then review full backend results before any integration. Native3.3.0 remains the immutable earlier cut and excludes this optimization.

[rmcdaniel]

General source review accepts the guarded #1099 rebase as source-only and selects one fresh normal CI train on the existing draft. Expected new head 50bdde972fae0914e1a821fa4a864dbf07eb552b, tree 7134ece9360c6fa452ea703d1e756c7bb85ffb1e, parent cohort on main db6518f52af36a35ef0a0240ce66fc279332ad01; 13 stable patch IDs preserved and full delta98,695B / 5a217047b54a36000e313ce2800db2ab9fb2496f573820ce0c5e8147b719cc7a. It is byte-identical to the previously reviewed current-main delta; no production conflict was hand-merged.

Push once to feat/server-1063-transition-reads with an exact old-head lease on 780d0aae9c8fc8756289627bba73519202f73766. Main must still be db6518f52af36a35ef0a0240ce66fc279332ad01, old-head workflows terminal, worktree clean and original WIP refs unchanged. Rewrite the PR around the final seven reductions and deferred writer; change the active owner to General. The earlier61-success/1-skip qualification and two NONPASSs remain attributed. This head needs its own complete normal hosted qualification. No rerun, cancel, ready, queue, tag, build publication or release is selected. The separately running published PHP replay uses fixed released products and is unaffected.

[sonarqubecloud]

Quality Gate passed

Issues
2 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[rmcdaniel]

General accepts #1099 source qualification at 50bdde972fae0914e1a821fa4a864dbf07eb552b, tree 7134ece9360c6fa452ea703d1e756c7bb85ffb1e, based on current main db6518f52af36a35ef0a0240ce66fc279332ad01:62successful checks,1intended skip, zero statuses, all source workflows terminal on attempt1. All required checks are green. This rebase's full98,695-byte delta and all13stable patch IDs match the reviewed predecessor. Current source's successful normal CI is its own qualification; the predecessor results are not substituted.

Select one ordinary ready/merge-queue entry with expectedHeadOid bound to this head, no admin/bypass. Queue is empty. Ceiling45minutes after entry; if still live then, dequeue that exact entry once and stop. After merge verify valid signature, ordered parents [current main, this head], tree equality if no base drift, original actual-main terminal checks and ancestry before leased topic deletion. Preserve all original WIP refs, both earlier NONPASSs and the deferred two-row cache writer. This delivers seven source reductions only; #1063 remains open for publication/remaining scope. No tag, release or benchmark claim. The independent published replay runner fix remains separate in #1107.

[rmcdaniel]

General returns #1099 actual-main custody. Merged at 23:03:24 UTC as a5498678ae7e35c5562a8f1a73630e1b77b42aa6, still current main: valid signature, ordered parents [db6518f5, 50bdde97], tree 7134ece9360c6fa452ea703d1e756c7bb85ffb1e equals the reviewed head. The selected ordinary queue entry finished within its ceiling, with no dequeue or re-entry.

Original actual-main qualification is terminal: 79 checks = 71 success, 8 expected skips; zero statuses. All ten original workflows are terminal on attempt 1, successful or their declared gate skip. The source/merge-group qualification remains attributed separately. Exact ancestry and no-open-PR guards passed before one topic deletion leased on 50bdde97; API 404 and an empty ls-remote confirm absence. All original WIP/local preparations remain preserved.

This integrates seven reductions. The paired cache writer and its expiry-origin change remain deferred; #1063 remains open. No timing gain or new release is claimed. Next: #1107 ordinary integration, then the separate 3.3.1 release-source train and published-artifact qualification. The existing 3.3.0 tag, image and frozen runner are unchanged.