Repository navigation
Qualify paired namespace signals and reset cancelled MySQL sessions - #429
Conversation
|
The paired namespace-signal reference baseline now passes all nine published PHP/Python/Rust client/author directions and their canonical embedded counterparts on SQLite, at source The actual reference checks cover 36 original runs, 108 commands and 108 tasks, 72 signals and waits, and 432 persisted history rows. Wrong-scope clients receive the original HTTP404/typed SDK refusal; 36 real compatible foreign PHP Worker polls are empty before the correct author claims the ready target task. Independent row checks bind each actual task to the correct original run/namespace, one attempt and zero repairs, with exactly one start plus two signal commands per run. Full response bytes, 153 original HTTP sidecars, ordered histories and exact wait/signal/command links are retained. Embedded uses explicit public Two recorder assumptions surfaced in the actual PHP baseline: colon-bearing workflow IDs are percent-encoded in signal paths, and the published SDK throws Frozen Server 2.5.20 / Workflow 2.5.10 / PHP 2.2.7 / Python 2.5.4 / Rust SDK 3.4.4 and resource limits are unchanged. SDK preparation compiles the actor in Rust 1.86. Its artifact 11767819229 independently passes 297 package hashes, 13 actor source hashes and exact archive/VCS identities; those actor hashes match the local reference source. Actual artifact producer The earlier default-profile CI run passed five databases and failed MySQL 8.4. Its original record is preserved and the actual failure cause remains under review. Cleanup also referenced the old pending failed-update manifest after #427's promotion; source This profile remains pending. Native namespaced signal admission/waits/wakeup, the full new six-database matrix, broader namespace/auth, consumers, migration and matched performance are still ahead. No package/image publication, shared dependency adoption, native Nexus capability or release/cutover claim. |
|
Original MySQL 8.4 run 38294049008, job 114961521246, artifact 11767904539 now has an independently audited failed original. PHP and embedded pass all nine failed-update cases. Native passes four, then Python/Python's later good update and changed pending duplicate receive four HTTP 503 The exact missing lookup and cause are still unestablished. This does not establish the older MariaDB failure's cause, storage pressure, a provider/network fault or a published SDK defect. Next add a separate bounded stderr event |
|
The next #429 checkpoint wires the pending paired namespace-signal profile into each existing isolated PHP/native/embedded job for SQLite, PostgreSQL 16/17, MySQL 8.0/8.4 and MariaDB 10.11. It reuses the same compiled SDK artifact, typed database/bootstrap paths, recorder, comparator and task resource cleanup. Completed default signal/query/failed-update reference servers stop before the additional profile starts. The new workload adds nine two-run pairs per target. Only the whole-job ceiling for these expanded jobs changes from 30 to 35 minutes; original run budgets stay 35 seconds, the new pair ceiling stays 70 seconds, and SDK/native/database CPU, memory and swap limits stay unchanged. No earlier failed/cancelled run is retried, relabeled or given more time. This is a bounded qualification allowance, with no benchmark or performance claim. The fixtures remain pending until the actual original artifacts qualify. Lookup-stage diagnostics are separately awaiting runtime evidence; neither subsequent clean checks nor this expanded workload establish repair of the old MySQL/MariaDB failure. |
|
The nine paired namespace fixtures are now wired into all six existing database jobs, with the previously recorded 35-minute whole-job ceiling. Original 35-second run / 70-second pair deadlines, resources and frozen packages stay unchanged; the fixture directory remains pending. First native regression run 38301631466 compiled and passed the other 77 execution tests on SQLite, but failed because wrong-namespace signal refusal omitted PHP's canonical Current candidate |
|
The canonical refusal correction compiled and passed the new metadata assertions. The next original native regression 38303618712 correctly rejected the test's missing typed signal argument contract with Current |
|
Current candidate The actual SDK artifact 11771374917 passes independent uploader/API/download identity, 297 packaged-member hashes, 13 current actor-source hashes and nine retained metadata comparisons. Producer All 4,755 reader tests and the strengthened raw-reference recheck pass. The prepared artifact audit additionally binds the 153 original reference sidecars / 1,051 full HTTP exchanges to both original runs, actual SDK outputs and paired controls. PHP associative metadata projection and recorder-added installed identity are accounted for explicitly; original wire bytes and typed Avro semantics remain unchanged. All local transport, extraction, SDK dependency/binary and audit resources are retired; useful thin/raw proof stays retained. Frozen tuples, original failures, per-run deadlines and runtime caps are unchanged. Broader #325, release, migration, performance and cutover remain open. |
|
Exact Original MariaDB artifact 11773300722 is independently verified: 279,906,602 bytes, SHA-256 The first native namespace case separately reaches Before changing the recorder, the correction contract is: write the control observer journal and request-arrival journal directly into retained fixture evidence, checkpoint setup stage before blocking observer/SDK namespace calls, and consolidate/remove only after a complete final control receipt is saved. A forced child timeout must leave those partial original bytes and the last stage. Capture only bounded stage/namespace/workflow identity and existing sanitized HTTP data, with no credentials. Add always-run native container state/limit receipts so a later stalled request can be distinguished from process loss. No SDK workaround, application SQL/HTTP change, retry/deadline extension or fixture retuple is authorized by this evidence. The native missing receipt still needs a controlled storage reproduction. Original PostgreSQL 17 artifact 11772334653 independently passes 27 namespace pairs / 54 original runs / 306 full HTTP sidecars / 1,866 actual exchanges, all 1,678 member hashes and 78 native execution tests. Its archive is 269,424,334 bytes / SHA-256 |
|
Current A real locked PHP SDK 2.2.7 / source The original five-pass/MariaDB-failure cohort remains unchanged. Missing receipt |
|
The #429 matrix remains held. Original SQLite job 115016752915 in run 38310847080 failed before the namespace profile, in the frozen published PHP patch-replay control. Artifact 11774657948 is retained unchanged: 8,160,199 bytes, SHA-256 cc68f330b8731845b35f3b928674823d52b94c10a88b699f416b9e4373082dcc, all 100 members independently hashed. The actual PHP SDK 2.2.6 / 290fa427a76ffc6ea226bcd59749374aa7a89efa original worker received one 503 on workflow polling: task=null, backend_lock_pressure, SQLite lock_pressure=true, matching namespace/queue/protocol/discovered capabilities and a one-second body retry hint. It emitted worker.retrying, then polled successfully with the same worker, queue and options and a new poll_request_id, as this unmodified PHP Worker actually does. The original activity checkpoint and cold Python replacement completed the same run 01m4s0ryr4s1y72gkm0wm0frkf. The original record remains product-fail because the reader rejected that response as “all other observed I/O succeeds.” Contract decision before correction: extend the existing full-response PHP workflow-poll pressure classifier to the actual PHP original phase, excluding legacy-marker profiles whose original producer is Python/Rust. Keep its exact response inventory, null task, typed SQLite pressure, bounded positive retry hint, original registration/capabilities, actual retry diagnostic, no transport failure, same authority/options and successful recovery in that same phase. PHP may produce its next poll identity; published Python/Rust replacement retries must still preserve the original request identity and positive Retry-After header. Generic storage_unavailable, other endpoints/statuses, missing recovery, changed authority and hidden worker failures remain failures. The probe's shared original 30-second deadline and all runtime caps remain unchanged. Add positive original-phase and adversarial guards, replay the immutable original observation against old and corrected readers, then run the full readers and a fresh exact matrix. A corrected-reader replay is a new check, never a relabel of the original failed job/record. This is a shared-reader correction, with no PHP/SDK/native runtime, dependency or frozen tuple change. The separately diagnosed MariaDB postcommit receipt_read failure remains unexplained and unqualified. |
|
Exact reader correction ddecb91 / tree 3bad9323af5381962f7223bd5b26e3ab73e0def2 is pushed. It extends the existing full-response PHP workflow-poll pressure check to the actual original PHP phase, excluding legacy language producers. No native SQL, PHP/SDK/runtime implementation, tuple, resource cap or deadline changed. The unchanged original SQLite observation passes both the corrected patch checker and full observation checker, while the original reader still rejects “all other observed I/O succeeds.” Four actual-observation corruption guards pass. All 4,817 pure-reader guards pass; the SDK-dependent forced-loss test is skipped in the pure Node runtime and retains its unchanged earlier real locked-SDK qualification. Original artifact/job/record remain failed and byte-identical. The correction contract was posted before implementation at #429 (comment) . All five original 6bac positive namespace artifacts are now independently audited: 135 paired observations / 270 original runs / 1,530 full HTTP sidecars / 9,585 exchanges / 8,352 original member hashes. MariaDB's original postcommit receipt_read missing row remains unexplained. The fixture profile and PR remain pending/held; no six-database promotion or merge claim. Fresh exact Shared 38313701622 and PHP/source 38313701591 are live. Original 24af run 38310847080 is preserved: SQLite failed earlier in the patch reader, MySQL 8.0 passed, other original jobs remain live. There was no cancellation/restart request. General's direct acknowledgment preserves this reader scope and frozen references. |
|
The MariaDB receipt lookup remains unexplained. I am preparing a bounded isolated check of the exact SQLx 0.9.0 connection pool's behavior when a normal BEGIN reply is delayed and its future is dropped. A transparent TCP gate will delay only the downstream reply; SQL remains ordinary BEGIN, with no custom transaction statements. The check must inspect actual transaction state, compare a pooled read with an independently committed row and a fresh connection, and include an ordinary uninterrupted control. It uses disposable data and changes no native or reference runtime, isolation setting, retry rule or acceptance contract. A positive library-level reproduction would establish that mechanism only; it would still need evidence connecting it to the original native failure. The original failed job and missing-row record remain unchanged. No driver or host cause is claimed before this check. |
|
The exact SQLx 0.9.0 / Tokio 1.53.2 MariaDB experiment reproduces a pooled stale read after an ordinary BEGIN reply is held and its future is dropped. The returned session has a physical transaction open, logical is_in_transaction=false and REPEATABLE-READ. A pooled lookup reports RowNotFound for a separately committed row that a fresh connection sees. Explicit rollback restores visibility; an uninterrupted BEGIN/COMMIT control passes. The same experiment with rollback on pool return leaves no open transaction and observes the committed row. This proves the injected mechanism, not the cause of the original #429 CI failure. Correction contract before native changes: return MySQL/MariaDB pooled sessions only after ordinary ROLLBACK successfully clears pending physical transaction state. Existing SQLx pool behavior closes a session when the release hook fails. Bound reset to five seconds, matching the existing acquisition budget; reset timeout returns an error so SQLx closes the session immediately instead of attempting graceful reuse. Preserve isolation, the four-connection pool, acquisition/lock budgets, original durable receipts and all public HTTP/retry/deadline contracts. This adds internal session hygiene and an extra return-time database statement, which must remain visible to later performance qualification. No PHP counterpart or new capability is required for this internal pooling behavior. Add a real MySQL/MariaDB qualification regression through the native pool factory: delay an ordinary BEGIN reply, cancel its future, then require a reused session to observe an independently committed row. Also hold a COMMIT acknowledgment and cancel that future after the row is independently visible; pool reset must preserve the already committed row. The prior pool factory must fail the fresh-read assertion. Run the exact three MySQL-family configurations, existing execution suites and full differential matrix before claiming qualification. Original missing-row and namespace-timeout records stay unchanged and their native cause remains unestablished. |
|
The exact pool correction is pushed at e82d6e2 / tree cc3a6f1830edeb18c9afbc751b0c0aaba7c2992c. MySQL/MariaDB sessions now complete a bounded five-second ROLLBACK before pool return; reset failure disposes the session. The existing four-slot pool, isolation, acquisition/lock budgets and public receipt/retry/deadline contracts are unchanged. The extra return statement still needs later performance measurement. Exact SQLx 0.9.0/Tokio 1.53.2 isolated MariaDB controls reproduce physical transaction=1 with logical transaction=false and a stale pooled receipt read after losing a held ordinary BEGIN reply. Reset gives physical transaction=0 and fresh visibility. The native-factory regression passes with reset and fails RowNotFound without it. It also preserves rows committed before a lost COMMIT reply, and proves that an unreleased reply exhausts the reset budget, discards that connection and frees a slot for a different connection. All nine intended real MariaDB qualification cases pass, including catalog/refusal, concurrent bootstrap, real interruption, typed data/read-only roles and the new pool regression. Rust 1.99 formatting, diff and workflow resource isolation checks pass. The local 78-case HTTP execution suite is still running; full exact MySQL 8.0/8.4/MariaDB and six-database differential checks remain pending. The CI selector now includes the new regression while excluding the parent-driven bootstrap helper from direct invocation. Two existing native test files receive formatting only. The contract was recorded before implementation at #429 (comment) . This proves and corrects the injected pooling mechanism; it does not establish the cause of any historical CI 503 or namespace timeout. All original records remain unchanged. The PR/profile remain pending, with no dependency/SDK tuple, capability, image/package, migration, matched performance or release/cutover claim. |
|
Exact e82d6e2 / tree cc3a6f18 Shared 38316178557 fails all three MySQL-family jobs. Original tiny artifacts 11777041158 (MySQL 8.0), 11777435829 (MySQL 8.4) and 11777011231 (MariaDB) are retained unchanged and all 240 member hashes verified. Each passes eight existing database qualification cases and fails the new pool test at its final seven-second slot-disposal observation. Earlier fresh-read and lost-COMMIT assertions in that test pass. SQLite and PostgreSQL 16/17 jobs pass; PHP/source and boundaries pass. The profile and PR remain held. The same source locally passes all nine MariaDB qualification cases, 78 HTTP execution cases and 33 ordinary library cases. The local database used plain TCP; the CI databases negotiate TLS. That difference needs an exact TLS reproduction, not a relaxed assertion or disabled TLS. The pinned driver source shows that pool release errors call MySQL close_hard, which still awaits stream shutdown. Pool minimum size defaults to zero; automatic minimum-size replacement has not been established as the reason for the failed observation. Next retain a bounded exact-source TLS probe with reset-entry/result and transport-close evidence to distinguish failure to reset, shutdown waiting, and a gate that held the wrong response. Keep the original five-second reset, seven-second observation and fifteen-second test budgets. Any change to the test or runtime follows that evidence. No historical CI cause, PHP/SDK defect, host cause, tuple change or release claim is inferred. |
|
The exact TLS 1.3 probe reproduces the final CI disposal failure. The reset hook logs timeout after its original five-second budget, but the transport does not close and the pool retains the slot past the original seven-second observation. Explicitly flushing before the held BEGIN produces the same failure. Both original probe logs remain failed; the one-shot command retired its database and network. Pinned SQLx 0.9.0 defines close_hard as immediate connection closure without graceful shutdown. Its MySQL implementation nevertheless awaits stream shutdown, and the Rustls transport performs asynchronous TLS I/O on that path. The current official sparse index has no later stable sqlx-mysql release; the current upstream MySQL close_hard code still has this await. Next control, before product changes: use the verified official sqlx-mysql 0.9.0 source archive (SHA-256 90b8020fe17c5f2c245bfa2505d7ef59c5604839527c740266ad2214acebea27, VCS 003b698e99e024f3621b8043a2426fde5b741171), changing only close_hard to drop its owned connection immediately. Keep graceful close, Rustls, certificate/hostname checks, reset and acquisition budgets, pool size, durable receipts, all original assertions and public contracts unchanged. Re-run the same real TLS test and require transport closure, freed capacity, a different replacement connection and the previously committed row. This is ordinary socket ownership cleanup, with no replacement protocol implementation or TLS downgrade. Only after that control passes, adopt the minimal native dependency patch with explicit archive/VCS provenance, license files and exact source-difference verification, then qualify all three MySQL families, both PostgreSQL versions, SQLite and existing TLS trust/hostname cases. No PHP/Composer/shared SDK tuple adoption is involved. Historical native receipt failures remain separate and unexplained; no release/cutover claim. |
|
Exact native driver correction 64ca245 / tree 70eed06bfdea1048d8cbf96050947d09cd258a9e is pushed. It retains the five-second session reset and changes only the pinned MySQL driver's hard-close operation to immediate owned transport disposal. Graceful close, Rustls, certificate/hostname checks, pool/acquisition/lock budgets, original test assertions and all public contracts remain unchanged. Original exact TLS 1.3 diagnostics reproduce the CI failure: reset times out, but socket closure and pool-slot release do not happen within the original seven-second observation. The flush-before-gate control fails identically. With only the driver hard-close correction, both controls pass (5.52 / 5.50 seconds), showing transport closure, freed capacity, a different replacement connection and the prior committed row. Final repository source, without diagnostic overlays, passes the unchanged real TLS regression twice (5.56 / 5.54 seconds). All 33 ordinary library tests pass; the 18 backend qualifications are intentionally ignored in that ordinary invocation and remain required in their actual matrix jobs. Rust 1.99 format, locked offline compilation and diff checks pass. All 78 official upstream crate members are checked against archive SHA-256 90b8020fe17c5f2c245bfa2505d7ef59c5604839527c740266ad2214acebea27 / VCS 003b698e99e024f3621b8043a2426fde5b741171. Only src/connection/mod.rs changes code. Full upstream MIT/Apache texts replace packaged license references, and DURABLE-PATCH.md records the scoped override. The native lockfile changes only that package's source; no Composer, shared package or frozen SDK tuple changes. The correction contract precedes adoption: #429 (comment) . Fresh exact six-database, source and TLS trust/hostname qualification remain pending. Original e82 MySQL-family jobs and both failed TLS probes remain failed; the historical receipt/namespace failure causes stay unestablished. The profile/PR remain held. No package/image release, migration, matched performance or cutover claim. |
|
Exact candidate Independent original-archive review across SQLite, PostgreSQL 16/17, MySQL 8.0/8.4 and MariaDB 10.11 verifies all 10,053 member hashes, 162 paired namespace observations, 324 original runs, 1,836 complete HTTP sidecars and 11,421 wire exchanges. All nine PHP/Python/Rust client/author directions and canonical embedded counterparts preserve the two namespace values, original commands/runs, cold author replay, foreign signal refusal and empty foreign-Worker claims. Each configuration also passes 78 native execution tests with zero ignored tests. All three MySQL families pass the nine schema/pool qualifications, including the unchanged cancelled-BEGIN, committed-row/lost-COMMIT and bounded TLS disposal regression. Pressure, role/fencing, real node recovery, PHP takeover refusal and certificate trust/hostname gates pass. PHP/source passes 2,744 tests / 61,718 assertions, with six existing deprecations and one skip. All three original SDK preparation artifacts independently match their frozen installation/source/archive identities, packaged hashes and actor source. The pinned Rust direction worker hash remains unchanged. Boundary checks pass. This qualifies the candidate paired namespace slice and scoped native pool/driver correction. The fixture promotion and final reviewed head/merge/fresh-main gates are next. The original failed jobs and observations remain unchanged; the injected cancellation and TLS mechanisms do not establish the causes of the historical receipt or namespace failures. No Composer/shared package adoption, frozen retuple, native Nexus claim, image/package release, matched performance, complete parity, migration or cutover is established. |
rmcdaniel
changed the title
|
Final review checkpoint: current head 8250d90 is up to date with PHP main. The ancestry integration matched the previously reviewed CI merge tree exactly and changed no Composer identity. Review caught a stale variable-based namespace cleanup path after promotion; the one-line correction now resolves the qualified profile. A read-only audit exercises all 24 execution/cleanup profile/backend selectors and reproduces the old missing-manifest path, without invoking Docker cleanup. Fixture bytes, native runtime and frozen actor identities remain unchanged. Exact-head Shared 38336076237 / PHP 38336076386 are queued; both boundaries pass. Original 621/a8 runs stay intact and their actual outcomes remain pending. The fully audited original six-database qualification is recorded in comment 6136523527; final-head qualification, review/merge and fresh-main remain ahead. |
|
The original intermediate promotion run 38335399623 has a failed SQLite comparison. Its original artifact 11785235867 is independently verified at 7,795,400 bytes / SHA-256 The failure occurs in the frozen published PHP 2.5.14 manual-schedule lifecycle fixture, before the new namespace profile runs. The original access log identifies No trigger retry, workflow restart, workflow cancellation or frozen tuple change was introduced. This PHP reference finding is being handed to General. The current exact head remains |
|
Final-head SDK artifact qualification passes for exact Independent original artifact review verifies:
Each archive matches its original API size/digest, original job log and exact runner revision. Installed registry/package/source identities match their existing frozen manifests; all three compiled Rust Workers are byte-identical to the previously qualified cohort. No SDK, Composer or shared dependency adoption occurred. Final PHP/source checks also pass: 2,744 tests / 61,714 assertions, six existing deprecations and one skip. The source-qualification job verifies the successful repository-contract and feature jobs; it is not a separate runtime cohort. Boundary checks pass. The SDK review qualifies artifact identity only. The final differential matrix remains live, so merge and fresh-main qualification remain held. General acknowledged ownership of the original PHP schedule failure; its original outcome and unresolved cause stay intact. The next separate run-targeted signal coverage proposal is recorded in #325, followed by one published PHP/PHP reference observation. The new shared/embedded/native profile remains unimplemented and unqualified. |
|
The exact final-head PostgreSQL 17 job 115099907699 is failed and preserved. Its original artifact 11786171296 is independently verified at 211,405,986 bytes / SHA-256 The main comparison step failed at the namespace profile's forced setup-loss guard, before namespace runtime execution. All 48 pure namespace reader cases pass; the runtime test fails at its assertion that the held-request test server has received the original GET. The test had stopped waiting once the proxy's arrival journal became nonempty. The proxy records that journal before asynchronous outbound delivery, so journal visibility does not guarantee that the server request callback has run. The earlier assumption that this was a pressure-phase failure was incorrect; that phase runs afterward independently. Correction boundary before editing: await both the retained arrival and the original server request before forced child loss. Preserve the existing 10-second observation / 15-second test ceilings, all original stage/request/deadline/credential/absence assertions, real locked SDK and byte-preserving proxy. This is a test synchronization correction; no recorder, native runtime, fixture bytes, SDK tuple, retry or workflow budget change. Qualify corrected and prior guards under a controlled forwarding delay, then run the exact locked-SDK test normally. The original job remains failed; no rerun or causal attribution to historical native/database failures is authorized by this record. |
|
The retention test synchronization correction is committed/pushed at The guard now waits for both the locally retained arrival journal and the held-request server's actual request callback before forced child loss. With a controlled 250-ms forwarding delay, the prior guard fails at the same assertion as the original PostgreSQL 17 job; the corrected guard passes all original assertions. Three ordinary runs with the real locked PHP SDK each pass all 49 selected tests, zero skips. The original 10-/15-second test ceilings, fixture/run budgets, recorder, native runtime, SDK/Composer dependencies and frozen actors remain unchanged. The port log also records the independently inspected selected-run PHP reference and the Workflow 2.5.13 / pending Rust SDK 3.5.2 coordination. These are future separate profiles, not dependency adoption or native qualification. Fresh original checks: Shared 38339429192, PHP/source 38339429196, and boundary 38339429200; boundary 38339423364 passes. PR remains draft/held. Prior |
|
Independent review now confirms that the original Original artifact identities, checked against their job/API/upload receipts and every ZIP member:
The new guard and counterfactual preserve the existing timeout and assertions. The original three jobs remain failed; no rerun, cancellation, tuple change or historical database-cause attribution occurred. Temporary extraction is removed after retaining proof. Exact |
|
Exact All original API size/digest/upload-log/runner/source/installed identities match:
All three compiled Rust Workers are byte-identical to their previously qualified frozen cohorts. No SDK/package/Composer/shared dependency adoption occurred. The exact-head PHP/source suite now passes: 2,744 tests /61,714 assertions, six existing deprecations and one skip, with the original feature log retained. Both boundaries pass. This is SDK/source qualification only: all six database jobs remain live. Merge and fresh-main stay held; every original failed job retains its outcome. |
|
Final-head candidate qualification is complete at All six original jobs in Shared run 38339429192 pass without reruns: SQLite, MySQL 8.0/8.4, MariaDB 10.11 and PostgreSQL 16/17. Independent original-artifact review verifies API size/digest, matching job/upload identity, every archive member and the actual recorder output. Totals are 162 paired observations, 324 original runs, 1,836 complete HTTP sidecars, 11,583 byte-checked wire exchanges and 10,053 member hashes.
Each configuration passes 78 actual native execution tests and all 49 namespace reader/retention cases without skips. The three MySQL families each pass nine schema/pool qualifications, including cancelled BEGIN, lost COMMIT reply and bounded TLS disposal. Both PostgreSQL families pass six schema qualifications. Role-token, worker-fencing and declared database-pressure steps pass; non-SQLite configurations also pass actual trusted-certificate and wrong-trust/hostname refusal checks. PHP/source 38339429196 and both exact-head boundary runs pass. Its original feature log confirms 2,744 tests / 61,714 assertions, six existing deprecations and one skip. The same-head SDK audit verifies 892 archive members, 23 actor-source hashes, 25 thin comparisons and three unchanged compiled Workers. Static review is complete. Runtime, scoped SQLx patch and SDK actors remain byte-identical to the earlier qualified candidate; Composer identities are unchanged. The setup-loss regression's receiver/journal synchronization preserves the original limits and assertions. All three failed originals and the controlled counterfactual stay retained. Other historical failures, including the General-owned frozen PHP trigger fault, remain unchanged and do not acquire a causal explanation from this qualification. Merged at All local audit containers and extraction scratch are removed after retaining original archives, hashes, logs, scripts and receipts. This receipt does not establish fresh-main, publication, native Nexus, full parity, migration, matched performance or cutover. |
|
The automatic main SDK producer in Shared 38343065563 passes on exact merged Independent original-artifact review verifies all 892 archive member hashes, packaged manifests, 23 actor-source hashes, 25 thin byte comparisons, registry package references/archive hashes, original job/upload receipts and the three unchanged compiled Workers. Main Rust remains 3.4.4, legacy 3.4.3 and repeated-patch 3.4.2; existing Python/engine/Server cohorts remain frozen.
Thin identities, member hashes, original API/job/source records and audit scripts are retained. The 32 MiB temporary SDK archive directory and audit containers are removed at 03:49 UTC before the 04:00 cleanup boundary. No dependencies or compiler caches remain from this audit. Fresh-main PHP/source 38343065608 now passes. Original feature job 115118404535 confirms exact main producer The merged candidate receipt remains separate from this main SDK/source proof. This does not establish fresh-main database qualification, publication, migration, matched performance or cutover. Subsequent checkpoint: all six original main database jobs and independent artifact audits are complete. This SDK/source receipt retains its narrower proving scope; the linked delivery record establishes the full bounded main checkpoint. |
|
#429's exact fresh-main checkpoint is complete at All six original jobs in automatic main Shared 38343065563 pass without reruns. Independent original-artifact audits bind each API size/digest to its actual job/upload and producer, hash every member, run the strict three-runtime reader and compare complete original wire bytes. Totals are 162 paired observations, 324 original runs, 1,836 full HTTP sidecars, 11,567 byte-checked exchanges and 10,053 archive member hashes.
Every configuration passes 78 native execution tests and 49 namespace reader/retention cases without skips. The three MySQL families each pass nine schema/pool qualifiers, including cancelled BEGIN, lost COMMIT reply and bounded TLS disposal; both PostgreSQL families pass six schema qualifiers. Original role-token, worker-fencing, declared database-pressure and applicable trusted/wrong-trust/hostname TLS steps pass. Main PHP/source 38343065608 passes. Its original feature log confirms 2,744 tests / 61,718 assertions, six existing deprecations and one skip; source gate 115120582316 passes. The optional candidate-only source job is skipped and is not claimed as full cohort evidence. Boundary 38343065642 passes. The independent main SDK audit verifies 892 archive members, 23 actor hashes, 25 thin comparisons and all three unchanged compiled Workers. The completed remote head remains deleted/404. The clean source worktree and local branch are removed after proving they are reachable from the identical qualified main tree. All audit containers and extraction scratch are removed before the recorded cleanup boundary; original compressed archives, identities, logs, scripts, hashes and receipts remain retained. Composer and frozen actor/package tuples are unchanged. Original setup-loss failures and the controlled correction remain intact, alongside the older cancelled/failed PHP, database and omitted-input records. The injected reset/TLS proof does not establish their historical causes. This delivers the bounded paired-namespace signal and cancelled-session correction, not full namespace/auth or Rust Server acceptance. Next under #325: the separate selected-run signal family with real durable rejection records, then corrected marker/range-repair integration using its exact qualified published consumer tuple. General retains PHP dependency/expiry/query contracts; SDK owners retain their implementations. Full capability/consumer parity, external payloads, migration, multi-node/fairness, matched performance, operations/images and release/cutover remain open. No native Nexus, publication or full-goal claim follows from this checkpoint. |
2 participants
SDK signal coverage previously exercised only the default namespace. This change adds nine paired PHP/Python/Rust client/author fixtures that park two original runs in distinct namespaces with the same type and queue. Foreign signals receive canonical 404 refusals and compatible foreign Workers poll empty; cold authors complete both original runs with their own exact Avro values. Native current-run signal lookup and untimed condition waits support this boundary. The shared recorder retains setup and HTTP journals through forced child loss.
A cancelled BEGIN can leave a physical MySQL/MariaDB transaction on a returned pooled connection. The native pool now rolls back before reuse within five seconds and disposes failures. Exact TLS probes also exposed SQLx 0.9.0 hard close awaiting unavailable TLS I/O; the scoped native driver override drops the owned transport immediately. Official source provenance and licenses are retained, with one upstream source-code file changed. Graceful close, certificate validation, pool size and acquisition/lock budgets remain fixed. The extra return-time statement needs later performance measurement.
The retention regression now waits for both its retained arrival journal and the actual receiver callback within the original 10/15-second limits. A controlled 250 ms forwarding delay reproduces the prior assertion failure and passes the corrected guard; three ordinary locked-SDK runs pass all 49 cases without skips. Recorder behavior, fixture budgets and original failed records remain unchanged. Regression evidence.
Final head
7faa58d018eb063759766cbcca4ba4ec8553fd53incorporates current PHP main, promotes byte-identical fixtures and updates execution/cleanup paths. Its actual CI producera4115e22a8fc9e72106b402ea932bdb91b772722has the identical tree62736363c1e03ceae091ea578d307fb4d2370054. Composer identities are unchanged. Runtime, vendored driver and SDK actor bytes match the pre-promotion candidate.Validation:
6299289185caea472b2738a5d4442a617edf0b8ahas the identical reviewed tree. Its automatic fresh-main qualification also passes all six configurations, PHP/source, SDK and boundary; independent main audits verify 162 pairs, 1,836 sidecars, 11,567 exchanges and 10,053 member hashes. Completed branches, worktree and audit resources are removed.Frozen tuples and earlier failed originals remain intact. The injected session/TLS regressions do not establish the causes of historical database failures. This PR does not qualify full parity, migration, matched performance or cutover, and makes no package/image publication or native Nexus capability claim.