Skip to content

Support poll hints and bound activity completion uploads - #458

Merged
rmcdaniel merged 31 commits into
mainfrom
rust-325-activity-completion-uploads
Oct 15, 2026
Merged

rmcdaniel merged 31 commits into
mainfrom
rust-325-activity-completion-uploads

Conversation

@rmcdaniel

@rmcdaniel rmcdaniel commented Oct 14, 2026 •

Copy link
Copy Markdown
Member

Opted-in workers receive a null-task wake hint when the opposite task kind changes on the same namespace and queue. Activity workers can upload version-1 local result bytes bound to their original live task, attempt and owner, then reconcile a content-verified matching upload after closure. Neither operation grants completion authority.

Upload validation runs before body/provider work and after transfer, preserves the original lease, requires exact stored identities and PHP-compatible Unicode/context hashing, and enforces durable slot/byte allowances. Unsupported workflow/query/prepared contexts stay refused. Full external transport discovery remains false. PHP query middleware, OpenAPI 55 and frozen consumers are preserved. This includes #455.

Candidate 3409a776ef4f7fe7864b22b5d75c70cbb43fe29b has tree e213e9b1928621cd0d3b3b24c97983a3db3ce51e, identical to tested merge 3dd7b2f9bc1dd8d4e84958996efbeae6cde1c34b. Applicable normal source gates, all six execution database cells and six codec/deadline health cells pass in 38686850381. Attempt 2 reran only MySQL 8.0 with unchanged source, tuple and limits; the other cells retain their original executions. Each execution artifact reports 113 native tests and 25 comparisons passing. The original MySQL 8.0 query transport failure remains failed and unexplained. Later PHP source integration and fresh combined-tree qualification remain ahead; no main promotion is claimed.

Local controls cover native unit/SQLite HTTP behavior, MySQL 8.4/PostgreSQL 17 authority races, and 36 actual PHP/native originals with Unicode owners, typed values, verified bytes and cold-parent completion. Failed original controls remain failed. Automatic policy, S3, matched performance, migration, native release and cutover remain outside this change. Refs #325.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The raw-HTTP published PHP 2.6.2/Core 2.5.14 control passes a genuine original, 01m4ydwtcw0ngxa0cnkenqgcnc, within its original 30-second budget (1.44 seconds observed). Its version-1 activity-result completion context uploads 44 exact serialized Avro bytes, preserves the reference on an identical retry, rejects changed slot bytes and a foreign owner, reconciles the matching upload after activity closure, and completes the same cold parent with exact int64/null/Unicode values. A stopped-database copy confirms one completed activity, one ActivityCompleted event, three tasks and no repeated claim. The original database hash is unchanged by the copy audit. This is a public HTTP control, not an installed SDK-worker claim.

The unchanged native tree 6903b01a explicitly refuses the same context with 422/development_payload_completion_upload_unqualified after a genuine activity claim, run 01M4YDYQZBBSKFBFM2ZQRV37W7. This uses the existing official CI release artifact from 38647001027/11933721192, verified against its manifest, native tree and lock; executable SHA-256 6bff4423e6e3e5c0f95502c5a8fce0da6f7af23694f2b22379667db7705ecb3b. No new compile was needed.

Two earlier PHP originals are unsuccessful probe runs, not product defects: the first used an incorrect poll path (run 01m4ydtwdm3xfxkv7k28e5r5nf); the second submitted decoded binary bytes where this transport requires the serialized Avro text, and completion correctly refused that payload. The corrected third original does not rescue them. Initial malformed registration preparations preceded business admission. Probe source 253af8a preserves the corrected wire recipe.

Both application containers stop with explicit SIGTERM, exit 0 and no OOM. Containers/network are removed at 00:06:14 UTC, before the 00:15 boundary; the downloaded executable/archive, databases/provider bytes and raw outputs are deleted after this audit. Native implementation, allowance/expiry/race controls, all database families and broader completion contexts remain pending. Full native transport discovery, performance and release remain unqualified.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Native prototype 594eb45 is not a qualified full-suite result: its first full execution run has 112 passing cases and one failed assertion. The unchanged old byte-transport test supplied the malformed header {} and expected the former blanket development_payload_completion_upload_unqualified reason. The implemented parser now returns the published external_payload_completion_invalid reason for that malformed context. PHP's existing Context::parse and the external transport OpenAPI already define that 422 disposition.

Before updating the assertion, the intended boundary is explicit: malformed activity completion contexts use the existing canonical invalid-context diagnostic; a valid unsupported workflow/query/prepared context remains explicitly unqualified. Neither reads provider bytes or grants task authority. I will keep the malformed-context check and add a separate valid-but-unsupported context check. No frozen fixture or PHP contract is being relaxed, and the initial full-suite result remains failed.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The new canonical-ID control exposes a real defect in my unpublished prototype on actual MySQL 8.4.5: a case-variant task ID reaches the same activity through the database collation, but hashes into a different upload allowance. Its upload incorrectly returns 201 where canonical ownership requires 409. The regression reaches a genuine activity claim and receives reference ep_01M4YFB3YBA08T244Q6ZJE7EDJ. This is a new prototype failure, not a defect attributed to published PHP or an older frozen cohort. The earlier plaintext-auth setup was blocked before business admission; the actual counterexample used TLS-required MySQL.

The correction will require exact stored task/attempt IDs before body/provider work, matching PHP's existing strict task/attempt status bindings. It will preserve the submitted identity and refuse aliases, rather than silently rewriting a context into another allowance. Existing 113-case SQLite acceptance does not qualify that MySQL boundary. The failed original remains failed; fresh corrected backend controls and full regression are required.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Actual unpublished native UTF-8 counterfactual: a valid v1 activity completion context with lease_owner="foreign/λ😀" returns HTTP 422 / external_payload_unsupported before ownership evaluation. Expected established behavior is HTTP 409 / external_payload_completion_lease_rejected. The focused real HTTP test fails 0/1 on the existing a269 implementation plus this new test. The shared header accessor uses HeaderValue::to_str(), which rejects non-visible-ASCII bytes; the published completion header permits JSON up to 4096 UTF-8 bytes. Fix and actual published PHP/native Unicode consumer qualification follow separately. This failure remains a failed original; the prior ASCII-only 12-original consumer result does not qualify the forthcoming Unicode revision.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The UTF-8 header counterfactual remains failed as recorded above. After its correction, the first complete local suite is 112/113: its cold-reader helper still sent the old ASCII worker ID after this PR changed the registered fixture worker to Unicode. Commit 20cadd2 corrects only that helper's original identity; this is a test preparation error, not a runtime correction.

The complete 20cadd2 SQLite run is also failed, 112/113. All three new completion-upload cases pass, but the existing stalled_upload_keeps_original_timeout_and_releases_capacity_for_next_original control returns HTTP503 instead of201 on its second tiny upload under the unchanged 100ms operation timeout. Its stalled first request correctly refuses. This is the same existing short-timeout control previously observed failing in #450, but no common cause is established. Source, assertions, timeout and resources remain unchanged. One new unchanged-head complete run will follow the focused MySQL/PostgreSQL completion controls; any success is a separate execution and will not rescue either original.

The actual revised SQLite consumer cohort independently passes all twelve originals and ten reader refusals, including genuine unescaped UTF-8 headers on both published PHP and native, with successful explicit SIGTERM shutdown and deletion. Other local backend consumer outcomes and hosted whole-head qualification remain pending.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The first revised local MySQL consumer profile on 20cadd2 fails in the published-PHP sixth original 01m4ygqhbkvd03g7v3tkz3ssg7: first and duplicate upload each return201, then the raw probe's strict PHP-array comparison refuses their references. The prior five PHP originals complete; native consumer originals are not admitted after this failure. Both applications shut down0/noOOM, app databases/provider bytes are removed, and cleanup passes.

The assertion did not retain the two returned references, so this original does not establish lost identity or a Server defect. The current probe uses PHP === on JSON objects; MySQL JSON hydration may change object member order. Prospectively the probe will compare exactly the five reference members with strict value/type equality while allowing JSON object member order, retaining both returned references and whether member order changed. The independent reader will verify the same exact identities/types. This changes no meaningful reference field, original workflow/observation deadline, fixture expectation or runtime source. A fresh MySQL original must demonstrate the disposition; the original failed observation remains failed.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Local acceptance at recorder 29f5ba842d5605de4b5c4979591b76e4f5aa9b83 / tree e399bb635f93f40eb9460ea59fd62de33bb406a9: separate published PHP/native SQLite, MySQL8.4.5 and PostgreSQL17.11 each PASS twelve originals and twelve independent reader refusals (36 originals total). Every original keeps its 30-second bound. Each pair's actual completion-upload originals are:

Backend Published PHP Native
SQLite 01m4yh5ja0kc472v58zeppyf53 01M4YH69ZVCEB4ZQ1GAY6KTYZ6
MySQL8.4 01m4yh1cwmrdgat4jhzpdmjqkz 01M4YH25SKKR7V7F50TYGGWY5N
PostgreSQL17 01m4yh3n9ypd2xeaatweqxmab8 01M4YH4CJGWG6CAPFSWX6NTQF7

All six actually send unescaped UTF-8 completion JSON and preserve exact original task/attempt/owner, 44-byte Avro content/digest, typed int64/null/Unicode result, unchanged lease, matching duplicate/closed upload reference and one cold-parent completion. Published PHP/MySQL changes JSON member order on the duplicate; every field and type remains identical. The new reader refuses a changed duplicate ID and stringified size. The earlier strict-array observation stays failed and does not establish a PHP Server defect.

Native component d22375a69b03ba7bd497b099c60300b15c4f4018, binary SHA256 c9d6bb893d75b9b1d3828ee5b0708d7ff446eb14fee9ff93ab8921cc9f1a37c7: 70 library tests pass/18 explicit ignores; unchanged complete SQLite repeat113/113; focused real HTTP completion authority/race cases3/3 on MySQL and3/3 on PostgreSQL. The prior unchanged100ms control failure remains failed/unknown; the separate repeat does not rescue it. MySQL's functional cases use required TLS, not a certificate-trust qualification claim.

All six consumer applications exit0/noOOM with explicit SIGTERM, all three scoped cleanup steps pass, application databases/provider bytes are immediately removed. Database servers/compiler/dependencies/raw scratch are being retired before the agent's01:05 bound. Hosted final-head all-six/source acceptance and promotion remain pending. No workflow/query/prepared context, pressure bypass, S3, provider restart/backup/cleanup-fault, advertised full transport, release/cutover or performance claim. Frozen tuples remain unchanged.

@rmcdaniel rmcdaniel changed the title Qualify activity completion uploads against the published wire contract Support poll hints and bound activity completion uploads Oct 15, 2026
@rmcdaniel

Copy link
Copy Markdown
Member Author

Current combined main candidate is 93368d2 / whole e3629e43c4e8406f0d2a2a24afea8a4b87f82768 / native d22375a69b03ba7bd497b099c60300b15c4f4018. It integrates delivered main303 and updates the port guide; runtime is unchanged from 483/790. Current full qualification is 38670267284, with normal source/boundary checks pending. The superseded 790 shared run was cancelled during SDK setup; dependent database placeholders executed zero steps. Admitted 1fa/483 matrices continue unchanged. This stack delivers parent #455 and upload work through one current-main candidate, then requires independent fresh-main acceptance. No local runtime is live and no frozen workload, protection, capability or release claim changes.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The earlier admitted 483a1ca cohort is now fully green: shared 38663459733 all13/all6, normal feature38663459681 and both boundary checks pass. All six execution logs independently read back 45 passing records with no test-error markers, including cold external-consumer reads, restart/worker fencing and corrected marker/range/clock/expiry cases. This qualifies that original candidate. Current933 has identical native subtree d22375a6 but a changed whole tree and remains separately pending in38670267284; promotion and fresh-main acceptance are not carried over. Earlier failed originals and the limited activity-upload/discovery scope remain unchanged.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Exact combined candidate 93368d2 / whole tree e3629e43c4e8406f0d2a2a24afea8a4b87f82768 now passes all 13 shared jobs in 38670267284, plus every applicable normal feature/source/boundary check. All six decoded database logs have been independently read: each contains 45 passing scenario outcomes and no failed/error records. Conditional cancellation/prepared-local source qualification is explicitly skipped. Native subtree d22375a69b03ba7bd497b099c60300b15c4f4018.

This qualifies this candidate, including its #455 ancestry; it is not merged or fresh-main acceptance. Main has advanced through PHP documentation #461 and a coordinated PHP-only Core 2.5.21 / Server 2.8.1 window is active. I will integrate that exact fully qualified handoff once, then qualify the actual combined tree before promotion. Existing failed originals, frozen actors and capability/release limits remain unchanged. No source/timeout/fixture revision is made by this checkpoint.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Qualified PHP main ed803519000fac6f5b9a5039387e841293f8ecc7 is integrated at 3a2682000ff77674846b7d308f0548330c755f76, whole tree f5a6e13b1bbc32687fa4f7f1b505c5c8885ddbb4. The exact parents are the previously qualified 93368d2ae component and ed803519; native/shared/runner paths are unchanged from 933.

Fresh all-13/six-database qualification 38680689284 and normal source checks are running. Earlier candidate passes do not establish this combined-tree result. Promotion and its separate fresh-main checks remain pending. General retains PHP 2.8.1 publication/installed consumers; no image, native capability, frozen tuple or release claim follows from the merge.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Fresh combined head 3a26820 / run 38680689284 now has a failed PostgreSQL 17 job (116247159980). Its decoded original log proves the failure is the existing fresh capacity upload still using the preceding stalled request 100 ms test policy: expected 201, actual 503/external_payload_unavailable at external_payloads.rs:1379. The original stalled request timeout remains enforced. No current PHP cause is established. The subsequent corrected-marker call reaches a native target that its failed prerequisite never created and gets cURL 52; this is dependent absence, not evidence of a native crash. I am bringing forward the already prepared #460 fixture separation: unchanged 100 ms stalled original, independent normal-policy fresh upload sharing the exact same one-slot semaphore, plus prerequisite gating for dependent recovered-target checks. No runtime policy or resource budget changes. The admitted other cells keep their original deadlines; I will not push a new revision until those originals terminate. This failed head stays failed and later all-six checks must execute afresh.

@rmcdaniel

Copy link
Copy Markdown
Member Author

The original combined run 38680689284 is terminal: twelve jobs pass; PostgreSQL 17 fails the separate fresh-upload capacity assertion while still using the preceding stalled request's 100 ms policy. Five database cells pass all 113 execution tests; PostgreSQL 17 has 112 passes and that one failure. All six small execution artifacts were independently read. Their synthetic revision 32d24f8d has the exact 3a268200 tree f5a6e13b; later missing-target readers are dependent preparation absence, not evidence of a native crash. The original remains failed.

New candidate 3409a776ef4f7fe7864b22b5d75c70cbb43fe29b, whole tree e213e9b1928621cd0d3b3b24c97983a3db3ce51e, preserves the stalled original deadline and shared semaphore. Only the independent fresh request uses ordinary policy. Recovered-target readers now require successful target preparation. It also incorporates General's independently qualified PHP #436 source 35cd3642; its PHP contract tests and runtime bytes are preserved, with native empty-declaration parity still a separate open gate.

This is a new exact-source qualification, with no passes carried across heads or promotion/release claim. #463's already-admitted originals continue unchanged. Audited local ZIPs and all completed local runtime/build/raw resources have been deleted.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Original3409a776e shared38686850381 is terminal: twelve jobs PASS, MySQL8.0 execution cell FAILS. Its small original artifact11951413442 was independently read. Native113 execution tests pass; all frozen PHP and embedded nine-direction query observations pass, but the frozen native query recorder fails status0 versus required200 and sdk-queries-profile.status=1. This is a transport observation with no established cause; the thin artifact does not retain the original raw request identity. It is not a timer/upload, PHP/SDK or infrastructure diagnosis. All six codec/deadline health cells pass. Earlier failed records stay failed. One bounded exact same-head/tuple/limits MySQL8.0 job retry is requested now; the other five original database cells and health checks are not rerun, and this must not be described as a fresh all-six execution. No source, expectation or budget change.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Independent artifact audit for 3409a776ef4f7fe7864b22b5d75c70cbb43fe29b: the five completed SQLite/MariaDB/MySQL 8.4/PostgreSQL 16/PostgreSQL 17 artifacts each report 25 passing comparisons, 19 zero status markers and all 113 native execution tests passing. Their recorded revision is merge 3dd7b2f9bc1dd8d4e84958996efbeae6cde1c34b, with exact parents 35cd3642 / 3409a776 and tree e213e9b1928621cd0d3b3b24c97983a3db3ce51e, identical to the candidate tree. Separate database storage controls also pass. These are bounded summary audits, not reconstruction of deleted raw journals.

The original MySQL 8.0 query transport observation remains failed/unknown. Its one unchanged-head, unchanged-budget MySQL-only retry remains live in run 38686850381 attempt 2. The other jobs were not re-executed; this is not a fresh all-six run. No promotion or combined PHP 2.8.2 qualification is claimed. Downloaded audit ZIPs and disposable summaries are removed immediately after this audit.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Run 38686850381 attempt 2 is complete. Only the MySQL 8.0 execution job was rerun. Its artifact 11955197774 records all 113 native execution tests, 25 comparisons and 19 status markers passing. The other five execution jobs and six health jobs retain their original successful executions; they were not rerun.

All six execution artifacts identify tested merge 3dd7b2f9bc1dd8d4e84958996efbeae6cde1c34b, parents 35cd3642 / 3409a776, and tree e213e9b1928621cd0d3b3b24c97983a3db3ce51e, identical to candidate 3409a776ef4f7fe7864b22b5d75c70cbb43fe29b. The six independently inspected health artifacts also pass their database audit and three deliberate corruption refusals each, with 14 first-attempt tasks and zero repairs per cell. Their recorded native subtree is 7d0287cd8255e9a270010c2ab50c4cb21467f995 and binary SHA-256 is 6cedb86226a0d7d2b2552a8e345f4febec9277586bbbf37c9f1517a848bc7f3b. Their process JSON files are live snapshots, not graceful-shutdown evidence.

The original MySQL 8.0 native query transport failure remains failed, with cause unknown. This outcome qualifies this candidate component across the six databases. It does not qualify a new combined tree, current PHP main, migration, performance, release or cutover. Later PHP source integration and fresh combined checks remain ahead. Downloaded audit ZIPs and temporary summaries are deleted after this inspection; GitHub retains the hosted outcomes.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Fresh combined source is now 688ee3ea02a26501abf9fb3d506624f17e79a7c2, whole tree 3be0ab688a0ce47f15166a2b56557d692d52cc0e. It incorporates General's fully qualified PHP #469 source 1f994962b597e4cd24a4121a8a5c21b9c5ba1d27 / tree 647f0f8e43a3af4830f4523d4d580f46404d3eee, including the scoped sticky-affinity primary-key correction and current release metadata/numeric declarations.

WorkerController SHA-256 is exactly the handed-off 12b2fb5224e2a4de32e4113c8f29f7f1901841dd49fc9b6aba66e7bd341ed208. PHP application, Composer and release/chart bytes have no difference from that qualified source. Native source and frozen fixture tuples retain this candidate's earlier scope. The combined OpenAPI revision 55 retains the agreed PHP query additions and optional cross-kind hint documentation.

This push starts a new normal/shared qualification for the combined tree. The 3409a776 component's six-database outcome and MySQL-only retry remain exactly those earlier results. Its failed MySQL original, earlier PG17 fresh-upload failure and frozen reference failures are not relabeled or explained by the unrelated PHP correction. No pass is carried across trees. Current combined checks, ancestry/promotion and fresh-main qualification remain pending. General retains protected PHP 2.8.3 image/Helm publication and installed consumers; no published/native adoption follows from this source integration.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Independent audit of all six deadline-health artifacts in fresh combined run 38704820349 passes for candidate 688ee3ea / whole tree 3be0ab688a0ce47f15166a2b56557d692d52cc0e. The actual CI merge 93f080a9193402829826bc1e3b7ee4ce7825b112 has exact parents 1f994962 and 688ee3ea and the identical whole tree.

Artifact digests were independently checked before reading. Each database audit reports 14 original first-attempt tasks, zero repairs, completed timer/renewal and cancelled cleanup originals. Each also passes three actual SQL corruption refusals. The six build records agree on native subtree 7d0287cd8255e9a270010c2ab50c4cb21467f995, binary SHA-256 6cedb86226a0d7d2b2552a8e345f4febec9277586bbbf37c9f1517a848bc7f3b and locked source provenance. Artifacts: SQLite 11958302570, MySQL8.0 11957964011, MySQL8.4 11957664464, MariaDB10.11 11958731999, PostgreSQL16 11958407440, PostgreSQL17 11957809363.

Their process JSON still shows running snapshots, so it does not establish graceful shutdown. This is the six health components only; the six execution/differential jobs remain running, so all-13/all-six qualification, promotion and fresh main are pending. Earlier failures retain their original outcomes. The downloaded audit ZIPs are already deleted. No performance, release or native capability claim follows.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Exact final candidate 688ee3ea02a26501abf9fb3d506624f17e79a7c2 is qualified on whole tree 3be0ab688a0ce47f15166a2b56557d692d52cc0e. The tested synthetic merge 93f080a9193402829826bc1e3b7ee4ce7825b112 has that identical tree and exact parents qualified PHP main 1f994962 / candidate 688ee3ea.

Shared run 38704820349 passes all 13 jobs: installed adapters, all six execution configurations and all six deadline-health configurations. Normal feature/source 38704820742 and applicable boundary checks pass. Conditional prepared/cancellation remains skipped.

Independent readback verified all six execution artifact ZIP digests and their exact tested revision. Every cell retains 25 passing comparison records and 19 zero status files. The SQLite log separately confirms 70 library passes / 18 explicit ignores and 113 execution passes. Execution consumers identify native tree 7d0287cd8255e9a270010c2ab50c4cb21467f995 and binary 5590446ed00792d98e1d184ce90a3c4f92aa7794d6fc80cd48086eb4680917c5. The separately optimized deadline-health binary and its six original audits remain the distinct earlier receipt 6188016894.

Execution artifacts: SQLite 11960940288, MySQL 8.0 11960109308, MySQL 8.4 11960438209, MariaDB 11961425334, PostgreSQL 16 11960014366, PostgreSQL 17 11960998154. These are bounded outcomes and diagnostic tails. Full raw HTTP journals, payloads and databases are not retained. Downloaded audit copies are removed after this readback.

This qualifies the current upload/poll-hint source integration only. Earlier failed originals and the prior MySQL-only retry remain separate. Frozen tuples, the held query/upgrade/declaration preparations, unadvertised external capabilities, performance, migration and release/cutover limits are unchanged. Merge and fresh-main verification are next.

@rmcdaniel
rmcdaniel marked this pull request as ready for review October 15, 2026 09:24
@rmcdaniel
rmcdaniel merged commit 3013761 into main Oct 15, 2026
20 checks passed
@rmcdaniel
rmcdaniel deleted the rust-325-activity-completion-uploads branch October 15, 2026 09:24
@rmcdaniel

Copy link
Copy Markdown
Member Author

#458 is merged at 30137616bb8fa77d4b5417bda7448c15dd034181 (09:24:15 UTC), with exact parents 1f994962 / 688ee3ea and whole tree 3be0ab688a0ce47f15166a2b56557d692d52cc0e, identical to the qualified candidate and tested synthetic merge. The remote #458 head is absent. Included #455 commits are also on main, and its remote head is removed.

Fresh-main shared 38712683859, feature/source 38712683837 and boundary 38712683826 are pending. Candidate passes do not establish these new executions. Completed worktrees and downloaded audit ZIPs are removed. The earlier failed originals, frozen actors and capability/migration/performance/release limits remain unchanged.

@rmcdaniel

Copy link
Copy Markdown
Member Author

Final fresh-main checkpoint: 30137616bb8fa77d4b5417bda7448c15dd034181, whole tree 3be0ab688a0ce47f15166a2b56557d692d52cc0e, now passes shared run 38712683859, all 13 jobs: installed adapters, six execution cells and six deadline-health cells. Exact-head PHP feature run 38712683837 and boundary run 38712683826 also pass. Parents are qualified PHP 1f994962 and reviewed candidate 688ee3ea; #455 source 1fa1cc2e is an ancestor too.

All six original bounded execution artifacts were independently downloaded, digest-checked and inspected. Each summary identifies exact revision 3013761 and has 25 passing comparison JSONs and 19 zero status files. Native tree is 7d0287cd8255e9a270010c2ab50c4cb21467f995, execution binary SHA-256 5590446ed00792d98e1d184ce90a3c4f92aa7794d6fc80cd48086eb4680917c5.

Backend Original artifact
SQLite 11963683037
MySQL 8.0 11964069077
MySQL 8.4 11964314185
MariaDB 10.11 11964372688
PostgreSQL 16 11963483859
PostgreSQL 17 11964535789

Both merged remote heads remain absent and completed worktrees are retired. Downloaded audit ZIPs are deleted immediately after inspection; code and this readable outcome remain. Earlier failed originals stay failed with their recorded limits/unknown causes. This qualifies only this delivered foundation and exact 301 tree. It does not transfer passes to current PHP security source, open #463/#470 candidates, native discovery, migration, performance or a release.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants