Implementation status

Modernization acceptance inventory

Current candidate: 1.26.0-rc.1, v2 namespace, PHP 8.2+/PhpRedis 6+.

Current workflow (2026-10-05): the owner has now explicitly requested code changes without running tests. Runtime, lint, static analysis and browser checks are deferred; no new passing result is inferred from source inspection. Earlier TTL/WordPress/Redis/minimum-runtime/concurrency/commerce results remain historical evidence for the exact revisions in RELEASE-COMPLETION.md. The subsequent CDN, reset handling and static-contract source edits have not been executed. Optional Query Monitor analysis declarations and guarded Members calls have since been added in source; whole-project analysis has not been rerun. Query Monitor now handles missing/replaced collectors and labels retained command data/omissions. Counter/remember stress, matched HTTP performance, backend/migration cases and operator-controlled rollout remain open.

The approved wp redis profile <url> method is implemented in source: one real same-site anonymous GET, a five-minute HMAC ticket consumed through captured-primary WATCH, early request-only 100% collection, actual site confirmation and a bounded signed Redis result consumed by CLI. Default 2,000 events, maximum 10,000/2 MiB; table/JSON expose omissions, logical size coverage and separate API/HTTP timing. Forced samples bypass hourly Insights persistence. Response bodies stream to an owned temporary file and cleanup is reported. HTTP-PROFILE.md records scope, failover/expiry, late shutdown, page-cache and privacy limits, plus deferred tests. A full signed profiling GET acceptance remains pending; shared observer syntax/static and regression checks have since run as recorded in RELEASE-COMPLETION.md. The HTTP profile now also records bounded confirmed-write TTL evidence from the existing write path: requested seconds, capped/sent ranges and nine buckets, with a maximum of 256 inspected replies per API call. Runtime-only and preserved counter expiry are explicit. Existing hourly hashes, cache semantics and Redis command counts are unchanged by this observation; actual overhead is unmeasured. The owner explicitly approved §9 on 2026-10-05. The separate opt-in TTL history store now has source implementation: request-wide 256 replies/64 rows, site/hour 64 named groups + other/64 KiB, captured-primary WATCH with finite EX and absolute EXPIREAT, 192-hour retention, bounded completed-hour reporting and explicit CLI/REST/admin loading. Existing hourly counters remain unchanged. See TTL-HISTORY.md for passing regression, real Redis, REST and browser results; extended backend/failure parity and overhead acceptance remain open.

The optional ttl_jitter setting defaults to zero and accepts integer percentages from zero to 100 (including valid integer strings through the existing configuration parser). It shortens effective positive TTLs only after the maximum/query limits, uses integer arithmetic without overflow and never creates a zero TTL. Single and bulk writes plus split alloptions share the policy; bulk keys are sampled separately. Effective zero/one-second TTLs, counter mutation expiry, nonpersistent groups and internal metadata/prefetch retention keep their existing behavior. Configuration export/diagnostics include the percentage. Existing keys are not rewritten; turning the option off does not undo earlier expiry. Jitter is not stampede protection; the separate recomputation/remember API now has its own approved contract below. Final validation must cover configuration boundaries, integer limits, NX/XX, zero/one/positive TTL, maximum/query caps, independent bulk draws, split alloptions, counter retention, both real backends and five-repeat off/on performance comparison.

The owner approved the separate ocp_remember() API. Source implementation adds default-empty stampede_groups, bounded leases/waiting on the existing captured primary transaction interface and conditional publication after at most one callback invocation outside transactions. Configured mutations use a confirmed lease cancellation before their watched value mutation; bulk operations use that per-key path. Group/site/namespace cleanup and stale-query pruning cancel old computations before scanning data. Cache conflicts do not replay mutations; unconfirmed results are evicted from local memory. Ordinary remember/sear helpers retain their acquisition/wait behavior, with the false-hit bug fixed. Nonpersistent groups retain runtime caching. Full details, Redis 6.0.9+ lease requirements, uniform-worker configuration, partial EXEC/failover limits and deferred final validation are recorded in STAMPEDE.md. No syntax/runtime/concurrency or performance test was executed for this feature; real backend parity is not established.

Cluster and cache-enabled Relay now retain up to eight direct transaction sockets per owning connection object, with oldest-idle eviction. Reuse requires unchanged connection settings and clean, unchanged native database/codec/retry/timeout state. Unsupported state inspection keeps the former one-operation socket lifetime. Exceptions, error-counter growth or uncertain cleanup retire the socket. Cluster still discovers CLUSTER SLOTS before every transaction; Relay conditional reads still bypass its local value cache. Neither path replays a transaction. Observation deltas avoid counting earlier operations twice; a new lease clears lastCommand. Initial codec setup now rejects failed native option changes and closes the client. Final tests must cover reuse/eviction, configuration changes, conflicts and error cleanup, metrics, unavailable capabilities, real Relay/Cluster failover/resharding, and five-repeat connection-cost/performance comparisons. No speedup is established.

Drop-in file mutations now require nonempty product-name and URI headers matching the bundled stub. Legacy upstream compatibility and the compatibility filter do not grant ownership. Protected files are explained in management/health screens; CLI disable also refuses to remove them. CLI enable –force retains its explicit replacement behavior and warns on unowned files. It does not create a backup. The filesystem guard refreshes cached identity, but is not an atomic compare-and-swap against other processes or a cryptographic integrity check. Existing drop-in and metadata fixtures need ownership cases before the final test phase; no prior passing report validates these changes.

The drop-in migration wizard is now implemented in source using the existing settings pages and WordPress filesystem API. It requires core POST nonces, cache management and plugin-installation capabilities, plus a network super administrator on multisite. Preview/backup/install bind both current and replacement fingerprints; the download issues a ten-minute, user/network-bound content-free confirmation. Backups contain signed, unencrypted PHP bytes (or an absent-file record), bound to the content path, network URL and auth signing key. Restore accepts the exact recorded replacement or an already-restored state, and never flushes Redis or rewrites provider configuration. The operator must prepare the fresh rollback namespace. Staging/verification/cleanup use WordPress’s transport; concurrent external writers and transport move/copy failures do not have an atomic rollback guarantee. See DROPIN-MIGRATION.md for limits and manual recovery. Core nonce and capability checks, download/upload flows, multisite, tampered/stale/oversized/link cases and actual filesystem/OPcache failures remain final-phase validation gates.

withoutMutations() now captures the physical client’s actual serializer and compression settings, validates option changes, restores both options despite an individual cleanup failure, attempts to close an uncertain client and retains the first error. Replicated scopes unwind the originally captured nodes, including partial setup and nested calls. Final verification must cover PhpRedis/Relay, replica setup/cleanup failures, callback/close error precedence, nested calls and Sentinel replacement during a raw callback. Active raw scopes now enroll each discovered primary/replica before it is exposed to callers. Nested scopes enroll new nodes from outermost to innermost and restore each captured physical client, including retired nodes, even after a setup/callback/cleanup error. Direct pinned Cluster/Relay transactions inherit their owner’s raw scope and restore normal options before a socket can reenter the pool. This closes the source-level gap where replacement/direct sockets used normal codecs inside a raw callback. Runtime evidence remains pending: cover read retry after rediscovery, a callback that catches a write failure and continues on the new primary, direct primary access, nested discovery, duplicate node references, raw/normal pooled operations, and native option failures. The Sentinel read/write replay policies are unchanged.

Historical validation below applies to its recorded revisions. The earlier error-buffer reporting update passed 38 PHP unit regression cases and 152 assertions on PHP 8.3.0, including 24 focused bounds/reporting checks. Ten PHP source/template/test files passed syntax checks. The seven-file PHPStan scope reports eight findings in unchanged API/diagnostic code; an immutable-source comparison reproduces all eight on 42f7acc, with no new findings. This is not a clean static-analysis result. The real WordPress evidence below predates this reporting update and does not validate its behavior in a live admin request.

This file records source implementation and its verification limits. Following the owner’s later authorization to run PHP tests after these updates, the local PHP 8.3.0/PHPUnit 9.6.37 suite passed 37 regression cases and 148 assertions on 2026-10-04. One real WordPress/Docker feature test was initially skipped. The owner then authorized the real WordPress run. A preceding fresh Docker fixture passed all seven integration scripts and the lifecycle suite, including that skipped test’s contracts.php scenario. All 37 regression cases also passed inside PHP 8.3.35. The fixture used WordPress 6.8.3, WooCommerce 10.2.2, PhpRedis 6.1.0 and Redis 7.4.2. Installed runtime/drop-in identity was verified against 187 source files. This establishes the recorded standalone behavior, not Relay or production acceptance. Browser, heavy concurrency, scale and performance gates remain open. Prior reports retain their original revision scope.

Remaining implementation and evidence work

• The new wp redis why source probes one scoped key without reading its value, retries or a keyspace scan. Existing transaction routing and native type constants are reused. TTL explanation shares the existing jitter reduction arithmetic and draws no random number. Syntax/static/runtime verification has not run for this change. Component attribution and the Insights screen/rules now have source implementation under the separately approved bounded storage contract (§5 of REFACTOR-DECISIONS.md); their runtime and performance acceptance remains open. wp redis profile now has source implementation under the separately approved contract in §6 and HTTP-PROFILE.md. TTL/persistence classification and read-only evidence review now use the existing completed-hour counters without changing storage. Exact TTL/reuse/expiry/freshness measurement and numeric TTL proposals remain phase-two development work; the evidence layer does not complete adaptive TTL. None of these new source changes have runtime or performance acceptance.

• Key-explosion detection now has source implementation under the explicitly approved HLL/fingerprint/SCAN contract in REFACTOR-DECISIONS.md §7. Collection defaults off. HMAC observations, fixed site/hour slots, primary-side HLLs, fenced explicit scans, bounded history, scoped REST/admin/CLI output and opt-in doctor guidance are described in KEY-GROWTH.md. Source review only; no PHP, static, runtime, browser or performance checks have run for this change. Missing, partial, incompatible and approximate results are explicit; no automatic scan, cleanup, configuration change or production acceptance. Source follow-up: hourly JSON and scan-history writes now also confirm EXPIREAT at the approved absolute retention deadline. SET keeps its finite EX fallback; both replies are required. This removes relative-TTL transport/queue drift on successful publication, adds one expiry command per metadata write, and does not promise rollback or a durable marker for partial expiry failure. Delay/ACL/ clock-skew checks remain deferred in KEY-GROWTH.md.

• The owner approved REFACTOR-DECISIONS.md §8 on 2026-10-05. Configuration profiles now have source implementation: preserved early config, four built-ins/custom manifest data, explicit-field precedence, exact group TTL caps, bounded local import and JSON export, CLI differences and read-only current-site management previews. Existing bootstrap groups and counter TTL retention remain. No live configuration file was changed. CONFIGURATION-PROFILES.md records semantics, rollback and final schema/TTL/CLI/REST/backend/browser/performance gates. Tests have not run; source completion does not grant phase-two/production acceptance.

• Dedicated 20-process alloptions and 100k/1m cleanup sources are connected to the explicit run.py –acceptance option, with installed-source verification. They have not run. A separate –commerce source scenario now submits two real HTTP checkouts and checks session/order/stock isolation; it is also unexecuted.

• Local regression coverage includes scopes, budgets, new connection capabilities, cleanup failure precedence and request-memory invalidation. Old bulk-key/counter fixtures now follow the approved v2/transaction contracts. An IPv6 seed port bug found during the run was fixed. Eleven changed PHP files passed syntax checks; nine connection/configuration files passed PHPStan level 5 with zero errors.

• Equal-sized prefetch record/index containers can contain different members after a partial EXEC failure. Eviction now verifies the space it needs and repairs detected mismatches in that site’s metadata. Regression coverage includes a partially applied EXEC and fresh/stale native errors. Real Redis passed 50 checks with seeded inconsistent metadata, serialization, TTL and site/data isolation; it did not inject a real server EXEC failure. Three PHP files passed syntax checks and three related production files passed PHPStan level 5. Extra reads have not been benchmarked; failed EXEC still has no rollback guarantee.

• Freeze a new review ZIP from the final immutable source after remaining acceptance.

• The eight errors in log-bounds-static-comparison.json describe its historical baseline/candidate, not the current checkout. Source inspection on 2026-10-05 found the callable multisite initialization guard in api.php, the Basename bootstrap/dynamic-constant setup in phpstan.neon.dist and tests/PHPStan/constants.php, and the single-instance Relay adaptive-cache guard and removed obsolete ignores in Diagnostics.php already present. Do not implement those fixes a second time. A fresh final-stage static run is still required to establish the current result; source inspection and historical reports do not verify real Relay behavior.

• Cursor-based management group listing remains a proposal in REFACTOR-DECISIONS.md; the existing full-list behavior has not been changed without separate approval.

• Local PHP regressions and the default fresh WordPress integration run are now authorized and completed. Topology/TLS, browser, heavy load and performance gates remain open. Licensed Relay behavior is explicitly unverified.

• No new frontend framework, hosted CI provider, update server, automatic optimizer, production deployment or external telemetry is introduced in this release.

Local reports are under build/modernization: phpunit-final.txt/phpunit-final.xml, transaction-cleanup-syntax.json and transaction-cleanup-static.json. Counts and changed PHP source hashes are recorded in tests/integration/status.json. These results cover their stated scope only. Error-buffer evidence is in log-bounds-unit.xml, log-bounds-syntax.json, log-bounds-static.json and log-bounds-static-comparison.json; its source hashes and limits are under log_bounds_validation in status.json. The real WordPress evidence is in tests/integration/results/ocp-check-bf3a22fd1247/run.json; its digest, tested source hashes and summary are recorded under prefetch_membership_validation in status.json. The earlier ocp-check-cab5b3ed7a90 evidence retains its original scope. The latest run tested the checkout based on 5055471 plus the prefetch fix, not an exported ZIP. A commit/ZIP records a reviewable checkpoint, not successful production acceptance.