Release readiness

This report records the earlier baseline. See the 9 September release candidate report for account-name support and the latest verification status.

The SDK code passes local and hosted package, release-build, simulator, and performance verification. Live-provider compatibility remains unverified because no current SDK/demo session is saved on this Mac. The next prerelease is v2.0.0-alpha.27; its changelog entry and migration guidance are prepared. No release tag or publication has been created.

The storage-lock cancellation issue is fixed. Cancelled lock acquisitions and queued direct writes stop before acquiring a lease; commits already underway finish atomically. Shared preparation and runtime-owned accepted writes retain their completion and ordering guarantees. Ten new regression tests cover the original failure and cancellation through the storage queues.

Verification results

Check	Result	Scope
Full package suite, warnings as errors	527 executed: 521 passed, 6 opt-in skips, zero failures	Core, UI, SQLite, Realm, compatibility and regression tests, including six waves per adapter in the new concurrency harness
Optimized release build, warnings as errors	Passed	All four SwiftPM library products
Optimized performance/SDK/storage selection, warnings as errors	29 passed, zero failures	Image-heavy compaction, large histories, cancellation, parser, pipeline, SDK ownership, all ten storage regressions and three concurrent lifecycle tests
Sustained concurrent lifecycle workload	Passed on all three disk adapters	40 waves per adapter: 720 overlapping operations, 120 reader reopens, 60 owning-runtime replacements, and independently checked transcripts/attachments
Signed simulator execution	Passed	iOS 26.5 / iPhone 17 Pro; plain/structured completion, SQLite/Realm reopening, cancellation
Verification harness	7 passed	Build/run mode separation, portable build settings, failure details, runtime selection and report validation
Source-size guard	211 production files passed	208 Swift files and three verification scripts, each at most 600 physical lines
Source-size boundary checks	Passed	600 lines accepted; 601 rejected in an isolated fixture repository
Workflow YAML and whitespace	Passed	Both workflows parse; git diff --check clean
Optimized workflow failure propagation	Passed	The exact CI/release shell steps preserve success and failure exit codes through log capture using a local stub command
Hosted CI	All four jobs passed	Current package/release/optimized/simulator checks, macOS 14 with Swift 6.1.3, and a Swift 6.1 app build followed by actual iOS 17.0.1 execution
Live tests explicitly enabled	2 skipped	Neither known SDK/demo Mac Keychain entry contained a current session

The six ordinary-suite skips are two live tests and four opt-in performance workloads. The latest local package run completed after the cursor and Realm compiler fixes with zero failures in 152.749 seconds. The optimized 29-test selection then passed in 118.035 seconds, and all four library products passed a separate release build with warnings treated as errors. The signed iOS 26.5 simulator check was repeated after the cursor and Realm compiler fixes; run ID 84a31735-7d67-4ece-bdc2-8323af0d6b8b passed SQLite and Realm completion, reopening, and cancellation on 8 September at 06:13:55 UTC. Live-account access was disabled for that run. The portable archive was separately checked for both architectures and valid signing, extracted, and executed successfully; run ID 7d71d3b4-f7a8-488c-98ae-32dc6cb0aaa3 passed on iOS 26.5 at 06:35:28 UTC.

Hosted CI passed all four jobs for code revision d75130cfcfd6feaf1054c13bd9c00e55050092c2. This includes both package profiles, the Swift 6.1 simulator-app build, actual iOS 17.0.1 execution, and the current-system simulator. The final documentation-only follow-up records those results; it changes no tested code or verification workflow.

The macOS 14 package checks use Swift 6.1.3. The iOS 17 app is built with Xcode 16.4 on macOS 15, then installed on the macOS 14 runner's iOS 17.0.1 simulator. This separation is necessary because Xcode 16.2's embedded package resolver cannot load Swift 6.1 packages.

CI retains separate current/minimum package logs and iOS 17 build/runtime reports. Current-host and release jobs also retain the optimized test log, and log capture preserves failures. The simulator verifier rejects missing, stale, incomplete, or failed adapter reports, removes its temporary simulator, and explicitly disables live-session access. Manual release dispatch preserves the earlier verification paths for tags that predate the new stress tests or simulator verifier.

Hosted-verification fixes

Hosted checks exposed two additional SDK issues. History-page equality could fail intermittently because identical cursor payloads used different JSON key orders. Cursor generation now uses canonical key ordering for both cursor versions; three new tests verify stable output and continued decoding of previously issued cursors.

Swift 6.1 also rejected transaction closures that captured Realm values alongside the store actor. Synchronous persistence and attachment-reference helpers now operate on explicit inputs, and memory transactions capture their helper values directly. Realm opening and asynchronous writes stay on the owning actor. No Sendable checks were disabled, no dependency sources were edited, and no public adapter API or database schema changed. The complete local suite passed after this extraction.

The minimum profile uses Swift 6.1.3 because GRDB 7.10 already requires Swift 6.1. Model-catalog decoding was simplified to fit that compiler's type-checking budget. Xcode's Apple Clang builds Realm's C++20 dependency modules; the standalone Swift toolchain's Clang could not import s2geometry in this configuration.

Public API compatibility

The release review covered the accumulated core/UI API changes, the database-adapter source diff, CI/release gates, and their changelog/migration guidance.

Swift API Digester compared isolated core/UI modules against local tag v2.0.0-alpha.26, commit 87f4d5a700aff88a6da636eba05d48dc2def1e9c, using the arm64-apple-macosx14.0 target.

• Core: seven initializer signatures gained defaulted parameters. Representative ordinary calls compile in PublicAPICompatibilityTests; exact initializer-as-function references require closure adapters.
• Core: AgentRuntime.Configuration.authProvider and .secureStore changed to optional types. Host-managed sessions can omit both, so configuration inspection needs optional handling.
• Core: no new protocol requirements were reported, preserving existing backend conformances.
• UI: no removed or changed public declarations were reported.
• Database adapters: the accumulated source diff changes internal relationship filtering and Realm history-window selection, with no public declaration changes.

The migration guide includes examples and the reproduction command. API diagnostics live in .build/api-review. This review checks the macOS-visible source surface and an iOS consumer build; it is not a binary ABI guarantee or an exhaustive source-compatibility proof for every platform-specific client.

Behavioral changes also require attention: one-shot output now validates before persistence, raw schema support is explicit, skill fields are strict, cancelled startup cannot launch work, overlapping compaction reports busy, and runtime/transport limits have finite defaults. These are documented in the migration guide and changelog.

Remaining release gates

1. Supply a current SDK/demo session and pass the live plain/structured and image → compaction → database reopen → follow-up checks. The complete image matrix currently runs on the Mac package-test host; signing into a separate iOS device does not populate its Keychain.
2. Publish the prepared v2.0.0-alpha.27 after the live checks pass: update the README release links, bring the reviewed changes from codex/release-validation onto main, and push an annotated tag through the normal release process. The release workflow reruns package, optimized, and simulator verification before publishing the GitHub prerelease. Any later code changes must pass the same CI checks again.

The performance follow-up batches compact-response bytes, validates image references without expansion, and reads consecutive Realm history windows by primary key. Eight new regression tests passed. The observed eight-image compaction and 20,000-record paging times decreased by 48% and 79% respectively; these local measurements introduce no performance-based release blocker and make no live-provider or physical-device energy claim. These changes add no public API or database schema migration.