DX7: Java, .NET, C++, mobile (Swift and Dart), and the tier-3 kit #18

Lands with M8. The last of the tier-1 set, plus the tier-2, mobile and tier-3 machinery.

Scope

• yo-java, Java 25 LTS with FFM, Arena.ofConfined(), Linker.Option.critical(true) chosen per function from api.model.json
• --enable-native-access detected at Db.open with a message that says exactly what to add to the command line
• A JNI provider for Java 17 and 21 as a second artifact, with its numbers published even though they are worse
• yo-dotnet, net10.0, [LibraryImport], SuppressGCTransition, NativeAOT
• yo.hpp, header-only over the C ABI, YO_SHAPE and &User::score
• yo-kit: the generator, 7 emitted items, 10 named templates
• The agent conformance suite, roughly 40 tasks, published including the failures
• The full 2300-case conformance corpus green in all six tier-1 languages

Mobile, tier 2

Swift and Dart were tier-3 templates. dx/14-swift.md and dx/15-dart.md promote them to tier 2 on one argument: mobile is the only place where "no server" is a constraint rather than a preference, so an embedded typed database fits a phone better than it fits a server — and a template maintained by nobody is not a mobile story, it is a checkbox. Tier 2 and not tier 1 because the API may lag one minor, which is what keeps the release-train barrier affordable when Apple or Flutter ships in a release week.

Swift (tamnd/yo-swift)

• SwiftPM package: CYo system/binary target, Yodb idiom layer, YodbMacros
• libyo.xcframework as a .binaryTarget with a SHA-256 SwiftPM verifies itself
• Key paths are the typed surface. \.score is a KeyPath<User, Double>; order(by:) and the comparison operators are checked by the compiler with no macro, no string, no generated constant. The strongest realisation of P3 in the tree
• @Document / @Value / @ID / @Index macros, emitting only what a key path cannot: the canonical shape description and the field-index map
• swift-macro-testing snapshots for every expansion and every diagnostic, with fix-its. Swift's trybuild
• Typed throws: throws(YoError), Code generated and deliberately not @frozen
• Sendability stated per type; Transaction is ~Copyable so consuming func commit() makes use-after-commit a compile error
• Span/RawSpan for the borrowed view — the reason the floor is Swift 6.2
• Async only where the cost class admits unbounded work, on a dedicated TaskExecutor, never blanket-async
• Swift 6 language mode only, -strict-concurrency=complete, warnings as errors

iOS, and it is the point

• File protection set to completeUntilFirstUserAuthentication at creation. NSFileProtectionComplete + mmap is a SIGBUS, not an error, so protection: .complete disables mmap instead of crashing
• Checkpoint on background inside a beginBackgroundTask window; release cache on memory warning; opt-out, not opt-in
• PrivacyInfo.xcprivacy declaring FileTimestamp (C617.1) and DiskSpace (E174.1). A wrong manifest is an App Store rejection for the user's app caused by our package. CI diffs it against a symbol scan
• Static linking only; no JIT anywhere, asserted at build time
• Size budget as a hard gate: <= 1.6 MB KV+docs, <= 3.2 MB default, per thinned arm64 slice

Dart / Flutter (tamnd/yo-dart)

• FFI plugin covering iOS, Android, macOS, Windows, Linux, plus Dart standalone
• Binaries vendored, not downloaded at build time. ~30 MB package, accepted under P9. A post-install download is the top reason a Flutter native plugin fails in CI or behind a proxy
• build_runner + source_gen shapes. Dart has no macros — the proposal was cancelled — so the shape is a part file and the spec says so instead of claiming parity
• yodb_lint custom-lint so an un-generated @YoDocument gets one sentence naming the command, not an analyzer riddle
• @Native(isLeaf: true) on every O(1) function, marked from the cost class in api.model.json. ~10 ns overhead is what makes a synchronous get on the UI isolate defensible
• One long-lived worker isolate per Db, not Isolate.run per query. Scans never touch the root isolate, and that is not configurable
• sealed class YoException for exhaustive switch; NativeFinalizer + Yo.use closure + a debug leak detector, because Dart has no scoping construct
• Android: API 24, arm64-v8a and x86_64 only. armeabi-v7a dropped — a 32-bit address space caps a mapped .yo file, and silent failure past a size is worse than absence
• 16 KB page alignment and extractNativeLibs=false, CI-checked: a misaligned .so is a Play Console rejection for the user's app
• Web via the wasm build, conditional export, labelled tier 3 on every web-only member
• Flutter frame-budget test: 60 fps over a 100,000-row scan, as a release gate

Shared

• Static libs for aarch64-apple-ios, -ios-sim, x86_64-apple-ios, aarch64-linux-android, x86_64-linux-android in the release matrix
• Corpus per PR on desktop, per merge on the simulator and emulator, weekly on physical devices
• dx/01 snippet budgets: Swift 14 lines, Dart 16
• Predictions registered under Y26: Swift 200 ns, Dart 250 ns

On publishing failures

The agent suite publishes tasks agents get wrong. That is the point of it. A suite that only shows successes measures nothing, and the failures are the list of things to fix in the next release.

Exit

Six-language interop green across all 30 ordered pairs, and a .yo file written in Rust read correctly from Swift and from Dart. Mobile size budgets and the Flutter frame-budget test pass. The agent suite baseline is published, failures included.

Reference: dx/08-jvm-and-dotnet.md, dx/09-c-and-tier3.md, dx/10-agent-dx.md section 6, dx/14-swift.md, dx/15-dart.md.