ci(bindings): run the Kotlin and Swift wire-contract tests on the shipped Binding Sets

[zancas]

Step 1 of #2930. Each piece here adds coverage in zingolib. Nothing is deleted.

What was already true

zingo-ffi/lib is a root-workspace member, so test.yaml runs the wallet crate tests on every PR. The netutils job runs the proxy crate tests and the live mixnet start. The test a_mixnet_refusal_maps_to_the_owned_mixnet_marker has pinned the mixnet marker since a67de8d. Step 1 therefore reduces to the two wire-contract tests and the glossary.

The Kotlin wire-contract test

The Android binding build gains a test source set over contract-tests/kotlin, JUnit 4.13.2 as a test dependency, and the golden directory as the zingo.golden.dir system property. The package job of binding-layer-android.yaml runs testReleaseUnitTest after it bundles the AAR. The test then compiles against the Kotlin Binding Set that the AAR ships, so no host library is needed.

The Swift wire-contract test

The Swift test has never run anywhere. A test-only package at contract-tests/Package.swift depends on the ZingoBindings package by path, so the shipped Package.swift gains no test target. The test now imports that module. Its golden path was one directory short, and it gains the missing component. The iOS workflow runs the test under xcodebuild on an available simulator, because the XCFrameworks hold iOS slices only.

The glossary

zingolib/CONTEXT.md gains Binding Set, Binding Layer Endpoint and Bundle Pin. It restates Prebuilt Bundle to the packaged form that #2833 publishes. Ruling 9 of #2930 refers to these terms.

Verified

• Both workflow files parse.
• The Android binding build configures with the new wiring, and its unit-test classpath resolves JUnit and JNA.

Not verified here

• The Kotlin test compiles only against generated bindings, which need the proxy crate's host build. The package job proves it.
• The Swift step needs a Mac. The nightly and the publish workflows run it.

JUnit is a new Gradle test dependency. The test file imports it.

Refs #2930.

🤖 Generated with Claude Code

[juanky201271]

The Kotlin half holds: the test only uses allocationSize/write/read over a ByteBuffer, so it never touches UniffiLib and needs no host library. Two things on the Swift step, which the PR says no Mac has run, and one on docs.

Docs this PR leaves stale: contract-tests/README.md still says the app's Gradle unit suite runs the Kotlin test and that the Swift one joins the step-7 packaging (#2504); the headers of GoldenWireContractTest.kt ("zingo-mobile's test suite wires it in") and GoldenWireContractTests.swift say the same. zingolib runs both now; the README should name binding-layer-android.yaml and binding-layer-ios.yaml.

[juanky201271]

All three addressed, thanks. The Kotlin step compiles and runs in the package job.

I dispatched ci-nightly on this branch and the Swift step fails at link time: ~37 undefined SystemConfiguration symbols from hickory_resolver / system_configuration in libzingo_nym_proxy_ffi.a (job). The shipped bindings/swift/Package.swift doesn't link the framework; zingo-mobile only works because its pbxproj links it by hand. linkerSettings: [.linkedFramework("SystemConfiguration")] on the ZingoBindings target should fix it. Good catch by the test, but as is this breaks the iOS job of the nightly and /publish.

nit: xcodebuild warned "Using the first of multiple matching destinations". The parser already has the UDID, so -destination "platform=iOS Simulator,id=<udid>" avoids that.

[zancas]

Thanks for dispatching the nightly. The link failure is real, and 87c401f does not fix it. 85b499d adds linkerSettings: [.linkedFramework("SystemConfiguration")] to the ZingoBindings target in bindings/swift/Package.swift and pins the setting in the workbench manifest test.

The destination nit is already in 87c401f. The nightly you dispatched ran at ac55e65. The step now passes -destination "platform=iOS Simulator,id=<udid>".

87c401f also closes eight items from a pass over ac55e65:

1. Simulator choice. ios_simulator reads the runtime version of each simctl entry and skips runtimes below IOS_DEPLOYMENT_TARGET. It prefers a booted iPhone and otherwise takes the first eligible one. It returns the UDID.
2. Output check. The built-versus-consumed check of --out compares the two paths lexically. It no longer canonicalizes the absent bindings/swift/build, so a wrong --out gets the authored message instead of "cannot resolve".
3. Output check timing. That check now runs in build() for iOS, before any step. A wrong --out is refused before the thirty-minute build, not after it. The unit test asserts the refusal instead of a plan that must fail.
4. GoldenWireContractTests.swift no longer reads ZINGO_GOLDEN_DIR. xcodebuild forwards only TEST_RUNNER_ variables, so the override never worked. The test resolves the golden directory from its own path.
5. xcodebuild runs without -quiet. The step prints the log and parses the last "Executed N tests" line. It requires N to equal SWIFT_CONTRACT_TEST_COUNT, which is 3. A unit test pins that constant to the func test methods in the Swift file. Zero discovered tests now fail the step.
6. The contract-tests README states that Gradle sets zingo.golden.dir to the absolute path of the golden directory.
7. run_command and capture_command share one working-directory check. A missing contract-tests directory now names that directory instead of blaming xcrun.
8. The header of binding-layer-android.yaml is rewrapped.

No Mac has run the Swift step since 87c401f. The iOS job on 85b499d is the proof.

Requesting a new review.