build: add an R8 shrink-survival test module and ship consumer keep-rules

[OmarAlJarrah]

Summary

Adds consumer keep-rules and a regression guard so that downstream applications which shrink with R8/ProGuard don't strip or break the SDK's reflective and SPI surface.

Shipped consumer keep-rules

New META-INF/proguard/*.pro files in sdk-core, sdk-io-okio3, sdk-transport-okhttp, sdk-transport-jdkhttp, and sdk-serde-jackson. Because they ship under META-INF/proguard/, downstream R8/AGP applies them automatically. They keep the Io seam, IoProvider, the HttpClient/AsyncHttpClient/Serde SPIs, the HTTP models + builders, the Tristate hierarchy, kotlin.Metadata, and Jackson's reflective databind/core/annotation surface. Both reference transports protect their construction surface, so neither can lose its rules unnoticed.

sdk-shrink-test regression module (test-only, unpublished)

Bundles a small consumer program plus the full runtime classpath into a fat jar, runs R8 using the shipped keep-rules (extracted from the SDK jars on the classpath and asserted present, so the test guards the real shipped rules rather than a private copy), then runs the shrunk program and asserts it works. R8 genuinely shrinks the bundle, and the shrunk run exercises the kept surface for real: installing OkioIoProvider, an in-process HTTP POST round-trip through both OkHttpTransport and JdkHttpTransport (via the core HttpClient SPI), and a Tristate JSON round-trip through JacksonSerde (the most shrink-fragile, reflection-heavy path). So it proves kept members still function after tree-shaking, not merely that class names survive.

R8 is run in shrink-only mode (obfuscation disabled): the guard is that dead-code elimination doesn't strip the SDK's reflective/SPI surface. Obfuscation survival is out of scope here because renaming would also rename the bundled third-party libraries (OkHttp, Okio, Jackson), each of which ships its own consumer keep-rules that an obfuscating consumer applies; the SDK's own -keep ... { *; } rules already pin names against renaming.

Wired into the check/build lifecycle. The module targets JDK 11: it depends on the Java-11 sdk-transport-jdkhttp so it can drive that transport through R8, and the whole pipeline already runs on a JDK 11 (R8 is Java-11 bytecode and the shrunk program runs on the same launcher), so a consumer that uses the jdkhttp transport is an 11+ consumer by construction. The module is excluded from apiValidation and the Kover aggregate (matching the example module); ktlint and explicit-API strict mode stay on. Detekt is skipped here for the same JDK-25-system-toolchain reason as the other non-Java-8 modules.

Real gaps this surfaced while being made to pass

• A consumer must supply an SLF4J binding (it's compileOnly in the SDK); the harness bundles slf4j-nop.
• The initial Jackson keep-rules were too narrow — its SerializationConfig static init failed because annotation enums were stripped. The shipped sdk-serde-jackson.pro was widened to keep the Jackson annotation/enum surface, which is the conventional recommendation for that library and exactly the regression class this module now guards.

Notes

• The build now includes an R8 step that requires a JDK 11 toolchain and the Google Maven repo (for com.android.tools:r8); this is documented in CLAUDE.md. An offline build, or one that cannot provision JDK 11, fails on :sdk-shrink-test:r8Run — scope the build to skip it.
• R8 8.9.35 logs a non-fatal note that it can't parse Kotlin 2.3 metadata (its max is 2.1); the @Metadata annotation is retained via -keepattributes / -keep class kotlin.Metadata (proven by the passing serde round-trip), and a future R8 bump quiets the log.
• The issue's CI smoke job is blocked on CI existing (Add CI workflows to enforce the build gates on every PR #70); hanging the check off the Gradle lifecycle means CI picks it up automatically once that lands.

Touches settings.gradle.kts, so expect a rebase against the example-module (#102) and convention-plugin (#105) PRs depending on merge order.

Closes #72