kubuszok/balticporter/cross-platform-port

Three platforms, one port

A green JVM compile says nothing about the other two rows. Scala.js compiles everything and fails at LINK on an unknown method; Scala Native links whole classes and fails on a method the javalib never declared. Every platform question is answered the same way: read the javalib, record the fact, then decide.

1. Check the javalib, do not guess

J=$(find ~/Library/Caches/Coursier/v1 -name 'javalib_native0.5_3-*.jar' | sort | tail -1)
unzip -l "$J" | grep 'java/net/'                       # Native: which classes exist
JS=$(find ~/Library/Caches/Coursier/v1 -name 'scalajs-javalib-*.jar' | sort | tail -1)
unzip -l "$JS" | grep 'java/lang/ThreadLocal'          # JS: class present?
unzip -p "$JS" java/lang/ThreadLocal.sjsir | strings | grep -i withInitial   # JS: member present?

A Foo$.sjsir/Foo$.nir entry is the companion — no companion, no static factory.

Measured so far (JS 1.22.0, Native 0.5.11/0.5.12):

• java.net.URL: absent on BOTH (Native ships URI, URLEncoder/URLDecoder,

MalformedURLException). Catalog JS-P35; rule java.net.URL (both), Class#getResource (both), Class#getResourceAsStream (JS only — Native declares it and answers null).

• ThreadLocal: class on both; withInitial on Native only. A Java static final scratch instance

is shared by all threads, so confine it to a ThreadLocal subclass overriding initialValue rather than withInitial, which Scala.js lacks: spell new ThreadLocal[T] { override def initialValue(): T = e }.

• java.lang.invoke (MethodHandle): JVM only. Reflection (Class#getMethod…): JVM only.

2. Where a fact goes

• balticporter/api/…/catalog/ApiRows.scala: one row per API, availability + verdict per

platform. A rule that cites a row (at = p(n)) must not claim a platform the row calls Keep (PortabilityTargetsSpec enforces it, and its JS-only list must name every JS-only rule).

• balticporter/engine/…/tir/PortabilityCheck.scala: the rule, on = Rule.JsOnly/NativeOnly

or both, exactMember for one method. A more specific rule (java.net.URL) precedes its package prefix (java.net.).

• The port's answer is policy: a CallSiteSubstitutionTransform respelling

(getResource(x) != null → getResourceAsStream(x) closed and null-checked), a drop + inject, or a PortManifest.targets narrowing WITH its reason. Never a JVM-only dependency in the consumer (no jackson/ANTLR/reflection in ssg; "hand-maintained port" is not an answer).

3. The reference port already decided most of it

git -C ../sge show origin/master:<path> — master's migration notes name the platform decision (ThreadLocal … NOT the withInitial static factory (ISS-832 JS-link regression)). Read it before inventing a spelling; the port emits master's spelling, and the engine records why: the ThreadLocal confinement above, and a member rename attached to a redirect of an external type (such as Comparable.compareTo to Ordered.compare) must bind even though the run owns no declaration of that type, honouring the redirect's scope.

4. Verify on every row before pushing

sge: testCompile-jvm-3; testCompile-js-3; testCompile-native-3, then ci-jvm-3, test-js-3, test-native-3 (aliases). Native tests link: a suite that compiles can still fail to link, so the Native TEST run is the gate for injected code. Fixture differences between rows are real (DummyTextureData.type vs dataType) — fix the fixture, never skip the row.

5. Behaviour: java wins, per platform

A Native suite expecting a different exception than java (FileReadError vs GdxRuntimeException) is the hand port's divergence; the port keeps java's behaviour and the test is adapted, recorded as a finding. JVM-only fallbacks (classpath resources) stay counted on JS.