[ASDAlexander77]

Closes the last two open items in the own markers spec.

Ternary of an array literal or null. return i < 3 ? ["a"] : null failed MLIR verification ("operand #0 does not dominate this use") under every model, gc included. The cause:

• The ternary's type was const_array<string,1> | null.
• The generated ___cast function that makes it a string[] | null narrows t in its else branch. Narrowing a const_array gives the narrowed name storage of its own: a heap copy that it owns.
• That copy was released at the end of the function, outside the branch it was made in.

Now a literal array in a branch of ?: is widened to an array in that branch, as TypeScript types it. Each branch also hands its array out with a reference of its own, as new C() is (Retain plus __owned_result). That lets consumeConditionalResults settle the merge, so -mm=own accepts array ternaries. Before, own rejected even c ? ["a"] : ["b"]. rc passes --verify-ownership on the new test.

if (a) of a string[] | null didn't narrow a, because truthiness narrowing required a tagged union. An array keeps null in its own pointer, so a[0] indexed the union. Truthiness narrowing now also covers untagged unions of arrays and null. It deliberately leaves Shared<T> | null alone: own reads a counted handle through the variable itself, and own_shared_graph breaks otherwise.

Test: tests/00narrow_array_null.ts, under gc, rc, none and own, compiled and JIT. It covers ternaries over if, while and for, and push after churn.

Android own. android-link.cmake now links an executable and a shared library under own as well as gc. This needs the own Android default library from ASDAlexander77/TypeScriptCompilerDefaultLib#38. I ran an own executable on the API 29 x86_64 emulator: matchAll, padStart, repeat, arrays and classes all gave correct output. The README and the spec are updated.

Full suite: 4834/4834 (Release, Windows).

🤖 Generated with Claude Code