test(core): gate u8 branch tests by backend types

crates/cubecl-core/src/runtime_tests/branch.rs

@@ -283,29 +283,36 @@ macro_rules! testgen_branch {
         }
 
         #[$crate::runtime_tests::test_log::test]
-        fn test_if_literal_u8_true() {
+        fn test_switch_const() {
             let client = TestRuntime::client(&Default::default());
-            cubecl_core::runtime_tests::branch::test_if_literal_u8::<TestRuntime>(client, true);
+            cubecl_core::runtime_tests::branch::test_switch_const::<TestRuntime, FloatType>(client);
         }
 
         #[$crate::runtime_tests::test_log::test]
-        fn test_if_literal_u8_false() {
+        fn test_for_loop_with_break() {
             let client = TestRuntime::client(&Default::default());
-            cubecl_core::runtime_tests::branch::test_if_literal_u8::<TestRuntime>(client, false);
+            cubecl_core::runtime_tests::branch::test_for_loop_with_break::<TestRuntime, FloatType>(
+                client,
+            );
         }
+    };
+}
 
+#[allow(missing_docs)]
+#[macro_export]
+macro_rules! testgen_branch_u8 {
+    (u8) => {
         #[$crate::runtime_tests::test_log::test]
-        fn test_switch_const() {
+        fn test_if_literal_u8_true() {
             let client = TestRuntime::client(&Default::default());
-            cubecl_core::runtime_tests::branch::test_switch_const::<TestRuntime, FloatType>(client);
+            cubecl_core::runtime_tests::branch::test_if_literal_u8::<TestRuntime>(client, true);
         }
 
         #[$crate::runtime_tests::test_log::test]
-        fn test_for_loop_with_break() {
+        fn test_if_literal_u8_false() {
             let client = TestRuntime::client(&Default::default());
-            cubecl_core::runtime_tests::branch::test_for_loop_with_break::<TestRuntime, FloatType>(
-                client,
-            );
+            cubecl_core::runtime_tests::branch::test_if_literal_u8::<TestRuntime>(client, false);
         }
     };
+    ($uint:ident) => {};
 }

crates/cubecl-core/src/runtime_tests/mod.rs

@@ -83,6 +83,7 @@ macro_rules! testgen_all {
                 type UintType = $uint;
 
                 $crate::testgen_uint!();
+                $crate::testgen_branch_u8!($uint);
             })*
         }
         $crate::testgen_untyped!();

docs/superpowers/plans/2026-07-19-gate-u8-branch-tests.md

@@ -0,0 +1,62 @@
+# Gate `u8` Branch Runtime Tests Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Generate the `u8` branch runtime tests only for backends whose existing test type list includes `u8`.
+
+**Architecture:** Add a token-matching test macro beside the branch tests and invoke it from unsigned-type modules generated by `testgen_all!`. The backend type lists remain the single capability declaration.
+
+**Tech Stack:** Rust, declarative macros, Cargo tests, CubeCL WGPU runtime tests.
+
+## Global Constraints
+
+- Do not implement or emulate `u8` in WGSL.
+- Preserve the two runtime assertions on backends that list `u8`.
+- Keep the fork delta backend-agnostic and minimal.
+
+---
+
+### Task 1: Capability-gated test generation
+
+**Files:**
+- Modify: `crates/cubecl-core/src/runtime_tests/branch.rs`
+- Modify: `crates/cubecl-core/src/runtime_tests/mod.rs`
+
+**Interfaces:**
+- Consumes: the unsigned type tokens already passed to `testgen_all!`.
+- Produces: `testgen_branch_u8!(type_token)`, generating tests only for `u8`.
+
+- [ ] **Step 1: Verify the existing regression test fails**
+
+Run: `cargo test -p t4a-cubecl-wgpu --lib test_if_literal_u8 -- --nocapture`
+
+Expected: four failures ending in `U8 is not a valid WgpuElement`.
+
+- [ ] **Step 2: Add the selective macro**
+
+Move `test_if_literal_u8_true` and `test_if_literal_u8_false` out of
+`testgen_branch!` and into `testgen_branch_u8!`, with a `(u8)` arm containing
+the tests and a catch-all arm expanding to nothing.
+
+- [ ] **Step 3: Connect it to unsigned type generation**
+
+After each `$crate::testgen_uint!()` invocation inside the parameterized
+`testgen_all!`, add `$crate::testgen_branch_u8!($uint);`.
+
+- [ ] **Step 4: Verify the focused WGPU regression**
+
+Run: `cargo test -p t4a-cubecl-wgpu --lib test_if_literal_u8 -- --nocapture`
+
+Expected: zero matching tests, exit status 0.
+
+- [ ] **Step 5: Verify formatting, compilation, and the WGPU library suite**
+
+Run `cargo fmt --check`, `cargo check -p t4a-cubecl-core`, and
+`cargo test -p t4a-cubecl-wgpu --lib`.
+
+Expected: all commands exit with status 0.
+
+- [ ] **Step 6: Commit**
+
+Stage only the two source files and these design documents, then commit as
+`test(core): gate u8 branch tests by backend types`.

docs/superpowers/specs/2026-07-19-gate-u8-branch-tests-design.md

@@ -0,0 +1,28 @@
+# Gate `u8` Branch Runtime Tests by Supported Types
+
+## Problem
+
+`testgen_branch!` generates the `u8` literal branch tests once for every
+floating-point test module. WGSL does not support `u8`, so the WGPU test suite
+generates four runtime tests that can only panic in the WGSL compiler.
+
+## Design
+
+Keep the shared `u8` kernel and runtime assertion, but move test generation to
+a new helper macro that accepts one unsigned type token. The macro generates
+the two tests only for the `u8` token and expands to nothing for other unsigned
+types. `testgen_all!` invokes this helper from each generated unsigned-type
+module.
+
+This uses each backend's existing unsigned-type list as the capability source:
+CPU, CUDA, and SPIR-V retain one true/false pair under `u8_ty`, while WGSL and
+MSL generate no `u8` runtime tests. No backend-specific conditionals or `u8`
+emulation are introduced.
+
+## Verification
+
+- Before the change, the focused WGPU command must reproduce four failures.
+- After the change, the same filter must find zero WGPU tests and exit cleanly.
+- Macro expansion must compile for both a list containing `u8` and a list that
+  omits it.
+- The full WGPU library suite must pass on the local Metal backend.