diff --git a/crates/engine/src/cmd/diagnostics.rs b/crates/engine/src/cmd/diagnostics.rs
index b31a65619f..78660e0904 100644
--- a/crates/engine/src/cmd/diagnostics.rs
+++ b/crates/engine/src/cmd/diagnostics.rs
@@ -78,7 +78,10 @@ pub(crate) fn kernel_asked(choice: &Value, built: bool) -> Result<solidcraft_ker
     use solidcraft_kernel::kernel_route::Backend;
     let asked = choice.as_str().and_then(Backend::from_id).ok_or_else(|| bad("diagnostics.kernel", "`choice` is solidcore or smlib"))?;
     if asked == Backend::Smlib && !built {
-        return Err(EngineError::Disabled("diagnostics.kernel".into(), "this build doesn't include the C++ SMLib kernel; build with --features smlib".into()));
+        return Err(EngineError::Disabled(
+            "diagnostics.kernel".into(),
+            "this build doesn't include the C++ SMLib kernel; build with --features smlib".into(),
+        ));
     }
     Ok(asked)
 }
diff --git a/crates/kernel/src/step_in/tests.rs b/crates/kernel/src/step_in/tests.rs
index 043c639688..a86bac8d31 100644
--- a/crates/kernel/src/step_in/tests.rs
+++ b/crates/kernel/src/step_in/tests.rs
@@ -540,12 +540,18 @@ fn offset_surface_cube_import_is_closed_and_editable() {
     assert!(rel(measure(&result).unwrap().volume, 1000.0 - 40.0 * std::f64::consts::PI) < 1e-3);
 }
 
-/// A solid whose B-rep SMLib refuses, its faces too (here every SMLib call: the build has no
-/// SMLib), is refused with the kernel's reason, naming the solid. It once came in as a mesh of
-/// its faces, which the planner could still triangulate.
+/// A solid whose B-rep SMLib refuses, its faces too (here every SMLib call: the C++ kernel is asked
+/// for, SOLIDCRAFT_KERNEL=smlib, on a build without it), is refused with the kernel's reason, naming
+/// the solid. It once came in as a mesh of its faces, which the planner could still triangulate.
+/// On solidcore, the default, the cube imports, so there is nothing to refuse.
 #[cfg(not(feature = "smlib"))]
 #[test]
 fn a_solid_smlib_refuses_with_its_faces_is_refused_not_meshed() {
+    use crate::kernel_route::{Backend, choice};
+    if choice() != Backend::Smlib {
+        eprintln!("skipped: needs SOLIDCRAFT_KERNEL=smlib on a build without the C++ kernel");
+        return;
+    }
     let text = format!("{HEAD}{}#60=SHAPE_REPRESENTATION('',(#101),$);\n{TAIL}", cube_entities());
     let why = step_import(&text).map(|imp| imp.warnings).unwrap_err().to_string();
     assert!(why.contains("solid 'Cube': not imported: ") && why.contains("SMLib"), "{why}");
diff --git a/crates/smlib/src/core/blend_tests.rs b/crates/smlib/src/core/blend_tests.rs
index dade9f8168..e03e1e2312 100644
--- a/crates/smlib/src/core/blend_tests.rs
+++ b/crates/smlib/src/core/blend_tests.rs
@@ -929,10 +929,22 @@ fn edges_871() -> Vec<[f64; 3]> {
 /// #871's rounds at r 0.5 fit (the wall between the window and the block's side is 1.26 thick):
 /// 38 faces, and the closed form `issue_871_tests::rounds_that_fit_are_closed` derives. At r 1 they
 /// overlap, which is refused. With SMLib built, the C++ kernel gives the same.
+#[cfg(feature = "native")]
 #[test]
 fn rounds_on_the_871_part() {
-    #[cfg(feature = "native")]
     let _cpp = crate::route::on_cpp();
+    rounds_on_the_871_part_check();
+}
+
+/// The same on a build without the C++ kernel: solidcore's half alone. A test of its own, so that
+/// tests/KNOWN_FAILING_TESTS's entry for the C++ comparison (rounds_on_the_871_part) is graded only where it runs.
+#[cfg(not(feature = "native"))]
+#[test]
+fn rounds_on_the_871_part_on_solidcore() {
+    rounds_on_the_871_part_check();
+}
+
+fn rounds_on_the_871_part_check() {
     let p = part_871();
     let e = edges_871();
     let sizes = |r: f64| vec![r; e.len()];
diff --git a/crates/smlib/src/core/offset_tests.rs b/crates/smlib/src/core/offset_tests.rs
index 901a673d6a..6241ed1094 100644
--- a/crates/smlib/src/core/offset_tests.rs
+++ b/crates/smlib/src/core/offset_tests.rs
@@ -840,10 +840,22 @@ fn offset_faces_close_a_step_wall() {
 /// solidcore refuses to put the rounds back (solidcore#2020, open): SolidCraft says it does not
 /// make that move, where the C++ kernel gives the analytic value, also checked here. With SMLib
 /// built, the C++ kernel gives the same at every other move.
+#[cfg(feature = "native")]
 #[test]
 fn mitred_rounds_follow_a_moved_top() {
-    #[cfg(feature = "native")]
     let _cpp = crate::route::on_cpp();
+    mitred_rounds_follow_a_moved_top_check();
+}
+
+/// The same on a build without the C++ kernel: solidcore's half alone. A test of its own, so that
+/// tests/KNOWN_FAILING_TESTS's entry for the C++ comparison (mitred_rounds_follow_a_moved_top) is graded only where it runs.
+#[cfg(not(feature = "native"))]
+#[test]
+fn mitred_rounds_follow_a_moved_top_on_solidcore() {
+    mitred_rounds_follow_a_moved_top_check();
+}
+
+fn mitred_rounds_follow_a_moved_top_check() {
     use super::blend::{Corners, blend};
     use crate::Blend;
     let edges = [[20.0, 0.0, 10.0], [20.0, 30.0, 10.0], [0.0, 15.0, 10.0], [40.0, 15.0, 10.0]];
diff --git a/xtask/src/kernel_cases.rs b/xtask/src/kernel_cases.rs
index 5efcd9f823..35f3d66e06 100644
--- a/xtask/src/kernel_cases.rs
+++ b/xtask/src/kernel_cases.rs
@@ -229,6 +229,15 @@ pub(crate) fn run_all(o: &Options, list: &[PathBuf]) -> Vec<Value> {
 
 /// Every case on each backend, one job pool for all the runs; the results per backend, in order.
 fn run_on(o: &Options, list: &[PathBuf], kernels: &[&str]) -> Vec<Vec<Value>> {
+    // Without --cli each case is `cargo run`: build the CLI once first, so the first cases don't
+    // spend their timeout compiling it (the C++ kernel's build most of all).
+    if o.cli.is_none() {
+        let mut c = cargo();
+        c.args(["build", "-q", "--release", "-p", "solidcraft-cli"]).args(cli_features(kernels));
+        if let Err(e) = c.status().map_err(|e| e.to_string()).and_then(|s| if s.success() { Ok(()) } else { Err(s.to_string()) }) {
+            eprintln!("kernel-cases: building solidcraft-cli failed ({e}); each case will say why");
+        }
+    }
     let jobs = list.len() * kernels.len();
     let next = Arc::new(Mutex::new(0usize));
     let results = Arc::new(Mutex::new(vec![Value::Null; jobs]));