diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index dfa28c9f0..f9d0099b1 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -10,7 +10,7 @@ You need [Rust](https://rustup.rs/) 1.95 or newer.
 ```sh
 cargo run -p effectcraft -- --demo       # the desktop app, with the demo project
 cargo test --workspace
-cargo xtask ci                           # what every commit must pass
+cargo xtask ci                           # what every commit must pass (tests use --profile ci)
 ```
 
 ## Ground rules
diff --git a/Cargo.toml b/Cargo.toml
index 38ba10a0e..919b55d03 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -204,3 +204,10 @@ opt-level = 3
 [profile.release]
 lto = "thin"
 codegen-units = 1
+
+# `cargo xtask ci` builds clippy and tests with this profile. Tests don't need shipping-grade LTO:
+# thin LTO + one codegen unit across ~69 test binaries made the gate ~1.5 h while the tests run in ~75 s.
+[profile.ci]
+inherits = "release"
+lto = false
+codegen-units = 16
diff --git a/crates/engine/src/tests_settings.rs b/crates/engine/src/tests_settings.rs
index 1a445205b..4e1a8e391 100644
--- a/crates/engine/src/tests_settings.rs
+++ b/crates/engine/src/tests_settings.rs
@@ -545,11 +545,13 @@ fn open_recent_increment_and_revert() {
 fn memory_tick_never_waits_for_the_system() {
     let mut s = Session::default();
     let t0 = std::time::Instant::now();
-    s.memory_tick();
-    s.memory_tick();
+    for _ in 0..8 {
+        s.memory_tick();
+    }
     // Reading the system's memory starts PowerShell on Windows (about a second): the UI thread
-    // only takes the last reading.
-    assert!(t0.elapsed() < std::time::Duration::from_millis(200), "{:?}", t0.elapsed());
+    // only takes the last reading. Eight blocking ticks would take ~8 s; the bound is loose so a
+    // loaded machine doesn't flake it.
+    assert!(t0.elapsed() < std::time::Duration::from_secs(4), "{:?}", t0.elapsed());
     if crate::sysinfo::memory().is_some() {
         let mut got = None;
         for _ in 0..200 {
diff --git a/crates/gpu/src/tests_init.rs b/crates/gpu/src/tests_init.rs
index 9c579ae69..30a9ba6da 100644
--- a/crates/gpu/src/tests_init.rs
+++ b/crates/gpu/src/tests_init.rs
@@ -146,7 +146,8 @@ fn opengl_adapters_get_cpu_compositing_at_once() {
     let t = std::time::Instant::now();
     let error = crate::Gpu::new(&adapter, device, queue).err().expect("no GPU compositor on OpenGL");
     assert!(error.contains("(Gl)"), "{error}");
-    assert!(t.elapsed() < std::time::Duration::from_secs(1), "{:?}", t.elapsed());
+    // Generous bound (the failure mode took tens of seconds) so a loaded machine doesn't flake it.
+    assert!(t.elapsed() < std::time::Duration::from_secs(10), "{:?}", t.elapsed());
 }
 
 #[test]
diff --git a/crates/script/src/tests_ui_more.rs b/crates/script/src/tests_ui_more.rs
index 228a58016..87ff00e34 100644
--- a/crates/script/src/tests_ui_more.rs
+++ b/crates/script/src/tests_ui_more.rs
@@ -441,5 +441,6 @@ fn socket_edge_cases() {
         "net.jsx",
     );
     assert_eq!(o.result, json!(0), "{:?}", o.error);
-    assert!(t.elapsed().as_secs_f64() < 5.0);
+    // Doesn't hang; generous so a loaded machine doesn't flake it.
+    assert!(t.elapsed().as_secs_f64() < 30.0, "{:?}", t.elapsed());
 }
diff --git a/docs/testing.md b/docs/testing.md
index b053a3e8e..3c81c5524 100644
--- a/docs/testing.md
+++ b/docs/testing.md
@@ -51,6 +51,9 @@ Limits: a crate that reads a repo file at run time through a path that is not a
 sources is not detected, and building a subset of packages can unify features differently from the
 whole workspace. Run the full `cargo xtask ci` before a release.
 
+`xtask ci` builds clippy and tests with `--profile ci` (release optimisation without LTO, 16 codegen
+units, output in `target/ci/`): fast to build, still optimised enough for the pixel and perf tests.
+
 ## What the tests cover
 
 - **Engine commands** (`crates/engine/src/tests*.rs`): every command family, including undo and
diff --git a/xtask/src/main.rs b/xtask/src/main.rs
index ac1283408..0f422a493 100644
--- a/xtask/src/main.rs
+++ b/xtask/src/main.rs
@@ -9,7 +9,7 @@
 //! - `assets`: every asset file (image, icon, font, LUT, audio, video…) has a complete
 //!   `<file>.attribution` sidecar and an entry in `ATTRIBUTION.md` (AGENTS.md §1).
 //! - `ico <out.ico> <in.png>…`: pack PNGs into a Windows `.ico` (used by `packaging/icons.sh`).
-//! - `ci`: fmt check, clippy -D warnings, tests, layers, assets, wasm.
+//! - `ci`: fmt check, clippy -D warnings, tests (`--profile ci`), layers, assets, wasm.
 //! - `ci --affected [--base REV] [--dry-run]`: the same gates for only the crates the change since
 //!   REV (default: the merge-base with `main`) can affect; see `affected.rs` and docs/testing.md.
 
@@ -408,8 +408,8 @@ fn serve_dir(dir: &std::path::Path, port: u16) -> Result<(), String> {
 fn ci() -> Result<(), String> {
     let cargo = env!("CARGO");
     run(Command::new(cargo).args(["fmt", "--check"]))?;
-    run(Command::new(cargo).args(["clippy", "--workspace", "--all-targets", "--release", "--", "-D", "warnings"]))?;
-    run(Command::new(cargo).args(["test", "--workspace", "--release"]))?;
+    run(Command::new(cargo).args(["clippy", "--workspace", "--all-targets", "--profile", "ci", "--", "-D", "warnings"]))?;
+    run(Command::new(cargo).args(["test", "--workspace", "--profile", "ci"]))?;
     layers()?;
     assets()?;
     wasm()