diff --git a/.github/scripts/README.md b/.github/scripts/README.md
index d13e3185..36298667 100644
--- a/.github/scripts/README.md
+++ b/.github/scripts/README.md
@@ -7,6 +7,11 @@ The read-only hosted Linux job tests the PR merge result and records both head
 and checkout identities. No PR-target execution or signing/deployment credentials
 are used. The 150-minute cold-build limit is provisional for this new job layout.
 
+The CI environment disables release-profile LTO and uses 16 codegen units for
+`cargo xtask ci`. Tests retain release optimization and the existing gates;
+release packaging keeps the manifest's ThinLTO/single-unit configuration. The first
+hosted runs exhausted 150 minutes compiling test targets before any tests ran.
+
 The runner installs Mesa Lavapipe and probes its Vulkan adapter. A failed probe
 does not prevent CPU gates from running, but fails the final coverage report.
 `RUST_TEST_NOCAPTURE=1` exposes the existing GPU tests' adapter/parity/skip messages.
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index f15ecf83..28e82e67 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -15,6 +15,9 @@ env:
   CARGO_TERM_COLOR: always
   CARGO_BUILD_JOBS: "2"
   CARGO_INCREMENTAL: "0"
+  # Keep optimized tests; avoid release-package LTO costs for every test executable.
+  CARGO_PROFILE_RELEASE_LTO: "off"
+  CARGO_PROFILE_RELEASE_CODEGEN_UNITS: "16"
   RUST_BACKTRACE: "1"
   RUST_TEST_THREADS: "2"
   # Passing GPU tests must expose adapter/skip evidence instead of hiding it in captured output.