diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index f86899a..16059e7 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -239,7 +239,9 @@ jobs:
   # shared one, cargo's mtime check can reuse one tree's build for the other
   # and compare a binary against itself. Both trees build the PR's copy of
   # the example, so both sides run one workload; a PR whose example calls
-  # API the base lacks cannot be compared and reads INCOMPLETE.
+  # API the base lacks cannot be compared and reads INCOMPLETE. Both build
+  # with `--features macros`, so the cases behind it, the `#[cachekit]` path,
+  # are measured too.
   perf-ir:
     name: Instruction counts vs base
     if: github.event_name == 'pull_request'
@@ -320,8 +322,8 @@ jobs:
         env:
           BASE_LABEL: ${{ steps.base.outputs.label }}
         run: |
-          (cd "$RUNNER_TEMP/ir/head" && CARGO_TARGET_DIR=../head-target cargo build --locked --release --example bench_hot_path)
-          (cd "$RUNNER_TEMP/ir/base" && CARGO_TARGET_DIR=../base-target cargo build --locked --release --example bench_hot_path) || {
+          (cd "$RUNNER_TEMP/ir/head" && CARGO_TARGET_DIR=../head-target cargo build --locked --release --features macros --example bench_hot_path)
+          (cd "$RUNNER_TEMP/ir/base" && CARGO_TARGET_DIR=../base-target cargo build --locked --release --features macros --example bench_hot_path) || {
             {
               echo '### Instruction counts: INCOMPLETE'
               echo
diff --git a/README.md b/README.md
index 31942cb..d0fccf7 100644
--- a/README.md
+++ b/README.md
@@ -924,9 +924,10 @@ reports either one as exit 2.
 
 CI runs this comparison on every pull request (the **Instruction counts vs
 base** check): it builds the pull request and the commit it merges into, each
-in its own target directory, measures both on the same runner, and posts the
-verdict as one comment it edits in place (a pull request from a fork gets no
-comment; its verdict is in the run summary). A case it did not measure reads
+in its own target directory and with `--features macros` (the `#[cachekit]`
+cases need it), measures both on the same runner, and posts the verdict as
+one comment it edits in place (a pull request from a fork gets no comment; its
+verdict is in the run summary). A case it did not measure reads
 INCOMPLETE, never PASS. A release pull request is measured against the last
 release tag instead. The check is not required: a deliberate cost increase is
 accepted at review.