diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 73a38b4..328c5ef 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -105,6 +105,22 @@ jobs:
         run: echo "out=$(nix build --impure .#${{ matrix.target }} --print-build-logs --no-link --print-out-paths)" >> "$GITHUB_OUTPUT"
       - name: Put bin on PATH
         run: echo "${{ steps.build.outputs.out }}/bin" >> "$GITHUB_PATH"
+      # Confirmed root cause of CI's inflated startup numbers below (see
+      # PR #44): without this cache, plugin files land fresh from a network
+      # git clone on every run, and each cold file read shows up inside
+      # nvim's own require()/sourcing timings (--startuptime measures the
+      # whole call including the read, not just execution). That's what was
+      # producing 100-600ms readings vs. ~25ms locally for the identical
+      # build -- not a slow plugin, not CPU contention, not codesign/
+      # Gatekeeper overhead (all measured and ruled out, see below). Adding
+      # this cache alone took every matrix entry down to ~20-27ms, matching
+      # local numbers almost exactly. Keyed on runs-on because
+      # telescope-fzf-native's native build output is arch-specific.
+      - name: Cache lazy.nvim plugins
+        uses: actions/cache@v4
+        with:
+          path: ~/.local/share/nvim/lazy
+          key: lazy-${{ matrix.runs-on }}-${{ hashFiles('lazy-lock.json') }}
       - name: Restore plugins
         run: nvim --headless "+Lazy! restore" +qa
       - name: Run smoke test
@@ -116,11 +132,22 @@ jobs:
       # only ~2-3 minutes wall clock with plenty of concurrency headroom
       # (public repo, at most 3 test-matrix jobs running at once), so there's
       # no runtime pressure pushing this out into its own workflow.
+      #
+      # nvim's own --startuptime only instruments Lua/vimscript execution --
+      # it can't see process spawn or dynamic-linker/codesign overhead before
+      # that. Wrapping with `time` for an outside-in wall-clock figure ruled
+      # that out as the main story (gap vs. nvim's own reported mean was only
+      # ~5-40ms/run, largest on macOS as expected from Gatekeeper/AMFI, but
+      # nowhere near what was inflating CI's numbers before the plugin cache
+      # above). Kept as a standing signal: if a future regression shows up
+      # here instead of in nvim's own mean, that's process-launch overhead,
+      # not the config's fault.
       - name: Benchmark startup time
         run: |
           {
             echo "### Startup time — ${{ matrix.name }}"
             echo '```'
-            ./scripts/benchmark-startup 10
+            TIMEFORMAT='wall clock for all 10 runs: %R s (user %U s, sys %S s)'
+            { time ./scripts/benchmark-startup 10; } 2>&1
             echo '```'
           } >> "$GITHUB_STEP_SUMMARY"
diff --git a/scripts/benchmark-startup b/scripts/benchmark-startup
index 58b1315..1cd66af 100755
--- a/scripts/benchmark-startup
+++ b/scripts/benchmark-startup
@@ -17,6 +17,8 @@ for ((i = 1; i <= RUNS; i++)); do
   times+=("$(awk 'NF { last = $1 } END { print last }' "$LOG")")
 done
 
+echo "per run (chronological, ms): ${times[*]}"
+
 printf '%s\n' "${times[@]}" | sort -n | awk '
   { a[NR] = $1; sum += $1 }
   END {