diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 3aee47df..cff229ef 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -126,6 +126,13 @@ jobs:
           CODING_SH_REQUIRE_DIST: "1"
         run: npm run check
 
+      # The Site workflow runs this on main and stops the deploy when a page is over budget.
+      # Running it here, on the same OS and Node as that workflow, fails the pull request instead.
+      - name: Site perf budget
+        if: matrix.shard == 1 && runner.os == 'Linux' && matrix.node == '24.x'
+        working-directory: apps/site
+        run: npm run perf:budget
+
       # The built output holds nothing shaped like a secret.
       - name: Secret scan of the built output
         if: matrix.shard == 1
diff --git a/apps/site/scripts/perf-budget.mjs b/apps/site/scripts/perf-budget.mjs
index 6a991768..408593ea 100644
--- a/apps/site/scripts/perf-budget.mjs
+++ b/apps/site/scripts/perf-budget.mjs
@@ -21,11 +21,12 @@ const BUDGET = {
   // troubleshooting, which gains an entry per failure a new feature can show
   // (27.8KB with the 0.1.6 entries). Raised from 26KB for those three.
   // Troubleshooting is 30.5KB after the docs rebuild of 2026-11-11, so 33KB.
+  // /docs/api/read is 30.3KB with the weight edit and checkpoint fields, so 33KB.
   htmlBrotli: {
     "/": 24.5 * KB,
     "/docs/cli/": 30 * KB,
     "/docs/api/": 30 * KB,
-    "/docs/api/read/": 30 * KB,
+    "/docs/api/read/": 33 * KB,
     "/docs/api/read/search/": 30 * KB,
     "/docs/troubleshooting/": 33 * KB,
     other: 26 * KB,
diff --git a/docs/perf.md b/docs/perf.md
index 86168b5d..95d32d37 100644
--- a/docs/perf.md
+++ b/docs/perf.md
@@ -125,7 +125,7 @@ page has its own blocking-time budget on the slow phone. The rules are in
 |---|---|
 | Pull requests | GitHub's hosted runners. Load-time budgets are skipped: they only hold on our own machine. |
 | Pushes to main | The one desk runner labelled `perf`, so two perf runs never overlap. CI and deploys run beside it. |
-| Deploys | `npm run perf:budget` in apps/site checks page weight, inline CSS and script, fonts, scripts and images in the build. |
+| Pull requests and deploys | `npm run perf:budget` in apps/site checks page weight, inline CSS and script, fonts, scripts and images in the build. CI runs it on every pull request, and the Site workflow runs it again before a deploy. |
 
 The desk runs other jobs, so each sample records `cpu_busy`, the share of CPU in use while the
 page loaded. A load-time budget counts only samples at 0.5 or less (`--quiet-cpu`). When fewer