diff --git a/.github/workflows/lighthouse.yml b/.github/workflows/lighthouse.yml
index 5e01999..3b51d83 100644
--- a/.github/workflows/lighthouse.yml
+++ b/.github/workflows/lighthouse.yml
@@ -1,13 +1,11 @@
 name: Lighthouse
 
 on:
+  # Every pull request starts this workflow, and the `changes` job skips the
+  # audit when only bench changed. A skipped job reports success, so "Lighthouse
+  # audit" can be a required check; a workflow filtered out by `paths` would
+  # never report and would block the merge.
   pull_request:
-    # Lighthouse audits the marketing site only. It isn't a required check, so
-    # a workflow-level filter is enough here; ci.yml gates its required job with
-    # the changed-paths action instead.
-    paths-ignore:
-      - 'bench/**'
-      - '.github/workflows/bench-*.yml'
   workflow_dispatch:
 
 permissions:
@@ -21,10 +19,29 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
+  changes:
+    name: Detect changed paths
+    runs-on: ubuntu-24.04
+    timeout-minutes: 5
+    outputs:
+      site: ${{ steps.paths.outputs.site }}
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
+        with:
+          fetch-depth: 2
+          persist-credentials: false
+      - id: paths
+        uses: ./.github/actions/changed-paths
+
   lighthouse:
     name: Lighthouse audit
+    needs: changes
+    # Lighthouse audits the marketing site only. Skipped, which counts as
+    # passing, when nothing outside bench/ changed. If the path check itself
+    # fails, run anyway rather than skip.
+    if: ${{ !cancelled() && (needs.changes.result != 'success' || needs.changes.outputs.site == 'true') }}
     runs-on: ubuntu-24.04
-    timeout-minutes: 15
+    timeout-minutes: 20
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1  # v7.0.1
 
@@ -81,9 +98,16 @@ jobs:
           temporaryPublicStorage: true
 
       # Build a markdown table from the .lighthouseci/lhr-*.json reports.
-      # `unique_by(.url)` collapses duplicate rows if `numberOfRuns > 1`
-      # (the action writes the median LHR per URL, but be defensive);
-      # `sort_by(.url)` keeps row order deterministic across runs.
+      # lighthouserc.cjs collects three runs per URL and asserts on the best
+      # one, so the table shows each URL's best run by performance, the run
+      # the check is judged by. `group_by(.url)` also keeps row order
+      # deterministic. The linked report is the one the action uploads for
+      # that URL and may be a different run.
+      #
+      # Fork PRs, and Dependabot's PRs from this repository, get a read-only
+      # GITHUB_TOKEN that can't post comments. A failed comment must not fail
+      # this required check, so only same-repository, non-Dependabot PRs get
+      # one, and both comment steps are non-fatal for any other token limit.
       # `if: always()` so the PR comment still posts when assertions fail —
       # reviewers can see *which* category regressed without digging through
       # the workflow log.
@@ -93,8 +117,9 @@ jobs:
       # the 🟢 cutoff, so a cell can pass assertions and still show 🟡 — the
       # visual signals "we locked this in but it's improvable."
       - name: Format PR comment
-        if: always() && github.event_name == 'pull_request' && steps.lighthouse.outputs.links != ''
+        if: always() && github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository && github.actor != 'dependabot[bot]' && steps.lighthouse.outputs.links != ''
         id: lh-comment
+        continue-on-error: true
         env:
           LINKS_JSON: ${{ steps.lighthouse.outputs.links }}
         run: |
@@ -114,8 +139,8 @@ jobs:
                 bp: (.categories["best-practices"].score * 100 | floor),
                 seo: (.categories.seo.score * 100 | floor)
               })
-              | unique_by(.url)
-              | sort_by(.url)
+              | group_by(.url)
+              | map(max_by(.perf))
               | map(
                   "| `" + (.url | sub("^https?://[^/]+"; "")) + "`"
                   + " | " + cell(.perf)
@@ -133,6 +158,7 @@ jobs:
 
       - name: Post or update PR comment
         if: github.event_name == 'pull_request' && steps.lh-comment.outputs.body != ''
+        continue-on-error: true
         uses: marocchino/sticky-pull-request-comment@5770ad5eb8f42dd2c4f34da00c94c5381e49af88  # v3.0.5
         with:
           header: lighthouse
diff --git a/CLAUDE.md b/CLAUDE.md
index b47bb02..5eb4812 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -21,7 +21,7 @@ The repository holds two independent apps:
 They share one Bun workspace and the CI policy, and deploy separately:
 
 - **Toolchain.** `bench/web` is a member of the root Bun workspace: one `bun.lock`, one `bunfig.toml`, one `bun install` at the root. Bun's isolated linker gives each package its own dependencies; both apps run Next.js 16.4.1, React 19.3 and TypeScript 7. The root `biome.json` lints and formats both apps: a `bench/**` override keeps bench's old Prettier style (single quotes, trailing commas, 100 columns) and sets bench's lint rules (see `bench/AGENTS.md`), and `bun run check:ci` at the root checks all of it. Both `tsconfig.json` files extend the root `tsconfig.base.json`, which holds the compiler options the apps share, including `noUncheckedIndexedAccess` and `exactOptionalPropertyTypes`. Each app keeps its own target, paths and extra flags (bench's `noFallthroughCasesInSwitch` and `noUncheckedSideEffectImports`). The root `tsconfig.json` excludes `bench/`, and `bench/web/postcss.config.mjs` keeps Next.js from picking up the root's Tailwind config.
-- **CI.** `.github/actions/changed-paths` reports which app a change touches. `ci.yml` skips its site jobs on bench-only changes, and `bench-ci.yml` and `bench-deploy.yml` skip on site-only changes. The workspace's root files (`package.json`, `bun.lock`, `bunfig.toml`, `biome.json`, `tsconfig*.json`) count as both. They gate pull requests with `if:`, not `on.paths`, so a skipped required check still passes. `bench-deploy.yml` keeps `on.paths` for pushes to main, because a push run that only skips would still queue in its production concurrency group. Dependency review, the secret scan and REUSE run on every change.
+- **CI.** `.github/actions/changed-paths` reports which app a change touches. `ci.yml` and `lighthouse.yml` skip their site jobs on bench-only changes, and `bench-ci.yml` and `bench-deploy.yml` skip on site-only changes. The workspace's root files (`package.json`, `bun.lock`, `bunfig.toml`, `biome.json`, `tsconfig*.json`) count as both. They gate pull requests with `if:`, not `on.paths`, so a skipped required check still passes. `bench-deploy.yml` keeps `on.paths` for pushes to main, because a push run that only skips would still queue in its production concurrency group. Dependency review, the secret scan and REUSE run on every change.
 - **Deploys.** The `vortex` Vercel project builds the site through the git integration, and `scripts/vercel-ignore-build.sh` skips bench-only commits. bench deploys to the `benchmarks-website` Vercel project (Root Directory `bench/web`) from `bench-deploy.yml`, with the Vercel CLI, on changes to `bench/` and to the root `package.json`, `bun.lock`, `bunfig.toml` and `tsconfig.base.json`. Schema migrations go through `bench-schema-deploy.yml`, which only runs manually from `main` because the AWS role trusts nothing else.
 
 ## Architecture Overview
@@ -100,7 +100,7 @@ Three places must agree with each other and with what Vercel runs. Vercel offers
 - `@types/node` → `^26`, matching the runtime so `tsc` cannot green-light an API that production does not have.
 
 Keep them in lockstep when bumping, and move to a new major only after Vercel offers it; check https://vercel.com/docs/functions/runtimes/node-js/node-js-versions first.
-10. **Renovate auto-merge** is patch + minor only via `:automergeStableNonMajor`. Major bumps stay open for human review. The 7-day cooldown is the first gate, and the required checks on `main` are the second. Renovate merges through GitHub's auto-merge, which waits only for required checks: `Lint, build, verify` (lint, typecheck, build, `bun run verify` and Playwright), `dependency-review`, `secret-scan`, `DCO`, `REUSE compliance`, `Submit bun.lock`, `Bench: format, lint, build, test`, `Bench: deploy` and `Vercel`, plus a CodeQL code scanning rule. Lighthouse and `bun audit` are not required, so they don't hold an auto-merge. Lock file maintenance automerges too. Renovate cannot apply `minimumReleaseAge` to a Bun lock refresh, so `renovate.json` turns it off for that update type; `bunfig.toml` enforces the same wait while Bun regenerates `bun.lock`. `biome.json` takes its `$schema` from the installed `@biomejs/biome` package rather than a versioned URL, so a Biome update can't fail on a schema version mismatch, and it leaves Markdown and YAML alone, as Biome 2.5 did.
+10. **Renovate auto-merge** is patch + minor only via `:automergeStableNonMajor`. Major bumps stay open for human review. The 7-day cooldown is the first gate, and the required checks on `main` are the second. Renovate merges through GitHub's auto-merge, which waits only for required checks: `Lint, build, verify` (lint, typecheck, build, `bun run verify` and Playwright), `dependency-review`, `secret-scan`, `DCO`, `REUSE compliance`, `Submit bun.lock`, `Bench: format, lint, build, test`, `Bench: deploy`, `Lighthouse audit` and `Vercel`, plus a CodeQL code scanning rule. `bun audit` is not required, so it doesn't hold an auto-merge. Lighthouse takes the best of three runs per URL, so one slow shared runner doesn't fail it. Lock file maintenance automerges too. Renovate cannot apply `minimumReleaseAge` to a Bun lock refresh, so `renovate.json` turns it off for that update type; `bunfig.toml` enforces the same wait while Bun regenerates `bun.lock`. `biome.json` takes its `$schema` from the installed `@biomejs/biome` package rather than a versioned URL, so a Biome update can't fail on a schema version mismatch, and it leaves Markdown and YAML alone, as Biome 2.5 did.
 
 ### bench/ (Bun workspace member)
 
diff --git a/lighthouserc.cjs b/lighthouserc.cjs
index 9b68d77..4372e64 100644
--- a/lighthouserc.cjs
+++ b/lighthouserc.cjs
@@ -85,7 +85,13 @@ module.exports = {
       startServerCommand: "bun run start",
       startServerReadyPattern: "Ready",
       url: baseUrls,
-      numberOfRuns: 1,
+      // "Lighthouse audit" is a required check, so one slow shared runner
+      // must not fail it. With several runs, LHCI asserts on the best one
+      // (its default `aggregationMethod: "optimistic"`): a real regression
+      // shows in every run and still fails, an outlier run does not. One
+      // blog-post run fell to 0.74 performance on 2026-10-30 against a usual
+      // 1.00.
+      numberOfRuns: 3,
       settings: {
         preset: "desktop",
         throttlingMethod: "simulate",