fix(benchmarks): complete local setup and CI scope

benchmarks/codex-mcp/scripts/setup.mjs

@@ -108,6 +108,11 @@ async function main() {
     'uv',
     ['sync', '--frozen', '--extra', 'local-worker', '--extra', 'mcp', '--extra', 'benchmarks'],
   );
+  run(
+    'uv',
+    ['run', '--no-sync', 'vidxp', 'init', '--yes'],
+    { env: commandEnvironment },
+  );
   run('npm', ['ci'], { cwd: benchmarkRoot });
 
   const glibcVersion = process.report?.getReport().header.glibcVersionRuntime;

docs/benchmarking/agent_ablation.md

@@ -101,7 +101,10 @@ therefore does not replace LongVALE in this ablation.
 Promptfoo 0.122.2 requires Node.js 22.22.0 or newer. The benchmark-local
 `.npmrc` enforces that requirement so an unsupported runtime fails during
 installation instead of failing after Codex runs have begun. You also need
-`uv` and the Codex CLI on `PATH`.
+`uv` and the Codex CLI on `PATH`. The setup verifies FFmpeg and ffprobe and,
+when they are absent, installs them through a supported package manager. On a
+fresh macOS machine, install Homebrew before running setup so VidXP can install
+FFmpeg automatically.
 
 From the repository root, run the automated setup:
 
@@ -110,12 +113,13 @@ npm --prefix benchmarks/codex-mcp run setup
 ```
 
 The command installs the pinned Python and Node dependencies, creates isolated
-state outside the checkout, opens Codex login when authentication is absent,
-downloads and verifies the pinned LongVALE archive, copies the five pilot
-videos, prepares the four required capabilities, indexes the media, saves the
-evaluation environment in the ignored `benchmarks/codex-mcp/.env` file, and
-runs preflight. Accept the LongVALE dataset terms before running it. Do not copy
-or commit the generated `auth.json`.
+state outside the checkout, initializes the system media runtime, opens Codex
+login when authentication is absent, downloads and verifies the pinned
+LongVALE archive, copies the five pilot videos, prepares the four required
+capabilities, indexes the media, saves the evaluation environment in the
+ignored `benchmarks/codex-mcp/.env` file, and runs preflight. Accept the
+LongVALE dataset terms before running it. Do not copy or commit the generated
+`auth.json`.
 
 By default, mutable state goes under the operating system's user data
 directory. Set only `VIDXP_EVAL_ROOT` when it needs to live elsewhere:

tests/test_ci_scope.py

@@ -20,6 +20,12 @@ def test_tests_and_desktop_changes_skip_container_builds(self):
             Scope(run_suite=True, run_container=False, run_desktop=True),
         )
 
+    def test_benchmark_changes_skip_product_artifact_builds(self):
+        self.assertEqual(
+            classify(["benchmarks/codex-mcp/scripts/setup.mjs"]),
+            Scope(run_suite=True, run_container=False, run_desktop=False),
+        )
+
     def test_product_and_workflow_changes_validate_containers(self):
         for path in (
             "src/vidxp/new_feature.py",

utils/ci_scope.py

@@ -27,7 +27,9 @@ def _is_documentation(path: str) -> bool:
 
 
 def _is_container_neutral(path: str) -> bool:
-    return path.startswith((".agents/", "desktop/", "plugins/", "tests/"))
+    return path.startswith(
+        (".agents/", "benchmarks/", "desktop/", "plugins/", "tests/")
+    )
 
 
 def _affects_desktop(path: str) -> bool:
@@ -47,6 +49,7 @@ def _is_unknown_product_path(path: str) -> bool:
         (
             ".agents/",
             ".github/",
+            "benchmarks/",
             "desktop/",
             "plugins/",
             "src/",