fix(ci): add imageio runtime dep, narrow types for ty

- pyproject: `imageio[ffmpeg]>=2.20` is now a runtime dep so the
  shared `EpisodeVideoRecorder` works on a base install (and CI's
  bare `uv sync --all-extras --dev` env can run the recorder tests).
- recording.py: rewrite `_resolve_filename` with `isinstance(spec, str)`
  branches so ty can narrow `str | Callable[..., str]` at the call site.
- test_recording.py: add `assert final is not None` before `.exists()`
  in three tests so ty narrows `Path | None`. Replace `r.count_frames()`
  with iteration-based counting (imageio's `_BaseReaderWriter` stub
  omits `count_frames`) and `# ty: ignore[not-iterable]` for the
  same stub gap on `__iter__`.
- benchmark.py: drop unused `# type: ignore` on `mani_skill.envs.sapien_env`
  import (ty flagged it as unused after the runtime-dep refresh).

Author: claude

pyproject.toml

@@ -12,6 +12,7 @@ license = {text = "Apache-2.0"}
 dependencies = [
     "anyio>=4.0.0",
     "filelock>=3.0",
+    "imageio[ffmpeg]>=2.20", # Streaming mp4 writer for benchmarks/recording.py
     "lazyregistry>=0.4.0",  # Python 3.8 support
     "msgpack>=1.0.0",
     "numpy>=1.24",

src/vla_eval/benchmarks/recording.py

@@ -278,9 +278,10 @@ def discard(self) -> None:
         _safe_unlink(tempfile_path)
 
     def _resolve_filename(self, context: Mapping[str, Any]) -> str:
-        if callable(self._filename_spec):
-            return self._filename_spec(context)
-        return self._filename_spec.format(**context)
+        spec = self._filename_spec
+        if isinstance(spec, str):
+            return spec.format(**context)
+        return spec(context)
 
 
 def _safe_unlink(path: Path | None) -> None:

src/vla_eval/benchmarks/robomme/benchmark.py

@@ -342,7 +342,7 @@ def _patched_parse(sim_backend, render_backend):
 
             _backend_mod.parse_sim_and_render_backend = _patched_parse
 
-            import mani_skill.envs.sapien_env  # type: ignore
+            import mani_skill.envs.sapien_env
 
             mani_skill.envs.sapien_env.parse_sim_and_render_backend = _patched_parse
         except Exception as e:

tests/test_recording.py

@@ -29,10 +29,12 @@ def _frame() -> np.ndarray:
 
 
 def _count_frames(path: Path) -> int:
+    # imageio's `_BaseReaderWriter` type stub omits __iter__; the concrete
+    # Reader returned for mp4 is iterable in practice.
     import imageio
 
     with imageio.get_reader(str(path)) as r:
-        return r.count_frames()
+        return sum(1 for _ in r)  # ty: ignore[not-iterable]
 
 
 # ---------------------------------------------------------------------------
@@ -58,6 +60,7 @@ def test_save_uses_status_in_filename(tmp_path: Path) -> None:
     rec.start({"task_name": "T", "episode_idx": 7})
     rec.record(_frame())
     final = rec.save(status="fail")
+    assert final is not None
     assert final == tmp_path / "T_ep7_fail.mp4"
     assert final.exists()
 
@@ -78,6 +81,7 @@ def test_consecutive_episodes_each_produce_their_own_file(tmp_path: Path) -> Non
         rec.record(_frame())
         rec.record(_frame())
         final = rec.save(status="success")
+        assert final is not None
         assert final == tmp_path / f"T_ep{ep}_success.mp4"
         assert final.exists()
     assert sorted(p.name for p in tmp_path.glob("T_ep*.mp4")) == [
@@ -101,6 +105,7 @@ def test_filename_template_with_subdirectories(tmp_path: Path) -> None:
     rec.start({"suite": "Counting", "task_name": "PickX", "episode_idx": 12})
     rec.record(_frame())
     final = rec.save(status="success")
+    assert final is not None
     assert final == tmp_path / "Counting" / "PickX_ep0012_success.mp4"
     assert final.exists()