fix(runs): address CodeRabbit review findings

- cli.py: `--only` filter now preserves `shared_args`; previously the
  filtered Preset dropped them, silently losing every job's shared flags.
- loader.py: `--set` now rejects unknown top-level keys. A typo like
  `--set job.rf3.gpus=0` (note missing 's') used to auto-create an unused
  `job` dict and silently no-op; now it raises KeyError with the valid
  keys listed.
- runner.py: handle partial spawn failures. If one job fails to spawn
  midway, already-launched jobs are terminated and joined instead of
  orphaned. Also wraps Popen in try/close to avoid log-file handle leak
  if the subprocess fails to start.
- presets/rf3_partial_chiral_off.toml: change `output_subdir = "."`
  (write to RESULTS_DIR root) to `"rf3"` (subdir under RESULTS_DIR) for
  consistency with the other RF3 presets and collision safety when
  RESULTS_DIR is overridden to a shared location.

Skipped (faithful to bash original): rf3_protenix.toml asymmetric
partial-diffusion-step — the source `run_rf3_protenix_mdc_actl.sh`
deliberately set it only for the Protenix job, not RF3.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: mag-astera

src/sampleworks/runs/cli.py

@@ -87,6 +87,7 @@ def _filter_only(preset: Preset, only: str) -> Preset:
         name=preset.name,
         description=preset.description,
         defaults=preset.defaults,
+        shared_args=preset.shared_args,
         jobs=keep,
     )
 

src/sampleworks/runs/loader.py

@@ -64,9 +64,17 @@ def _apply_overrides(raw: dict[str, Any], overrides: list[str]) -> dict[str, Any
     return raw
 
 
+_TOP_LEVEL_KEYS = frozenset({"description", "defaults", "shared_args", "jobs"})
+
+
 def _set_dotted(obj: dict[str, Any], dotted: str, value: Any) -> None:
     """Set ``obj`` at ``a.b.c`` to ``value``. Job name lookup is allowed under ``jobs``."""
     parts = dotted.split(".")
+    if parts[0] not in _TOP_LEVEL_KEYS:
+        raise KeyError(
+            f"--set: unknown top-level key {parts[0]!r} in {dotted!r}. "
+            f"Valid top-level keys: {sorted(_TOP_LEVEL_KEYS)}"
+        )
     cursor: Any = obj
     for i, part in enumerate(parts[:-1]):
         cursor = _index(cursor, part, where=".".join(parts[: i + 1]))

src/sampleworks/runs/presets/rf3_partial_chiral_off.toml

@@ -22,5 +22,5 @@ disable-chiral-features = true
 name = "rf3"
 env = "rf3"
 gpus = "5"
-output_subdir = "."
+output_subdir = "rf3"
 args = { model = "rf3", gradient-weights = "0.0 0.005 0.01 0.02 0.035 0.05 0.1 0.2 0.35 0.5", model-checkpoint = "${RF3_CHECKPOINT}" }

src/sampleworks/runs/runner.py

@@ -64,10 +64,26 @@ def run(preset: Preset, *, results_dir: Path, dry_run: bool = False) -> int:
         return 0
 
     _print_launch_summary(preset, invocations)
-    processes = [_spawn(inv) for inv in invocations]
+    processes: list[_RunningJob] = []
+    try:
+        for inv in invocations:
+            processes.append(_spawn(inv))
+    except BaseException:
+        _terminate_all(processes)
+        raise
     return _wait_all(processes)
 
 
+def _terminate_all(jobs: list[_RunningJob]) -> None:
+    """Terminate any already-launched jobs (used when a later spawn fails)."""
+    for j in jobs:
+        if j.proc.poll() is None:
+            j.proc.terminate()
+    for j in jobs:
+        j.proc.wait()
+        j.tee_thread.join()
+
+
 def _print_dry_run(inv: JobInvocation) -> None:
     print(f"# job: {inv.job.name}  (env={inv.job.env}, gpus={inv.job.gpus})", file=sys.stderr)
     print(f"# log: {inv.log_path}", file=sys.stderr)
@@ -99,13 +115,17 @@ class _RunningJob:
 def _spawn(inv: JobInvocation) -> _RunningJob:
     inv.log_path.parent.mkdir(parents=True, exist_ok=True)
     log_file = open(inv.log_path, "wb")
-    proc = subprocess.Popen(
-        inv.argv,
-        env=inv.env,
-        stdout=subprocess.PIPE,
-        stderr=subprocess.STDOUT,
-        bufsize=0,
-    )
+    try:
+        proc = subprocess.Popen(
+            inv.argv,
+            env=inv.env,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.STDOUT,
+            bufsize=0,
+        )
+    except BaseException:
+        log_file.close()
+        raise
     assert proc.stdout is not None
     thread = threading.Thread(
         target=_tee,

tests/runs/test_loader.py

@@ -130,6 +130,13 @@ def test_set_without_equals_raises(monkeypatch: pytest.MonkeyPatch) -> None:
         loader.load_preset("rf3_partial", overrides=["bogus_no_equals"])
 
 
+def test_set_with_unknown_top_level_key_raises(monkeypatch: pytest.MonkeyPatch) -> None:
+    """Typos like ``--set job.rf3.gpus=0`` (missing 's' in jobs) must not silently no-op."""
+    monkeypatch.setenv("HOME", "/home/test")
+    with pytest.raises(KeyError, match="unknown top-level key"):
+        loader.load_preset("rf3_partial", overrides=["job.rf3.gpus=0"])
+
+
 def test_bad_env_rejected(tmp_path: Path) -> None:
     bad = tmp_path / "bad.toml"
     bad.write_text(