fix: harden validate CLI + scrub registry path leaks

.github/workflows/ci.yml

@@ -43,7 +43,4 @@ jobs:
       - name: Install
         run: pip install -e packages/python
       - name: Validate all registry configs
-        run: |
-          for f in registry/**/*.yaml; do
-            python -m servingcard validate "$f"
-          done
+        run: python -m servingcard validate registry/

packages/python/servingcard/cli.py

@@ -218,17 +218,48 @@ def _fetch_remote_card(url: str) -> ServingCard:
 
 @app.command()
 def validate(
-    path: Path = typer.Argument(..., help="Path to a servingcard YAML file"),
+    paths: list[Path] = typer.Argument(
+        ...,
+        help="One or more servingcard YAML files (or directories to recurse)",
+        exists=True,
+        file_okay=True,
+        dir_okay=True,
+        readable=True,
+    ),
 ) -> None:
-    """Validate a servingcard YAML file."""
-    errors = validate_card(path)
-    if errors:
-        typer.echo(f"INVALID: {path}")
-        for err in errors:
-            typer.echo(f"  - {err}")
+    """Validate one or more servingcard YAML files.
+
+    Accepts files or directories. Directories are recursed for *.yaml and
+    *.yml. Exits 1 on any invalid file. Always loud about what was found —
+    silent success on a typo'd path is the failure mode this command is
+    built to prevent (CI used to silently iterate a literal glob).
+    """
+    targets: list[Path] = []
+    for p in paths:
+        if p.is_dir():
+            targets.extend(sorted(p.rglob("*.yaml")))
+            targets.extend(sorted(p.rglob("*.yml")))
+        else:
+            targets.append(p)
+
+    if not targets:
+        typer.echo("ERROR: no .yaml or .yml files found in given paths", err=True)
+        raise typer.Exit(code=2)
+
+    invalid = 0
+    for target in targets:
+        errors = validate_card(target)
+        if errors:
+            invalid += 1
+            typer.echo(f"INVALID: {target}")
+            for err in errors:
+                typer.echo(f"  - {err}")
+        else:
+            typer.echo(f"VALID: {target}")
+
+    typer.echo(f"\n{len(targets) - invalid}/{len(targets)} valid")
+    if invalid:
         raise typer.Exit(code=1)
-    else:
-        typer.echo(f"VALID: {path}")
 
 
 # ---------------------------------------------------------------------------

packages/python/tests/test_validate.py

@@ -5,9 +5,13 @@
 from pathlib import Path
 
 import yaml
+from typer.testing import CliRunner
 
+from servingcard.cli import app as _app
 from servingcard.validate import validate_card
 
+_runner = CliRunner()
+
 # ---------------------------------------------------------------------------
 # 1. Valid config returns empty errors list
 # ---------------------------------------------------------------------------
@@ -167,3 +171,37 @@ def test_speculative_without_benchmark(tmp_path: Path, minimal_card_dict: dict)
     p.write_text(yaml.dump(d))
     errors = validate_card(p)
     assert any("speculative" in e.lower() for e in errors)
+
+
+# ---------------------------------------------------------------------------
+# Spec 009 follow-up: validate CLI must fail loudly on missing paths and
+# accept directories. Regression for the "silent literal-glob" CI failure.
+# ---------------------------------------------------------------------------
+
+
+def test_validate_cli_fails_on_missing_path() -> None:
+    result = _runner.invoke(_app, ["validate", "/no/such/path/at/all.yaml"])
+    assert result.exit_code != 0
+
+
+def test_validate_cli_recurses_directory(tmp_path: Path, tmp_valid_yaml: Path) -> None:
+    nested = tmp_path / "nested" / "dir"
+    nested.mkdir(parents=True)
+    (nested / "card.yaml").write_text(tmp_valid_yaml.read_text())
+    result = _runner.invoke(_app, ["validate", str(tmp_path)])
+    assert result.exit_code == 0, result.stdout
+    assert "VALID" in result.stdout
+    assert "2/2 valid" in result.stdout
+
+
+def test_validate_cli_directory_with_no_yaml_errors(tmp_path: Path) -> None:
+    result = _runner.invoke(_app, ["validate", str(tmp_path)])
+    assert result.exit_code == 2  # distinct from validation failure (1)
+
+
+def test_validate_cli_accepts_multiple_paths(tmp_path: Path, tmp_valid_yaml: Path) -> None:
+    p2 = tmp_path / "second.yaml"
+    p2.write_text(tmp_valid_yaml.read_text())
+    result = _runner.invoke(_app, ["validate", str(tmp_valid_yaml), str(p2)])
+    assert result.exit_code == 0, result.stdout
+    assert "2/2 valid" in result.stdout

registry/deepseek-coder-v2-lite/gb10-fp8-baseline.yaml

@@ -15,7 +15,7 @@ huggingface:
 method_detail: PawBench full suite (PawStyle scenarios, 2 runs, concurrency 1/2/4/8)
 serving:
   engine_args:
-    model: /home/vvladescu/models/deepseek-coder-v2-lite-fp8
+    model: ${MODEL_DIR}/deepseek-coder-v2-lite-fp8
     quantization: fp8
     gpu_memory_utilization: 0.8
     max_model_len: 65536

registry/devstral-small-24b/gb10-baseline.yaml

@@ -15,7 +15,7 @@ huggingface:
 method_detail: PawBench full suite (PawStyle scenarios, 2 runs, concurrency 1/2/4/8)
 serving:
   engine_args:
-    model: /home/vvladescu/models/devstral-small-24b
+    model: ${MODEL_DIR}/devstral-small-24b
     quantization: bf16
     gpu_memory_utilization: 0.8
     max_model_len: 65536

registry/qwen3-coder/gb10-fp8-baseline.yaml

@@ -16,7 +16,7 @@ huggingface:
   url: https://huggingface.co/Qwen/Qwen3-Coder-Next
 serving:
   engine_args:
-    model: /home/vvladescu/models/qwen3-coder-next-fp8
+    model: ${MODEL_DIR}/qwen3-coder-next-fp8
     quantization: fp8
     gpu_memory_utilization: 0.8
     max_model_len: 131072

registry/qwen3-coder/gb10-fp8-eagle3-spec3.yaml

@@ -16,7 +16,7 @@ huggingface:
   url: https://huggingface.co/Qwen/Qwen3-Coder-Next
 serving:
   engine_args:
-    model: /home/vvladescu/models/qwen3-coder-next-fp8
+    model: ${MODEL_DIR}/qwen3-coder-next-fp8
     quantization: fp8
     gpu_memory_utilization: 0.8
     max_model_len: 131072