Validate FitConfig parameters and fail fast on invalid ranges (#4)

FitConfig and EmbeddingConfig previously accepted out-of-range values
silently (e.g. target_teacher_agreement=1.5, empty frontier_targets,
max_fit_labels=0, batch_size=0). The bad value only surfaced — if at
all — as a confusing failure deep inside the fit.

Add __post_init__ validation that raises ValueError with a clear,
specific message at construction time, plus tests. Non-breaking: all
existing valid configs still construct fine.

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: anwesh2410

src/tracer/config.py

@@ -14,6 +14,10 @@ class EmbeddingConfig:
     batch_size: int = 128
     normalize: bool = True
 
+    def __post_init__(self) -> None:
+        if self.batch_size <= 0:
+            raise ValueError(f"batch_size must be > 0, got {self.batch_size}")
+
 
 @dataclass
 class FitConfig:
@@ -35,3 +39,28 @@ class FitConfig:
     # on ~3k × 22-class embedding inputs). See `tracer.fit.surrogate._candidates`
     # for the full set of names.
     skip_candidates: Tuple[str, ...] = ()
+
+    def __post_init__(self) -> None:
+        if not 0.0 < self.target_teacher_agreement <= 1.0:
+            raise ValueError(
+                "target_teacher_agreement must be in (0, 1], got "
+                f"{self.target_teacher_agreement}")
+        if not self.frontier_targets:
+            raise ValueError("frontier_targets must not be empty")
+        for t in self.frontier_targets:
+            if not 0.0 < t <= 1.0:
+                raise ValueError(
+                    f"frontier_targets values must be in (0, 1], got {t}")
+        if not 0.0 <= self.min_deploy_coverage <= 1.0:
+            raise ValueError(
+                "min_deploy_coverage must be in [0, 1], got "
+                f"{self.min_deploy_coverage}")
+        if self.max_fit_labels <= 0:
+            raise ValueError(f"max_fit_labels must be > 0, got {self.max_fit_labels}")
+        if not 0.0 <= self.explore_rate <= 1.0:
+            raise ValueError(f"explore_rate must be in [0, 1], got {self.explore_rate}")
+        if self.retrain_every <= 0:
+            raise ValueError(f"retrain_every must be > 0, got {self.retrain_every}")
+        if self.min_labels_to_start < 0:
+            raise ValueError(
+                f"min_labels_to_start must be >= 0, got {self.min_labels_to_start}")

tests/test_fit.py

@@ -399,6 +399,55 @@ def test_public_exports():
     assert hasattr(tracer, "RepresentativeExample")
 
 
+# ── Config validation ───────────────────────────────────────────────────────
+
+def test_fitconfig_rejects_out_of_range_ta():
+    """target_teacher_agreement must be in (0, 1]."""
+    from tracer.config import FitConfig
+    with pytest.raises(ValueError):
+        FitConfig(target_teacher_agreement=1.5)
+    with pytest.raises(ValueError):
+        FitConfig(target_teacher_agreement=-0.2)
+    with pytest.raises(ValueError):
+        FitConfig(target_teacher_agreement=0.0)
+
+
+def test_fitconfig_rejects_empty_frontier_targets():
+    """frontier_targets must be non-empty and within (0, 1]."""
+    from tracer.config import FitConfig
+    with pytest.raises(ValueError):
+        FitConfig(frontier_targets=())
+    with pytest.raises(ValueError):
+        FitConfig(frontier_targets=(0.9, 1.2))
+
+
+def test_fitconfig_rejects_bad_scalars():
+    """min_deploy_coverage, max_fit_labels, explore_rate, retrain_every are guarded."""
+    from tracer.config import FitConfig
+    with pytest.raises(ValueError):
+        FitConfig(min_deploy_coverage=1.5)
+    with pytest.raises(ValueError):
+        FitConfig(max_fit_labels=0)
+    with pytest.raises(ValueError):
+        FitConfig(explore_rate=2.0)
+    with pytest.raises(ValueError):
+        FitConfig(retrain_every=0)
+
+
+def test_fitconfig_accepts_valid_config():
+    """Default config and a typical override construct without error."""
+    from tracer.config import FitConfig
+    FitConfig()
+    FitConfig(target_teacher_agreement=0.95)
+
+
+def test_embeddingconfig_rejects_nonpositive_batch_size():
+    from tracer.config import EmbeddingConfig
+    with pytest.raises(ValueError):
+        EmbeddingConfig(batch_size=0)
+    EmbeddingConfig(batch_size=64)
+
+
 if __name__ == "__main__":
     import traceback
     tests = [
@@ -423,6 +472,11 @@ def test_public_exports():
         test_skip_candidates_removes_models,
         test_demo_runs,
         test_public_exports,
+        test_fitconfig_rejects_out_of_range_ta,
+        test_fitconfig_rejects_empty_frontier_targets,
+        test_fitconfig_rejects_bad_scalars,
+        test_fitconfig_accepts_valid_config,
+        test_embeddingconfig_rejects_nonpositive_batch_size,
     ]
     for t in tests:
         try: