fix: preserve backend version input errors in predict

python/aisim/src/aisim/capacity.py

@@ -147,8 +147,9 @@ def finish_lowering(value):
                 finally:
                     model.close()
             if diagnostics.get("readiness") != "ready":
-                from .estimator_readiness import unready_estimator_message
+                from .estimator_readiness import reject_unlisted_estimator_version, unready_estimator_message
 
+                reject_unlisted_estimator_version(diagnostics)
                 raise ValueError(unready_estimator_message(diagnostics))
             canonical_result = dict(raw)
             resolved = diagnostics["provenance"]["config"]

python/aisim/src/aisim/estimator_readiness.py

@@ -92,6 +92,32 @@ def perf_data_missing_message(error: BaseException | str) -> str | None:
     )
 
 
+def reject_unlisted_estimator_version(diagnostics: dict[str, Any]) -> None:
+    """Preserve a version-slot input refusal before formatting an unready estimator."""
+
+    provenance = diagnostics.get("provenance") or {}
+    if not provenance.get("selection_failures") and not diagnostics.get("last_warning"):
+        # Explicit regression identities need not use an op-data version slot.
+        return
+    config = provenance.get("config") or {}
+    if not all(config.get(key) for key in ("system", "backend", "backend_version")):
+        return
+    from aisim_core.sdk.perf_database import UnlistedQueryVersionError, resolve_query_version
+
+    try:
+        resolve_query_version(
+            config["system"],
+            config["backend"],
+            config["backend_version"],
+            systems_paths=config.get("systems_paths") or None,
+        )
+    except UnlistedQueryVersionError:
+        raise
+    except ValueError:
+        # Other lookup failures retain the existing missing-data diagnostic.
+        pass
+
+
 def _unavailable_version_message(
     system: str | None,
     backend: str | None,

python/aisim/src/aisim/predict.py

@@ -17,6 +17,7 @@
 from typing import Any
 
 from aisim_core.sdk.errors import UnsupportedVisionFamilyError
+from aisim_core.sdk.perf_database import UnlistedQueryVersionError
 from aisim_core.sdk.utils import get_model_config_from_model_path
 
 from .compiler import prediction_to_replay_spec
@@ -110,6 +111,8 @@ def run_prediction(
         spec, and raw runner report.
 
     Raises:
+        UnlistedQueryVersionError: If timing preparation rejects a backend version
+            outside its queryable slots. Propagates unchanged as an input error.
         UnsupportedVisionFamilyError: If native VL replay cannot support the
             requested image processor. This ValueError subtype propagates unchanged.
         PredictionExecutionError: If the runner fails with another execution error.
@@ -147,11 +150,13 @@ def run_prediction(
     try:
         try:
             report = runner.run(spec, output_requirements=output_requirements)
-        except (KeyboardInterrupt, ResourceLimitError, UnsupportedVisionFamilyError):
+        except (KeyboardInterrupt, ResourceLimitError, UnsupportedVisionFamilyError, UnlistedQueryVersionError):
             # ResourceLimitError from a guarded runner must reach the caller
             # unwrapped so it can be handled distinctly from execution failures.
             # A known unsupported vision family is likewise an input rejection,
             # even though its geometry is resolved while preparing the runner.
+            # Version-slot refusals discovered during timing preparation also
+            # retain their configuration-error type and complete repair guidance.
             raise
         except Exception as exc:
             raise PredictionExecutionError(