diff --git a/aether/__init__.py b/aether/__init__.py
index 21f309a..190946a 100644
--- a/aether/__init__.py
+++ b/aether/__init__.py
@@ -1,3 +1,12 @@
+"""Aether-ML: a deep learning framework built on NumPy and CuPy, with hand-written
+CUDA and ROCm kernels.
+
+Model, layers, losses, metrics, optimizers and preprocessing transforms are loaded
+lazily on first attribute access, so ``import aether`` stays cheap.
+
+The array backend is selected with ``Model.to()``: ``"numpy"`` runs on the CPU, and
+``"cupy"`` runs on the GPU using the hand-written kernels where the device supports them.
+"""
 from importlib.metadata import PackageNotFoundError
 from importlib.metadata import version as _dist_version
 
diff --git a/aether/__init__.pyi b/aether/__init__.pyi
index d04e571..c90169d 100644
--- a/aether/__init__.pyi
+++ b/aether/__init__.pyi
@@ -5,6 +5,16 @@
 # gets real autocomplete/type info despite the runtime 
 # __getattr__ lazy loading.
 
+"""Aether-ML: a deep learning framework built on NumPy and CuPy, with hand-written
+CUDA and ROCm kernels.
+
+Model, layers, losses, metrics, optimizers and preprocessing transforms are loaded
+lazily on first attribute access, so ``import aether`` stays cheap.
+
+The array backend is selected with ``Model.to()``: ``"numpy"`` runs on the CPU, and
+``"cupy"`` runs on the GPU using the hand-written kernels where the device supports them.
+"""
+
 from aether.layers.activations import LeakyReLU as LeakyReLU
 from aether.layers.activations import ReLU as ReLU
 from aether.layers.activations import SoftMax as SoftMax
diff --git a/aether/_utils/progress.py b/aether/_utils/progress.py
index 0630f17..ad0f931 100644
--- a/aether/_utils/progress.py
+++ b/aether/_utils/progress.py
@@ -10,7 +10,8 @@
 import shutil
 import sys
 import time
-from typing import Callable, Literal
+from collections.abc import Callable
+from typing import Literal
 
 from aether._utils.null_objects import _NullProgress
 
diff --git a/aether/custom_kernels/batchnorm_kernel.py b/aether/custom_kernels/batchnorm_kernel.py
index fa6477e..15fbba8 100644
--- a/aether/custom_kernels/batchnorm_kernel.py
+++ b/aether/custom_kernels/batchnorm_kernel.py
@@ -281,7 +281,7 @@ def get_bn_backward_finalize_kernel(variant: str):
 
 # --- Availability Helper --------------------------------------------------------
 
-def is_gpu_bn_available(variant: str = None) -> bool:
+def is_gpu_bn_available(variant: str | None = None) -> bool:
     """Checks if CuPy hardware support and BatchNorm kernels are available."""
     if variant is None:
         variant, _ = config.resolve_gpu_launch_geometry()
diff --git a/aether/layers/activations.py b/aether/layers/activations.py
index 6757f88..3b920a3 100644
--- a/aether/layers/activations.py
+++ b/aether/layers/activations.py
@@ -118,9 +118,9 @@ class SoftMax(Layer):
 
     Notes
     -----
-    - **Expected input shape:** ``(batch_size, n_classes)``
-    - **Output shape:** ``(batch_size, n_classes)``
-    - **Precision policy:** Exempt from mixed-precision downcasting; forces
+    - Expected input shape: ``(batch_size, n_classes)``
+    - Output shape: ``(batch_size, n_classes)``
+    - Precision policy: Exempt from mixed-precision downcasting; forces
       intermediate arrays and outputs to ``float32``.
     """
     _precision_exempt = True
diff --git a/aether/layers/normalization.py b/aether/layers/normalization.py
index ec2aace..fce991f 100644
--- a/aether/layers/normalization.py
+++ b/aether/layers/normalization.py
@@ -22,13 +22,13 @@ class BatchNorm(Layer):
 
     Notes
     -----
-    - **Expected layout:** Channels-last. Operates over 2D inputs of shape
+    - Expected layout: Channels-last. Operates over 2D inputs of shape
       ``(batch_size, n_features)`` or 4D spatial inputs of shape
       ``(batch_size, height, width, n_channels)``.
-    - **Shape contract:**
+    - Shape contract:
         - Input: ``(d_0, d_1, ..., d_{k-1}, C)``
         - Output: ``(d_0, d_1, ..., d_{k-1}, C)`` (identical to input)
-    - **Training vs. Evaluation:**
+    - Training vs. Evaluation:
         - In training mode (``training=True``), mean and variance are computed
           from the current mini-batch across all leading axes, and running
           statistics are updated.
diff --git a/aether/losses/categorical_crossentropy.py b/aether/losses/categorical_crossentropy.py
index 8506e3b..f19fe03 100644
--- a/aether/losses/categorical_crossentropy.py
+++ b/aether/losses/categorical_crossentropy.py
@@ -49,7 +49,7 @@ def validate_graph(self, layers: list):
         last_layer = layers[-1]
         if isinstance(last_layer, self.prohibited_preceding_layers):
             prohibited_names = ", ".join(cls.__name__ for cls in self.prohibited_preceding_layers)
-            raise ValueError(
+            raise ValueError(  # noqa: TRY004 - graph errors are a ValueError by contract
                 f"[aether] '{type(self).__name__}' operates directly on unnormalized logits. "
                 f"Do not add an explicit '{prohibited_names}' layer immediately preceding this loss."
             )
@@ -126,7 +126,7 @@ class CategoricalCrossEntropy(Loss):
 
     Notes
     -----
-    - **Expected inputs:**
+    - Expected inputs:
         - ``y_pred``: Probability distributions of shape ``(batch_size, n_classes)``.
           Must already be normalized (e.g., via preceding ``SoftMax`` layer).
         - ``y_true``: Target labels as sparse integer indices ``(batch_size,)`` or
@@ -187,11 +187,11 @@ class SoftmaxCategoricalCrossEntropy(Loss):
 
     Notes
     -----
-    - **Expected inputs:**
+    - Expected inputs:
         - ``logits``: Raw, unnormalized activations of shape ``(batch_size, n_classes)``.
         - ``y_true``: Target labels as sparse integer indices ``(batch_size,)`` or
           one-hot distributions ``(batch_size, n_classes)``.
-    - **Graph constraint:** Operates directly on logits; preceding this loss with an
+    - Graph constraint: Operates directly on logits; preceding this loss with an
       explicit ``SoftMax`` layer is prohibited to prevent redundant double normalization.
     """
     prohibited_preceding_layers = (SoftMax,)
diff --git a/aether/model.py b/aether/model.py
index 061ae8c..55c364a 100644
--- a/aether/model.py
+++ b/aether/model.py
@@ -745,10 +745,9 @@ def train(
                 Defaults to True.
             print_every (int): Step interval frequency for logging telemetry.
                 Ignored when `verbose=0`.
-            verbose (int, optional): Verbosity mode for training telemetry.
-                - `0`: Silent mode (no output printed).
-                - `1`: Dynamic graphical progress bar with metrics (default).
-                - `2`: Plain text summary per epoch without carriage returns/ANSI escapes.
+            verbose (int, optional): Telemetry verbosity: 0 is silent, 1 shows a dynamic
+                progress bar with metrics (default), 2 prints a plain text summary per epoch
+                without carriage returns or ANSI escapes.
             validation_data (tuple, optional): Raw (X_val, y_val) tuple evaluated per epoch.
             fit_preprocessor (bool, optional): Whether to fit the preprocessor
                 (None=only if unfitted, True=always, False=never).
@@ -911,10 +910,8 @@ def evaluate(self, X, y, *, batch_size=None, verbose=1):
                 Raw/un-preprocessed when a preprocessing pipeline is attached.
             y (ndarray): Ground truth labels or targets.
             batch_size (int, optional): Mini-batch size. Defaults to None (full-batch).
-            verbose (int | bool, optional): Verbosity mode for evaluation.
-                - `0` / `False`: Silent mode (no output printed).
-                - `1` / `True`: Prints the final evaluation loss and accuracy metrics.
-                Defaults to 1.
+            verbose (int | bool, optional): Verbosity for evaluation: 0 / False is silent,
+                1 / True prints the final evaluation loss and accuracy metrics. Defaults to 1.
         Returns:
             tuple: (val_loss, val_acc) representing the computed validation metrics.
         Raises:
diff --git a/aether/preprocessing/transforms.py b/aether/preprocessing/transforms.py
index 997ca58..9b4e4be 100644
--- a/aether/preprocessing/transforms.py
+++ b/aether/preprocessing/transforms.py
@@ -1,5 +1,6 @@
 import warnings
-from typing import Any, Optional, Sequence, Tuple, Union
+from collections.abc import Sequence
+from typing import Any
 
 import numpy as np
 
@@ -20,10 +21,10 @@ def _dtype_name(dt):
 
 def to_tensor(
     *args: Any,
-    target_device: Optional[str] = None,
-    dtype: Union[str, type, np.dtype, Sequence[Union[str, type, np.dtype]], None] = None,
+    target_device: str | None = None,
+    dtype: str | type | np.dtype | Sequence[str | type | np.dtype] | None = None,
     preserve_integers: bool = True
-) -> Union[Any, Tuple[Any, ...], None]:
+) -> Any | tuple[Any, ...] | None:
     """Converts array-like inputs into backend tensors (NumPy or CuPy).
 
     Flexible API for precision handling and hardware device placement. Accepts
@@ -566,7 +567,7 @@ def from_config(cls, cfg):
         return cls(transforms=[deserialize(entry) for entry in cfg.get("transforms") or []])
 
 
-def deserialize(entry: dict | None) -> Optional[Preprocess]:
+def deserialize(entry: dict | None) -> Preprocess | None:
     """Reconstructs a Preprocess transform or pipeline from a serialized configuration dictionary.
 
     This function acts as the universal factory and resolver for preprocessing 
@@ -576,9 +577,9 @@ def deserialize(entry: dict | None) -> Optional[Preprocess]:
 
     When is this used?
     ------------------
-    - **Internally:** Called automatically by `Model.load()` to restore attached 
+    - Internally: Called automatically by `Model.load()` to restore attached 
       data preprocessing pipelines from saved `.aether` model archives.
-    - **Standalone:** Called by users to load or transfer preprocessor pipelines 
+    - Standalone: Called by users to load or transfer preprocessor pipelines 
       independently of a model (e.g., loading fitted scalers into external APIs, 
       ETL data workers, or configuration-driven workflows).
 
diff --git a/examples/cifar10/v2_cifar10.py b/examples/cifar10/v2_cifar10.py
index 42ce51d..31722a3 100644
--- a/examples/cifar10/v2_cifar10.py
+++ b/examples/cifar10/v2_cifar10.py
@@ -4,7 +4,6 @@
 from pathlib import Path
 
 import cupy as cp
-import numpy as np
 
 import aether as ae
 
diff --git a/pyproject.toml b/pyproject.toml
index 8b48252..744426b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -60,6 +60,12 @@ extend-exclude = ["notebooks", "examples/legacy", "*.ipynb"]
 # Allow `import aether.config as config`; preferred here for readability.
 ignore = ["PLR0402"]
 
+[tool.ruff.lint.per-file-ignores]
+# Skeleton templates keep placeholder imports and `pass` bodies on purpose.
+"tests/integration/_integration_test_skeleton.py" = ["F401", "PIE790"]
+# Stub `__all__` lists are grouped by topic, not sorted.
+"*.pyi" = ["RUF022"]
+
 [tool.ruff.lint.flake8-comprehensions]
 # Allow dict(key=value) for kernel template/kwargs bundles.
 allow-dict-calls-with-keyword-arguments = true
diff --git a/tests/base_case.py b/tests/base_case.py
index 844743e..da8651b 100644
--- a/tests/base_case.py
+++ b/tests/base_case.py
@@ -1,8 +1,6 @@
 import unittest
 import warnings
 
-import numpy as np
-
 import aether.config as config
 from aether.config import set_backend
 
@@ -55,7 +53,7 @@ def setUp(self):
 
     def shortDescription(self):
         """Docstrings are ommitted when running verbose -v unit tests"""
-        return None
+        return
 
     def make_built_layer(self, layer_cls, input_shape: tuple[int, ...], seed: int | None = None, **kwargs):
         """Constructs, binds to the active device backend, and builds a layer instance.
diff --git a/tests/config/test_dtype_policy.py b/tests/config/test_dtype_policy.py
index b8b75d2..bd7ec16 100644
--- a/tests/config/test_dtype_policy.py
+++ b/tests/config/test_dtype_policy.py
@@ -2,9 +2,8 @@
 
 import numpy as np
 
-import aether.config as config
 import tests.base_case as base_case
-from aether.config import COMPUTE_DTYPE, PARAM_DTYPE, DTypePolicy
+from aether.config import PARAM_DTYPE, DTypePolicy
 
 
 class TestDTypePolicy(base_case.AetherBaseTestCase):
@@ -34,17 +33,15 @@ def test_invalid_string_raises_value_error(self):
         """Ensure passing strings outside COMPUTE_DTYPE raises ValueError."""
         invalid_dtypes = ['int8', 'int32', 'float8', 'complex64', 'invalid_type', 'bfloat16']
         for dt in invalid_dtypes:
-            with self.subTest(dtype=dt):
-                with self.assertRaises(ValueError):
-                    DTypePolicy(compute_dtype=dt)
+            with self.subTest(dtype=dt), self.assertRaises(ValueError):
+                DTypePolicy(compute_dtype=dt)
 
     def test_non_string_non_none_raises_type_error(self):
         """Ensure non-str and non-None inputs raise TypeError."""
         invalid_types = [123, 32.0, ['float32'], {'dtype': 'float32'}, True]
         for val in invalid_types:
-            with self.subTest(val=val):
-                with self.assertRaises(TypeError):
-                    DTypePolicy(compute_dtype=val)
+            with self.subTest(val=val), self.assertRaises(TypeError):
+                DTypePolicy(compute_dtype=val)
 
     def test_valid_initializations(self):
         """Verify valid compute_dtypes set attributes correctly."""
diff --git a/tests/integration/model_base_suite.py b/tests/integration/model_base_suite.py
index eb37907..7b2acd2 100644
--- a/tests/integration/model_base_suite.py
+++ b/tests/integration/model_base_suite.py
@@ -4,7 +4,6 @@
 import math
 
 import aether as ae
-import aether.config as config
 import tests.base_case as base_case
 
 
@@ -37,9 +36,9 @@ def build_cnn_model(self, *, device=None, precision=None, input_dim=(32, 32, 3))
             optimizer=ae.Adam(lr=0.001, decay=5e-5),
             accuracy=ae.CategoricalAccuracy(),
         )
-        model.to(device if device is not None else self.backend_name)
+        model.to(target_device)
         model.manual_seed(seed=42)
-        
+
         return self._place_and_finalize(model, precision=precision, input_dim=input_dim)
 
     def build_mlp_model(self, *, device=None, precision=None, input_dim=(32, 32, 3)):
@@ -54,7 +53,7 @@ def build_mlp_model(self, *, device=None, precision=None, input_dim=(32, 32, 3))
             optimizer=ae.Adam(lr=0.001, decay=5e-5),
             accuracy=ae.CategoricalAccuracy(),
         )
-        model.to(device if device is not None else self.backend_name)
+        model.to(target_device)
         model.manual_seed(seed=42)
         return self._place_and_finalize(model, precision=precision, input_dim=input_dim)
 
diff --git a/tests/layers/test_batch_normalization.py b/tests/layers/test_batch_normalization.py
index b0a226c..6886de4 100644
--- a/tests/layers/test_batch_normalization.py
+++ b/tests/layers/test_batch_normalization.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import aether.custom_kernels.relu_kernel as gpu_relu
 import tests.base_case as base_case
 from aether.layers.activations import LeakyReLU, ReLU
diff --git a/tests/layers/test_flatten.py b/tests/layers/test_flatten.py
index d0a26f7..20c8c01 100644
--- a/tests/layers/test_flatten.py
+++ b/tests/layers/test_flatten.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.linear import Flatten
 
diff --git a/tests/layers/test_leaky_relu.py b/tests/layers/test_leaky_relu.py
index 9bb5f88..900d9a6 100644
--- a/tests/layers/test_leaky_relu.py
+++ b/tests/layers/test_leaky_relu.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.activations import LeakyReLU
 
diff --git a/tests/layers/test_max_pool2d.py b/tests/layers/test_max_pool2d.py
index 13aecb0..105d7b0 100644
--- a/tests/layers/test_max_pool2d.py
+++ b/tests/layers/test_max_pool2d.py
@@ -244,9 +244,9 @@ def test_same_padding_boundary_scaling(self):
 
     def test_build_invalid_padding_raises_value_error(self):
         with self.assertRaises(ValueError):
-            layer = self.make_built_layer(
-                MaxPool2d, 
-                input_shape=self.INPUT_SHAPE[1:], 
+            self.make_built_layer(
+                MaxPool2d,
+                input_shape=self.INPUT_SHAPE[1:],
                 filter_size=(2, 2), stride=(2, 2),
                 padding='not_a_real_padding')
 
diff --git a/tests/layers/test_softmax.py b/tests/layers/test_softmax.py
index 895356b..b53731c 100644
--- a/tests/layers/test_softmax.py
+++ b/tests/layers/test_softmax.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.activations import SoftMax
 
diff --git a/tests/layers/test_spatial_dropout.py b/tests/layers/test_spatial_dropout.py
index 0ef8c14..24bd610 100644
--- a/tests/layers/test_spatial_dropout.py
+++ b/tests/layers/test_spatial_dropout.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.dropout import SpatialDropout
 
diff --git a/tests/losses/test_categorical_crossentropy.py b/tests/losses/test_categorical_crossentropy.py
index 66dc794..efe64e7 100644
--- a/tests/losses/test_categorical_crossentropy.py
+++ b/tests/losses/test_categorical_crossentropy.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.linear import Dense
 from aether.losses.categorical_crossentropy import CategoricalCrossEntropy
diff --git a/tests/losses/test_softmax_categorical_crossentropy.py b/tests/losses/test_softmax_categorical_crossentropy.py
index 47238c4..2a64167 100644
--- a/tests/losses/test_softmax_categorical_crossentropy.py
+++ b/tests/losses/test_softmax_categorical_crossentropy.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.activations import SoftMax
 from aether.layers.linear import Dense
diff --git a/tests/model/test_evaluate.py b/tests/model/test_evaluate.py
index cca4f73..087fcf9 100644
--- a/tests/model/test_evaluate.py
+++ b/tests/model/test_evaluate.py
@@ -1,6 +1,5 @@
 import contextlib
 import io
-import unittest
 
 import numpy as np
 
diff --git a/tests/model/test_preprocessor_serialization.py b/tests/model/test_preprocessor_serialization.py
index ee7d6f8..6add677 100644
--- a/tests/model/test_preprocessor_serialization.py
+++ b/tests/model/test_preprocessor_serialization.py
@@ -7,7 +7,6 @@
 
 import numpy as np
 
-import aether.config as config
 import tests.base_case as base_case
 from aether._utils import NullPreprocessor
 from aether.layers import Dense, ReLU
diff --git a/tests/optimizers/adam_base_suite.py b/tests/optimizers/adam_base_suite.py
index 0e6e98f..f376fe4 100644
--- a/tests/optimizers/adam_base_suite.py
+++ b/tests/optimizers/adam_base_suite.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.layers.conv import Conv2d
 from aether.layers.linear import Dense
@@ -275,10 +274,7 @@ def test_gpu_kernel_execution(self):
         optimizer.init_params([layer])
         optimizer._compile_for_device("cupy")
 
-        try:
-            optimizer.step()
-        except Exception as e:
-            self.fail(f"CuPy kernel failed during optimizer.step(): {e}")
+        optimizer.step()
 
         self.assertTrue(self.xp.all(self.xp.isfinite(layer.weights)))
         self.assertTrue(self.xp.all(self.xp.isfinite(layer.biases)))
diff --git a/tests/optimizers/test_optimizer_adam.py b/tests/optimizers/test_optimizer_adam.py
index 2737eb5..9208d15 100644
--- a/tests/optimizers/test_optimizer_adam.py
+++ b/tests/optimizers/test_optimizer_adam.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.optimizers.adam import Adam
 from tests.optimizers.adam_base_suite import BaseTestOptimizerAdam
diff --git a/tests/optimizers/test_optimizer_adam_w.py b/tests/optimizers/test_optimizer_adam_w.py
index 68d0f8d..0216b0c 100644
--- a/tests/optimizers/test_optimizer_adam_w.py
+++ b/tests/optimizers/test_optimizer_adam_w.py
@@ -1,4 +1,3 @@
-import aether.config as config
 import tests.base_case as base_case
 from aether.optimizers.adam import AdamW
 from tests.optimizers.adam_base_suite import BaseTestOptimizerAdam
diff --git a/tests/preprocessing/test_deserialize.py b/tests/preprocessing/test_deserialize.py
index 6f04ce2..fc17eae 100644
--- a/tests/preprocessing/test_deserialize.py
+++ b/tests/preprocessing/test_deserialize.py
@@ -1,9 +1,7 @@
-import numpy as np
 
 import tests.base_case as base_case
 from aether.preprocessing.transforms import (
     Compose,
-    Preprocess,
     Rescale,
     StandardScaler,
     ToTensor,
diff --git a/tests/preprocessing/test_rescale.py b/tests/preprocessing/test_rescale.py
index 4984d9c..48192b6 100644
--- a/tests/preprocessing/test_rescale.py
+++ b/tests/preprocessing/test_rescale.py
@@ -1,6 +1,5 @@
 import numpy as np
 
-import aether.config as config
 import tests.base_case as base_case
 from aether.preprocessing.transforms import Rescale