What does the method get_weights return in the case of Bayesian layers?

If you execute the following code

import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp


input = tf.keras.layers.Input(shape=(1,))
layer = tfp.layers.DenseReparameterization(2)

x = layer(input)

model = tf.keras.models.Model(inputs=input, outputs=x)

inp = np.array([[1]])
print(model.predict(inp))

print("weights =", layer.get_weights())

print("layer.kernel_posterior.mean() =", layer.kernel_posterior.mean())
print("layer.kernel_posterior.stddev() =", layer.kernel_posterior.stddev())

print("layer.kernel_prior.mean() =", layer.kernel_prior.mean())
print("layer.kernel_prior.stddev() =", layer.kernel_prior.stddev())

print("layer.bias_posterior.mean() =", layer.bias_posterior.mean())
print("layer.bias_posterior.stddev() =", layer.bias_posterior.stddev())

You will get something like

[[0.07018379 0.13365692]]

weights = [
array([[0.01465803, 0.04191336]], dtype=float32), 
array([[-2.9708524, -2.859167 ]], dtype=float32), 
array([0.11541488, 0.1064656 ], dtype=float32)
]

layer.kernel_posterior.mean() = tf.Tensor([[0.01465803 0.04191336]], shape=(1, 2), dtype=float32)
layer.kernel_posterior.stddev() = tf.Tensor([[0.04998922 0.05573414]], shape=(1, 2), dtype=float32)

layer.kernel_prior.mean() = tf.Tensor([[0. 0.]], shape=(1, 2), dtype=float32)
layer.kernel_prior.stddev() = tf.Tensor([[1. 1.]], shape=(1, 2), dtype=float32)

layer.bias_posterior.mean() = tf.Tensor([0.11541488 0.1064656 ], shape=(2,), dtype=float32)
layer.bias_posterior.stddev() = tf.Tensor([0. 0.], shape=(2,), dtype=float32)

So, layer.get_weights(), where layer = tfp.layers.DenseReparameterization(2), returns a list of 3 arrays. The first array corresponds to the means of kernel_posterior and the last array corresponds to the means of bias_posterior. What does the second array array([[-2.9708524, -2.859167 ]], dtype=float32) correspond to?

Furthermore, why aren't the standard deviations also weights?

Run the minimal reproduction in the issue and inspect DenseReparameterization, get_weights, kernel_posterior, and bias_posterior behavior. Done when the documentation or issue response clearly identifies the second returned array and explains why posterior standard deviations are not returned as separate weights.