Input initial_level_prior inconsistent with initial_state_prior (sts.LocalLevel)

This might turn out to clarification, but feel free to let me know and I can close it.

In the simplest cases of
LocalLevel
LocalLinearTrend

the highlighted lines above define initial_state_prior through a MultivariateNormalDiag with mean and scale defined by initial_level_prior or initial_slope_prior.

May I know why are the priors defined as MultivariateNormalDiag even though the initial_level_prior/initial_scale_prior inputs may not be Normal?

For example, suppose I have the following simple random walk model, where my initial_level_prior is a LogNormal distribution.

trend = sts.LocalLevel(
    observed_time_series = co2_by_month_training_data,
    initial_level_prior=tfd.LogNormal(
        loc=2.,
        scale=1.
    )
)

If I check the initial_state_prior, as expected, the prior has been redefined as MultiNormal (which in this is simply a normal distribution)

trend.initial_state_prior

<tfp.distributions.MultivariateNormalDiag 'MultivariateNormalDiag' batch_shape=[] event_shape=[1] dtype=float32>

My questions are:

1. What is the rationale for overwriting the initial_state_prior with MultivariateNormalDiag as opposed to a joint independent distribution?
2. Is there a way for me to enforce the initial_state_prior to user defined distribution?

Thank you.

Start with the linked initial_state_prior definitions in tensorflow_probability/python/sts/components/local_level.py and local_linear_trend.py, then reproduce the LogNormal example from the issue. Compare the documented behavior of initial_level_prior and initial_state_prior, and establish whether the desired outcome is clarification or support for preserving a user-defined distribution.