All categorical values are displayed in the legend when source dataframe is polars.DataFrame

This is the example that produces all categorical values in the legend, which is unexpected. I copied it from a notebook, so might need tweaking to display elsewhere.

import polars as pl
import polars.selectors as cs
import datetime as dt
import random

import plotnine
from plotnine import *

import mizani
from mizani.formatters import log_format, custom_format, date_format, label_number, label_percent, label_bytes
from mizani.breaks import date_breaks

# turn on String Cache, otherwise we can't concat different dataframes into one
pl.enable_string_cache()


date = dt.datetime(2025, 8, 29)          # any date is fine
start = dt.datetime(date.year, date.month, date.day, 6, 0, 0)
end   = dt.datetime(date.year, date.month, date.day, 18, 0, 0)

# timestamps: every 5 minutes, inclusive of both ends (06:00 and 18:00)
timestamps = pl.datetime_range(
    start=start,
    end=end,
    interval="5m",
    closed="both",
    eager=True
)

random.seed(42)

sources = pl.Series("source", ["u", "w", "v"], dtype=pl.Categorical)
dirs = pl.Series("dir", ["incoming", "outgoing"], dtype=pl.Categorical)

df = (
    pl.DataFrame({"source": sources})
    .join(
        pl.DataFrame({"five_min_slot": timestamps}),
        how="cross"
    )
    .join(
        pl.DataFrame({"dir": dirs}),
        how="cross"
    )
    .with_columns(
        values=pl.lit(100.0),
    )
    .with_columns(
        pl.col('values').map_elements(lambda v: v*random.random(), return_dtype=pl.self_dtype()),
    )
)

(
    ggplot(
        df.filter(
            pl.col('source').eq('u')
        )
        .sort('five_min_slot'),
        aes(x='five_min_slot', y='values', color='factor(dir)')
    )
    + geom_line()
    + theme_linedraw()
    + theme(
        figure_size=(16, 5),
        axis_text_x=element_text(rotation=45, ha='right'),
    )
    + scale_x_datetime(
        labels=date_format("%H:%M"),
        date_breaks="15 minute",
        date_minor_breaks="5 minute",
        expand=(0, 0),
    )
    + scale_y_continuous(
        labels=label_percent(scale=1),
    )

)

For that code, I get the following output:

I would expect the legend to contain only incoming and outgoing.

Start by running the provided Python example with the Polars dataframe and compare the generated legend with the expected incoming and outgoing entries. Trace how categorical values from the filtered dataframe are collected for the legend, then verify that unused values are excluded and the example produces only those two entries.