DOC: metadata must have the same number of rows as events

I am trying to create epochs with metadata from raw, but run into an error with a piece of code that worked fine about 2 months ago (some previous version, likely 0.20.dev).
I updated to the latest master.

I want to use a subset of events of interest, while my raw data contains more events.
Creating epochs without metadata works fine.

(I tried to comment out the size check in _check_metadata, but it breaks later then.)

File "/home/sh254795/anaconda3/lib/python3.7/site-packages/mne/utils/mixin.py", line 358, in _check_metadata
% (len(metadata), len(self.events)))

ValueError: metadata must have the same number of rows (38) as events (86)

Have there been any changes to this code that could explain the error?

Example:

import os
import numpy as np
import mne
import pandas as pd

sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder, 'MEG', 'sample',
                                    'sample_audvis_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file, verbose=False)
raw.crop(tmax=60).load_data()

# find all events
events = mne.find_events(raw, stim_channel='STI 014')

# define events of interest 
event_id = dict(A=4, B=3)

tmin = -0.1
tmax = 0.5

#%% create metadata for the events of interest

# count events of interest 
n_events = sum(events[:,2] == 3) + sum(events[:,2] == 4)

# create arbitrary metadata
METADATA = np.arange(0,n_events)
 
META_DF = pd.DataFrame(data=METADATA, columns = ['trial'])
   
# Epoch the data without metadata - FINE
print('creating epochs...')
epochs = mne.Epochs(raw, events, event_id, tmin, tmax)
      

# Epoch the data with metadata - BREAKS
print('creating epochs...')
epochs = mne.Epochs(raw, events, event_id, tmin, tmax, metadata=META_DF)

Start with the reproducible mne.Epochs call in the issue and inspect _check_metadata in mne/utils/mixin.py, comparing the metadata row count with the selected events. Verify the intended behavior for metadata built only from event codes 3 and 4, then confirm the corrected behavior with a regression test or the provided example.