[aaronkanzer]

In EDF+ files, most instantaneous annotations have no duration field, and pyedflib reports a missing duration as -1. EDFRawIO only treats a duration of exactly 0 as an event, so these annotations land on the "Epoch" channel with a duration of -1 s and the "Event" channel stays empty. Two smaller problems come with it:

• The time window uses strict bounds, so an annotation at t=0 is dropped.
• event_count returns the total number of annotations for both channels, so for a file with annotations neo's own rawio_compliance.read_events fails with "Wrong shape 0, 3".

import numpy as np, pyedflib
from neo.rawio import EDFRawIO

w = pyedflib.EdfWriter("t.edf", 1, file_type=pyedflib.FILETYPE_EDFPLUS)
w.setSignalHeaders([dict(label="ch", dimension="uV", sample_frequency=100,
                         physical_min=-100, physical_max=100,
                         digital_min=-32768, digital_max=32767)])
w.writeSamples([np.zeros(1000)])
w.writeAnnotation(0, -1, "start")     # no duration field in the file
w.writeAnnotation(2.5, -1, "stim")
w.writeAnnotation(0, 1.5, "baseline")
w.close()

r = EDFRawIO("t.edf"); r.parse_header()
for i in range(2):
    print(r.event_count(event_channel_index=i), r.get_event_timestamps(event_channel_index=i))

Observed (0.14.6 and master):

3 (array([]), None, array([]))
3 (array([2.5, 0. ]), array([-1. ,  1.5]), array(['stim', 'baseline']))

Expected:

2 (array([0. , 2.5]), None, array(['start', 'stim']))
1 (array([0.]), array([1.5]), array(['baseline']))

MNE reads the same file as start and stim with duration 0 and baseline with 1.5.

Fix: classify an annotation as an event when its duration is <= 0 and as an epoch otherwise, make the window bounds inclusive, clamp negative durations when computing an annotation's end, and count events and epochs separately. A regression test builds the file with pyedflib. This is separate from the gain change in #1920, which touched the same file but not the annotation handling.