When word_timestamps is False, segments do not have a words key. The cleanup pass for instantaneous or empty segments was unconditionally assigning segment['words'] = [], introducing the key only on those empty segments. This makes the segment schema inconsistent: some segments get a words key, others do not.

SubtitlesWriter (used by both WriteVTT and WriteSRT) decides which iteration path to take by checking 'words' in result['segments'][0]. If the first segment happens to be empty or instantaneous, that spurious key flips the writer into the word-level path, which then raises KeyError: 'words' on the next non-empty segment that lacks the key.

Minimal reproducer:

from whisper.utils import WriteSRT
import io

result = {
    'segments': [
        {'id': 0, 'seek': 0, 'start': 0.0, 'end': 0.0, 'text': '',
         'tokens': [], 'words': [],  # would be set by transcribe.py
         'temperature': 0.0, 'avg_logprob': -0.5,
         'compression_ratio': 1.5, 'no_speech_prob': 0.1},
        {'id': 1, 'seek': 0, 'start': 1.0, 'end': 3.0, 'text': 'Hello',
         'tokens': [1, 2, 3],  # no 'words' key, word_timestamps=False
         'temperature': 0.0, 'avg_logprob': -0.5,
         'compression_ratio': 1.5, 'no_speech_prob': 0.1},
    ],
}
WriteSRT('/tmp').write_result(result, file=io.StringIO())
# KeyError: 'words'

Fix: only reset words when the key is already present, matching how add_word_timestamps populates it. Behavior when word_timestamps=True is unchanged (the key is present, gets cleared to [] as before).