[Dev-next-gen]

In the transcription viewer, editing a segment's start or end cell and typing 00:00:05.5 stores 5005 ms, not 5500 ms. On focus out the cell is redrawn as 00:00:05.005. Typing 00:00:05.1234 stores 6234 ms, so the extra digit pushes the segment more than a second later.

parse_timestamp reads everything after the dot as a whole number of milliseconds (int("5"), int("1234")), whatever its length. The editor itself writes the fraction as three digits (to_timestamp), so 05.5 and 05.005 end up as the same value even though the cell shows 05.500 for half a second.

The fix reads the fraction as a decimal: it is padded or cut to three digits before converting, so .5 is 500 ms, .45 is 450 ms and .1234 is 123 ms. Three-digit input such as 01:23:45.678 gives the same result as before.

I added parser tests for short and long fractions and a TimeStampLineEdit test that types 00:00:05.5 and fires a focus out. Before the change the three new tests fail (5005 != 5500, 6234 != 5123, and the widget's get_milliseconds() returns 5005). After it they pass, along with the other tests in transcription_segments_editor_widget_test.py (44 passed).

Found by a defect-hunting pipeline I build and run (Dev-next-gen), using Claude Code with Anthropic's Claude Opus 5.5.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.20%. Comparing base (053b388) to head (8df1b6c).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1660      +/-   ##
==========================================
- Coverage   83.88%   83.20%   -0.69%     
==========================================
  Files         115      115              
  Lines       13834    13835       +1     
==========================================
- Hits        11605    11511      -94     
- Misses       2229     2324      +95     

Flag	Coverage Δ
Linux	83.20% <100.00%> (+0.04%)	⬆️
macOS	?

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Dev-next-gen]

The red test (macos-latest) stops on tests/widgets/speaker_identification_widget_test.py::TestSpeakerIdentificationWidget::test_identification_worker_run (assert 0 == 1). The worker logs Can't load tokenizer for 'kredor/punctuate-all' right after fetching that model from the Hugging Face cache during the test (job log). This PR only touches the timestamp editor and its tests. Because pytest.ini has -x, the run ended there at 978 passed, before it reached transcription_segments_editor_widget_test.py. On Ubuntu (1368 passed) and Windows (1316 passed), the four timestamp tests show PASSED.

I couldn't find a main run where this test fails: the last two test (macos-latest) runs on main and #1659 passed. So I read it as a download hiccup, not as something this change can cause. Re-running that one job would settle it, if you have a moment.

Found by a defect-hunting pipeline I build and run (Dev-next-gen), using Claude Code with Anthropic's Claude Opus 5.5.