fix(subtitle): convert wvtt cue times by the media timescale

Cue times were multiplied by the mdhd timescale instead of scaled to µs, so the pycaption path put cues hours out of sync for any timescale but 1000.

tests/tracks/test_subtitle_convert.py

@@ -351,6 +351,24 @@ def test_segmented_fvtt_parses_and_converts(tmp_path, monkeypatch):
     assert cue_count(out) == 2
 
 
+def with_mdhd_timescale(data: bytes, timescale: int) -> bytes:
+    """The fixture with its version-0 mdhd timescale replaced; sample times stay in the same units."""
+    at = data.index(b"mdhd") + 16
+    return data[:at] + struct.pack(">I", timescale) + data[at + 4 :]
+
+
+@pytest.mark.parametrize("timescale", [1000, 2000, 90000])
+def test_segmented_fvtt_cue_times_follow_the_media_timescale(timescale):
+    data = (FIXTURES / "segmented.wvtt.mp4").read_bytes()
+    assert with_mdhd_timescale(data, 1000) == data  # the helper's offset hits the fixture's timescale
+
+    captions, _ = Subtitle.merge_segmented_wvtt(with_mdhd_timescale(data, timescale))
+
+    # the fixture's cues span 1000-2000 and 3000-4000 media units
+    times = [t for c in captions for t in (c.start, c.end)]
+    assert times == pytest.approx([t * 1_000_000 / timescale for t in (1000, 2000, 3000, 4000)])
+
+
 def test_segmented_fttml_joins_a_cue_split_at_a_segment_boundary():
     data = build_stpp_mp4(
         ttml_fragment(("00:00:01.000", "00:00:02.000", "Carried")),
@@ -449,17 +467,25 @@ def test_dash_fttml_cues_lose_the_presentation_time_offset(tmp_path, monkeypatch
     assert (caption.start, caption.end) == times
 
 
-def test_dash_fvtt_cues_lose_the_presentation_time_offset(tmp_path, monkeypatch):
-    template = '<SegmentTemplate timescale="1000" presentationTimeOffset="1000" duration="2000"/>'
-    data = (FIXTURES / "segmented.wvtt.mp4").read_bytes()
+@pytest.mark.parametrize("method", ["subby", "pycaption"])
+@pytest.mark.parametrize(
+    "mdhd_timescale, pto, cues",
+    [
+        (1000, 1000, ("00:00:00.000 --> 00:00:01.000", "00:00:02.000 --> 00:00:03.000")),
+        (2000, 500, ("00:00:00.000 --> 00:00:00.500", "00:00:01.000 --> 00:00:01.500")),
+    ],
+)
+def test_dash_fvtt_cues_lose_the_presentation_time_offset(tmp_path, monkeypatch, method, mdhd_timescale, pto, cues):
+    template = f'<SegmentTemplate timescale="1000" presentationTimeOffset="{pto}" duration="2000"/>'
+    data = with_mdhd_timescale((FIXTURES / "segmented.wvtt.mp4").read_bytes(), mdhd_timescale)
     sub = dash_subtitle(tmp_path, monkeypatch, Codec.fVTT, data, template)
+    monkeypatch.setattr("unshackle.core.config.config.subtitle", {"conversion_method": method}, raising=False)
 
     sub.download(None, None)  # type: ignore[arg-type]
 
     assert sub.codec == Codec.WebVTT
     text = sub.path.read_text("utf8")
-    assert "00:00:00.000 --> 00:00:01.000" in text
-    assert "00:00:02.000 --> 00:00:03.000" in text
+    assert all(cue in text for cue in cues)
 
 
 def ttml_fragment(*cues: tuple[str, ...]) -> str:

unshackle/core/tracks/subtitle.py

@@ -1111,8 +1111,8 @@ class Subtitle(Track):
 
                     if nodes:
                         caption = Caption(
-                            start=start_time * timescale,  # as microseconds
-                            end=end_time * timescale,
+                            start=start_time * 1_000_000 / timescale,
+                            end=end_time * 1_000_000 / timescale,
                             nodes=nodes,
                             layout_info=layout,
                         )