add <track> for time-aligned media transcripts

[vagenas]

• New <track> semantic element: <bdiv/>-delimited cue blocks, each with a mandatory start time and optional end time, an optional <frame> and <audio>, and a transcript of <voice>-attributed turns
• Semantics: cue blocks cover inclusive [start, end] intervals (missing end means [start, start]) and are ordered by non-decreasing start time, overlaps allowed
• Validation: XSD elements/groups + top_level_cat entry + range facets; Schematron rules for cue-block structure, timestamp order, and block sequence
• Archive: whole-track media in audio/ and video/ ({N}.{ext} by <track> position); pack() and doclang pack gain --audio/--video; media content types added; archive-demo gains a track and a spoken audio/1.wav
• Docs/tests: spec ### Tracks section, regenerated ### Reference (drops the abandoned <timestamp>/<hour>/… draft rows), token vocab, and track + packaging fixtures

[PeterStaar-IBM]

lgtm!

[ceberam]

I added a couple of comments below.

And just a reflection on the design choice for <voice>, compared to WebvTT <v> span (not highlighting any defect though and going very fine-grain). In WebVTT, the <v speaker> span wraps the spoken text inline, so speaker and utterance are in the same node. In DocLang's <track>, the <voice> precedes the text as a sibling. This means <voice> has no explicit close: a turn runs until the next <voice> or </bdiv>. The structural intent is the same, but the representation differs and it may create some gaps:

• no XML structure encodes that boundary: a tree walker (from a consumer application) has to reconstruct it by walking forward from each <voice> until the next one
• mixed attributed and unattributed content within the same cue could not be captured, like in the 3rd block of WebVTT EXAMPLE 6.
• word-level timing is not expressible (this is not supported in Docling's WebVTT parser either, by the way), just at cue block interval.

[vagenas]

@ceberam thanks for your comments

• no XML structure encodes that boundary: a tree walker (from a consumer application) has to reconstruct it by walking forward from each until the next one

This is consistent with DocLang's delimiter + virtual-text pattern (e.g. in lists, tables)

• mixed attributed and unattributed content within the same cue could not be captured, like in the 3rd block of WebVTT EXAMPLE 6.

It actually can with an empty <voice/> — I now even refined the explanation in the spec.

[vagenas]

• word-level timing is not expressible (this is not supported in Docling's WebVTT parser either, by the way), just at cue block interval.

This can indeed be done via cue blocks (with hardly any information loss), e.g.

00:00:00.030 --> 00:00:02.669 align:start position:0%
the<00:00:00.389> quick<00:00:00.750> brown<00:00:01.110> fox

from #175 could become:

<doclang>
  <track>
    <bdiv/>
    <seconds value="0"/>
    <msecs value="30"/>
    <seconds value="0"/>
    <msecs value="389"/>
    the
    <bdiv/>
    <seconds value="0"/>
    <msecs value="389"/>
    <seconds value="0"/>
    <msecs value="750"/>
    quick
    <bdiv/>
    <seconds value="0"/>
    <msecs value="750"/>
    <seconds value="1"/>
    <msecs value="110"/>
    brown
    <bdiv/>
    <seconds value="1"/>
    <msecs value="110"/>
    <seconds value="2"/>
    <msecs value="669"/>
    fox
  </track>
</doclang>