What about introducing DOM-based translation? The current system would probably need to remain, as DOM doesn’t cover everything (the cases it doesn’t cover are basically the cases where <translate nowrap> is used), but most cases could be covered in a very Parsoid-friendly way (and actually cases it wouldn’t cover are mostly cases where translation directly using VE is not daunting, but impossible: template parameters, extension tag content, HTML attributes etc.). For example, changes could be:

Old:

<translate><!--T:1--> Some.</translate> <translate><!--T:2--> Content.</translate>

New:

<span data-translate-id="1">Some.</span> <span data-translate-id="2">Content.</span>

Old:

<translate>
<!--T:1-->
Some. Content.
</translate>

New:

<p data-translate-id="1">Some. Content.</p> <!-- or <div>? -->

Old:

* <translate><!--T:1--> Item</translate>
* <translate><!--T:2--> Item</translate>
* <translate><!--T:2--> Item</translate>

New:

<ul>
<li data-translate-id="1">Item</li>
<li data-translate-id="2">Item</li>
<li data-translate-id="2">Item</li>
</ul>

Old (before marking for translation):

<translate>
Some. Content.
</translate>

New (before marking for translation):

<p data-translate-id>Some. Content.</p>