I think you are ignoring the main problem; what UX do you want to offer to your users?

But is the text that the user edits in any way valid or is any resemblance to syntactic credibility purely coincidental?

Allowing text entry is useful and Sirius surely has multi-line string editing, but as soon as you need semantics you start to need language tooling and that needs to integrate since presumably a name in the text needs to be spelled consistently wrt the declaration of that name elsewhere in the graphics so you need a conversion from your graphical model to one understood by the tooling

Your original question contrasting EOL and OCL suggests you have some flexibility so I must suggest that you consider the better integration of Xbase.

The EmbeddedXtextEditor in the OCL project was based on an EclipseLabs prototype. In principle it should work with any Xtext-defined language. I suspect that the EclipseLabs prototype has now forked a few times so you should find a version that already integrates with Sirius. So you are just left with the problem of converting the EVL grammar to Xtext form and integrating all the scope resolutions etc with your outer context.

(You will find that developing the embedded Xtext is much easier if you define a textual grammar for the graphics too, so that at least for testing you have a text-only Xtext regime. Your users may even like the interchangeable alternative of all-text rather than mixed text-and-graphics for some designs.)