Interesting - I hadn't really noticed how pronounced this dichotomy within dynamic languages is. On the one hand there are small languages designed for embedding and sandboxing (e.g. JavaScript, Lua, Tcl) and on the other hand, larger, more general-purpose languages (Perl, Python, Ruby).
I always assumed that a dynamic language could work well in both contexts, but in fact, most lie fundamentally on one side of the divide or the other. Only JavaScript, due to its immense popularity, has really managed (with Node.js) to expand from the first category into the second.
If Python were to be embedded in Excel, it would be expanding from the second category into the first. As you mention, to do this safely it may be necessary to create a special (subset) version of the language. Matz, the creator of Ruby, is trying to take his language in this direction with mruby [1] - a "lightweight implementation of Ruby complying to (part of) the ISO standard".
But, will these subset versions ever be popular? They necessarily leave the majority of the language's ecosystem behind - and knowledge of the full language will not necessarily transfer directly to the subset. Can a subset of an existing general-purpose language, even a widely-known one, compete against other languages that are specifically designed for safe embedding?
More generally, is it possible for a dynamic language to work well on both sides of the divide, or must all (even brand-new) languages choose one side or the other?