Recovery from data model errors

[catamorphism]

I think we haven't decided yet whether data model errors should be recoverable. The current version of formatting.md says they are not recoverable:

If the message being formatted has any Syntax Errors or Data Model Errors, the result of pattern selection MUST be a pattern resolving to a single fallback value using the message's fallback string defined in the formatting context or if this is not available or empty, the U+FFFD REPLACEMENT CHARACTER �.

However, #703 would delete "data model errors" from that text.

So, do we want to require the behavior as currently specified; or leave it up to the implementation whether to replace a smaller piece of the message with a fallback, or replace the entire message with a fallback?

[macchiati]

This feels like a post-45 issue.