Regression in content assist follow elements calculation

After updating to Xtext 2.25 (from 2.21) we see unwanted } proposals given the grammar

Specification:
	{Specification}
	'specification' name=ID '{'
	((a+=A) |
	(b+=B) |
	(c+=C) |
	(d+=D) |
	(e+=E) |
	(s+=S) |
	(f+=F) |
	(g+=G))*
	'}';

A:
	"A" name=ID;

B:
	"B" name=ID;

C:
	"C" name=ID;

D:
	"D" name=ID;

E:
	"E" name=ID;

Fqn:
	ID ('.' ID)*;

F:
	'pi' xxxxxx=ID name=ID '{'
	'}';

G:
	'something' 'pi' i=Fqn '{'
	'}';

S:
	'something' xxx=Fqn s=ID '{'
	'}';

and model

specification xxxx {
	something aaaa bbbb {
    }
}

invoking content assist at aaaa gives the unwanted } proposal.
debugging 2.21 vs 2.25 i can see a difference in
Dfa.error which leads to failedPredicateAtEOF be set to true respecively not set to true

which then leads to different paths in

org.eclipse.xtext.ide.editor.contentassist.antlr.internal.new StreamAdapter() {...}.announceEof(int)

cause seems to be this change:
https://github.com/eclipse/xtext-core/pull/1469

Start with the supplied grammar and model, then compare Xtext 2.21 and 2.25 around Dfa.error and failedPredicateAtEOF. Trace the content-assist path through StreamAdapter.announceEof(int) and review PR 1469. Done means invoking content assist at aaaa no longer proposes an unwanted closing brace.