Thank you for your interest in Marpa, your kind words, your interesting example and your careful write-up of it.
Briefly, precedenced statements work by applying precedence and associativity to instances of the LHS symbol on their RHS. In the 2nd example, <expr> occurs outside of the precedenced statement and accordingly precedence and associativity are not applied to those instances of <expr>, resulting in ambiguity.
What is going on is that precedenced rules get rewritten into BNF rules which enforce the associativity and precedence declared in the precedenced rule. BNF rules do not get rewritten.
Basically, the LHS of a precedenced rule should never occur in any rule which is a step of a derivation from the precedenced rule -- while in theory it could have a meaning, in practice it almost always will not do what the grammar writer actually intends.
In R3/Kollos, I will probably have a warning issue on likely problematic uses of the precedenced LHS outside of the precedenced rule. Or I may simply ban them.