Cap dice count, validate mi against die size, clarify two errors master
Four findings from the dice branch's final review. `4294967295d6` aborted the process trying to allocate billions of dice; parse_group now rejects any count over MAX_DICE (100), the same place the existing zero-count check lives. `d6mi100` parsed and silently produced a total no d6 can roll; parse_group now rejects a minimum greater than the group's die size. The MAX_REROLLS comment cited `d1r1` as an example of an unfailable condition, but d1 was dropped from VALID_SIDES earlier on this branch and is no longer valid notation; reworded to use only the still-valid r>0 example. And `2d6+5000000000` reported "is not a whole number" for a modifier that plainly is one; parse_term now checks the parse error's kind and reports "too large" instead when the value overflowed u32, so a model reading the error has something it can actually fix.