While I agree with you that the ballot should be physical, electronic counting defeats the purpose of transparency. Counting should also be manual in the presence of representatives of each candidate. Marking of ballots and voter eligibility checking do not present any issues if done electronically.
Maybe one way to address the issues of electronic counting would be to have parallel counts by each candidate (and their reps) using their own hardware and software with a manual count in case the electronic counts of every candidate do not match. But this means each candidate has to bring their own counting infrastructure which is prohibitive.
Another nice method to reduce risk from electronic counting is called the "Benaloh Challenge" (after Josh Benaloh, the inventor). The idea is that there are two steps to putting the paper ballot into the machine: first, the machine precommits to an encryption of its count (e.g. by printing some paper with evidence on it), and then the voter decides whether to spoil the ballot or actually cast it. If the voter spoils the ballot, then they get a new paper ballot to vote on, but they can retain the commitment. Voters may spoil any number of ballots. Decryptions of spoiled ballots along with enough information to check the machine's work are provided to the voter, either immediately or at the end of the day. This means that a cheating machine cannot cheat very much, though the whole thing also really relies on a verifiable privacy-preserving audit trail for the actual count (e.g. with homomorphic encryption). It at least means that nobody need trust the actual computers.
I have often imagined having say two scanners in physical sequence (so passing through a ballet is a single operation), and having more than one (competing) recorder for the output of each scanner head.
This would require some standardization of form factors and interfaces, but with small enough electronics units and scanners, in principle parallel data paths could be included in a single intentional "infrastructure".
Maybe one way to address the issues of electronic counting would be to have parallel counts by each candidate (and their reps) using their own hardware and software with a manual count in case the electronic counts of every candidate do not match. But this means each candidate has to bring their own counting infrastructure which is prohibitive.