Then it should use zero fees. In other cases, some miner may include it anyway, and then, it is no longer "free", when real satoshis are taken out of users. Also, the chain of transactions can form a package, so that after many free transactions, you can end the chain with some non-zero-fee transaction, bumping effective fee rate by CPFP. Also, by rejecting any free transactions upfront, miners can lose a lot of coins. This is how Garlo Nicon made his coins hard to steal: In practice, anyone could take 9950 tBTC. But because the initial transaction was free, and the next one paid absurdly high fees, he could be pretty sure, that nobody will take it, even if his node would happily share it with anyone, willing to request it. But nodes used 1 sat/vB as a minimum, so they didn't see the first transaction, so they never heard about the second one, and they became aware of it, only when they saw it included in a block. And even then, it was mined on just a CPU, but nobody was fast enough to notice it, and reorg the chain, before ASICs set all of that in stone, by mining on top of his block. This attack can be made stronger, by using different sighashes, than SIGHASH_ALL. Because then, after making a long chain of transactions, anyone would be able to break it at any point, just by adding more inputs or outputs. Normally, if signatures with different sighashes could be chained, it wouldn't be the case, but in existing network circumstances, if some previous output is referred by txid:vout, then exactly this txid:vout is actually signed, so if a single input or output in any previous transaction is altered in any way, then all following transactions are invalidated (which means they were propagated literally for free).
Then it should use zero fees. In other cases, some miner may include it anyway, and then, it is no longer "free", when real satoshis are taken out of users. Also, the chain of transactions can form a package, so that after many free transactions, you can end the chain with some non-zero-fee transaction, bumping effective fee rate by CPFP. Also, by rejecting any free transactions upfront, miners can lose a lot of coins. This is how Garlo Nicon made his coins hard to steal: In practice, anyone could take 9950 tBTC. But because the initial transaction was free, and the next one paid absurdly high fees, he could be pretty sure, that nobody will take it, even if his node would happily share it with anyone, willing to request it. But nodes used 1 sat/vB as a minimum, so they didn't see the first transaction, so they never heard about the second one, and they became aware of it, only when they saw it included in a block. And even then, it was mined on just a CPU, but nobody was fast enough to notice it, and reorg the chain, before ASICs set all of that in stone, by mining on top of his block. This attack can be made stronger, by using different sighashes, than SIGHASH_ALL. Because then, after making a long chain of transactions, anyone would be able to break it at any point, just by adding more inputs or outputs. Normally, if signatures with different sighashes could be chained, it wouldn't be the case, but in existing network circumstances, if some previous output is referred by txid:vout, then exactly this txid:vout is actually signed, so if a single input or output in any previous transaction is altered in any way, then all following transactions are invalidated (which means they were propagated literally for free).

Then it should use zero fees. In other cases, some miner may include it anyway, and then, it is no longer "free", when real satoshis are taken out of users. Also, the chain of transactions can form a package, so that after many free transactions, you can end the chain with some non-zero-fee transaction, bumping effective fee rate by CPFP. Also, by rejecting any free transactions upfront, miners can lose a lot of coins. This is how Garlo Nicon made his coins hard to steal: In practice, anyone could take 9950 tBTC. But because the initial transaction was free, and the next one paid absurdly high fees, he could be pretty sure, that nobody will take it, even if his node would happily share it with anyone, willing to request it. But nodes used 1 sat/vB as a minimum, so they didn't see the first transaction, so they never heard about the second one, and they became aware of it, only when they saw it included in a block. And even then, it was mined on just a CPU, but nobody was fast enough to notice it, and reorg the chain, before ASICs set all of that in stone, by mining on top of his block. This attack can be made stronger, by using different sighashes, than SIGHASH_ALL. Because then, after making a long chain of transactions, anyone would be able to break it at any point, just by adding more inputs or outputs. Normally, if signatures with different sighashes could be chained, it wouldn't be the case, but in existing network circumstances, if some previous output is referred by txid:vout, then exactly this txid:vout is actually signed, so if a single input or output in any previous transaction is altered in any way, then all following transactions are invalidated (which means they were propagated literally for free).

Then it should use zero fees. In other cases, some miner may include it anyway, and then, it is no longer "free", when real satoshis are taken out of users. Also, the chain of transactions can form a package, so that after many free transactions, you can end the chain with some non-zero-fee transaction, bumping effective fee rate by CPFP. Also, by rejecting any free transactions upfront, miners can lose a lot of coins. This is how Garlo Nicon made his coins hard to steal: In practice, anyone could take 9950 tBTC. But because the initial transaction was free, and the next one paid absurdly high fees, he could be pretty sure, that nobody will take it, even if his node would happily share it with anyone, willing to request it. But nodes used 1 sat/vB as a minimum, so they didn't see the first transaction, so they never heard about the second one, and they became aware of it, only when they saw it included in a block. And even then, it was mined on just a CPU, but nobody was fast enough to notice it, and reorg the chain, before ASICs set all of that in stone, by mining on top of his block. This attack can be made stronger, by using different sighashes, than SIGHASH_ALL. Because then, after making a long chain of transactions, anyone would be able to break it at any point, just by adding more inputs or outputs. Normally, if signatures with different sighashes could be chained, it wouldn't be the case, but in existing network circumstances, if some previous output is referred by txid:vout, then exactly this txid:vout is actually signed, so if a single input or output in any previous transaction is altered in any way, then all following transactions are invalidated (which means they were propagated literally for free).