fix(fund): assign multicall3 nonces up front to stop batches replacing each other

[praetoriansentry]

Description

polycli fund sends one multicall3 transaction per batch of wallets, each from its own goroutine. Every goroutine left Nonce nil, so each one asked the node for the pending nonce on its own. Two batches that read the same value collided: the second broadcast was rejected with replacement transaction underpriced and its wallets were never funded. If the second batch's fee estimate happened to come out higher, it instead silently replaced the first, and the receipt wait blocked on a hash that would never be mined. A 10,000-wallet run lost three of 25 batches this way.

Changes:

• Read the funder's pending nonce once and preassign a nonce to every batch. The concurrent fan-out stays, but the goroutines no longer touch the nonce endpoint, so neither timing nor load-balanced RPC routing matters.
• Retry each send up to three times with backoff before treating its nonce as a gap.
• When a send still fails, confirm the transactions below the gap, then return an error listing the batches that were not sent, how many accounts were not funded, and that the higher-nonce transactions stay pending until the gap is filled (pointing at polycli fix-nonce-gap). Previously a single send failure returned before confirming anything.
• ERC20 mode: a single approval for the total amount now covers every batch, instead of one approval per batch that each waited to be mined. util.Multicall3FundAccountsWithERC20Token no longer approves; the new util.Multicall3ApproveERC20Token does, and it returns an error when the approval is mined but fails instead of returning nil.
• Progress logs now report batch=N of=M nonce=... accounts=... instead of a wallet count that arrived out of order.

Breaking: the signature of util.Multicall3FundAccountsWithERC20Token changed. Its only caller was cmd/fund.

Jira / Linear Tickets

• N/A

Testing

• go build ./..., go vet ./..., shadow analyzer on changed packages
• New unit tests for batch splitting, gap detection, and the failure summary (go test -race ./cmd/fund/)
• Manual run against a live network with --key-file of 10,000 wallets (native token), confirming all 25 batches land and no replacement transaction underpriced errors
• Manual run in ERC20 mode (--token-address) confirming a single approval followed by the batch transfers

🤖 Generated with Claude Code