Summary

Two optimizations for Anvil fork mode that dramatically reduce eth_call latency for Multicall3 requests:

• Multicall3 Prefetch: When eth_call targets Multicall3 aggregate3, prefetches all touched storage slots via one eth_createAccessList + concurrent eth_getStorageAt into the BlockchainDb cache before EVM execution. Turns serial RPC round-trips into parallel batch fetches. Best-effort — failures silently fall back to on-demand fetching. Local dirty state always takes precedence over prefetched data.

• Multi-RPC Load Balancing: --fork-url now accepts multiple values. A new LoadBalancedTransport distributes requests via atomic round-robin across backends. The existing RetryBackoffLayer naturally fails over to the next backend on error.

Performance

9-subcall multicall against BSC mainnet fork with 3 RPC endpoints:

Usage

# Multi-RPC load balancing
anvil --fork-url https://rpc1.example.com --fork-url https://rpc2.example.com

# Single URL still works identically
anvil --fork-url https://rpc1.example.com

Test plan

• Unit tests: Multicall3 ABI decode (3 tests), LoadBalancedTransport round-robin (3 tests)
• Integration: Multicall3 eth_call succeeds in fork mode
• Data correctness — clean state: multicall returns correct upstream balance
• Data correctness — dirty state: after anvil_set_balance, multicall returns local modified value (NOT prefetched upstream value)
• Data correctness — mixed state: single multicall with both dirty + clean addresses returns correct values for both
• Integration: multi-URL fork bootstraps and reads state
• All existing fork tests pass (68/68, 6 pre-existing flaky failures unrelated)

Closes #14265