fix(runtime): exact-slot eth_getStorageAt, eth_getCode fallback for create2, and receipt-poll retry

[luiz-lvj]

Summary

Three independent runtime/bridge fixes surfaced while running the OZ tron-contracts suite against TRE. Together they clear ~25 deterministic failures (proxy/upgradeable + create2) and stop a transient socket drop from cascading an entire test worker.

1. stubProvider.getStorage — read the EXACT slot via eth_getStorageAt

getStorage POSTed debug_storageRangeAt(0, 0, addr, slot, 1) to /tre and returned Object.values(storage)[0] — i.e. the FIRST slot the node yields, not the requested one (java-tron's storageRangeAt doesn't filter to the key). So ERC-1967 reads came back as slot 0: getAddressInSlot(proxy, ImplementationSlot) returned the implementation's value field (10/100/42) instead of the implementation address.

This cascaded: TransparentUpgradeableProxy's tests read the ProxyAdmin address from the admin slot (a deliberate TVM workaround), got garbage, then impersonated the wrong address — so every admin-gated upgradeToAndCall reverted with data=0x and proxyAdmin.owner() reverted as call reverted (data=0x).

Fix: read the exact slot via eth_getStorageAt(hexAddr, slotHex, 'latest') on /jsonrpc (the same endpoint maybeAutoRegisterContract uses for eth_getCode), plus the EVM-predicted → TVM-actual remap that getCode/getBalance already apply, and normalize the result with zeroPadValue(..., 32).

2. stubProvider.getCode — fall back to eth_getCode for CREATE2-opcode contracts

getContract reads ContractStore, which doesn't contain contracts deployed via the CREATE2 opcode (e.g. RelayedCall's assembly-create2 relayer, or OZ Clones) — those live in CodeStore. When getContract returns empty, fall back to eth_getCode on /jsonrpc, which reads CodeStore. Fixes provider.getCode(relayer) returning 0x.

3. wait.js _getInfoBigSafe — retry transient socket errors on the receipt poll

The receipt poll did a bare fetch + resp.text() with no retry, unlike rpcCall (which already retries via isTransientFetchError). A G1GC stop-the-world late in a long run drops the keep-alive socket → TypeError: fetch failed / "other side closed" → the test fails. Worse, since every later fixture's waitForReceipt (and refundSigners) hits the same node, one drop cascades into failing the rest of the worker. Wrap the fetch+body-read in the same single-retry-with-100ms-backoff and export isTransientFetchError from cheatcodes.js for reuse. (Resilience for transient drops; node-level memory pressure is a separate, host-side concern.)

Files

• src/runtime/ethers-bridge.js — getStorage (eth_getStorageAt), getCode (eth_getCode fallback)
• src/runtime/wait.js — _getInfoBigSafe transient retry
• src/runtime/cheatcodes.js — export isTransientFetchError