fix: harden non-deterministic eth_call reads to prevent silent data corruption (#248)

[lukiod]

Summary

Fixes the root cause of POI divergence documented in issue #248: eth_call state reads that return wrong-but-valid values (from a degraded or lagging RPC backend) silently corrupt indexing state. The existing PR #249 fixed the abort on null Pool but did not address the silent corruption from bad RPC responses.

Changes

roundsManager.ts — newRound() handler

1. Use try_getNextTranscoderInPool() instead of the bare call (case 5): A reverted eth_call during transcoder enumeration is now logged as log.critical and breaks the loop, rather than silently returning EMPTY_ADDRESS and truncating the pool list.

2. Upgrade log.info → log.critical for getFirstTranscoderInPool and getTotalBonded reverts (cases 1, 2): Post-launch these reverts are impossible and indicate a determinism hazard (bad RPC backend), not legitimate control flow.

3. Carry forward protocol.totalActiveStake when getTotalBonded reverts or returns zero (case 2): Previously zeroing the value cascaded into zeroed participationRate and numActiveTranscoders for the entire round.

helpers.ts — getPriceForPair()

4. Cache last-known non-zero pool prices (DAI/ETH, LPT/ETH) in module-level variables (case 3): On eth_call revert, the cached price is carried forward instead of silently returning ZERO_BD, which would permanently under-count USD volume in 7 accumulators.

5. Only cache non-zero prices to avoid persisting the "pool not yet deployed" sentinel as the last-known value.

Verification

• yarn codegen succeeds
• graph build compiles without errors
• Existing test suite (docker compose -f docker-compose.test.yml up) requires a local graph-node deployment

References

• Issue #248 — Non-deterministic eth_call reads silently corrupt indexing (POI divergence)
• PR #249 — Previous fix for the null-Pool abort (merged)
• Fork PR #2 — Documents the cumulative factor corruption bug introduced by fix: create missing Pool on demand to avoid null-Pool abort #249 (separate from this fix)

[lukiod]

Pushed a fix (dd65b9e) and have one thing worth a second look before merge.

Fixed: LPT/ETH price cache did not survive a restart

flowchart LR
    A[slot0 reverts] --> B{in memory cache?}
    B -- yes --> C[use it]
    B -- no, just restarted --> D{protocol.lptPriceEth?}
    D -- yes, new --> E[use it]
    D -- no --> F[fall back to zero, old bug]

Loading

lastKnownLptEthPrice resets to zero on every restart (fresh WASM instance). Added the D/E step, reusing the entity backed field already written each round in roundsManager.ts. yarn codegen and graph build both pass.

Worth a look: log.critical on cases 1 and 5

flowchart LR
    R[eth_call reverts] --> Q{which case}
    Q -- "2, 3, 4" --> G[carry forward last value, indexing continues]
    Q -- "1, 5" --> H[log.critical, subgraph halts, needs manual redeploy]

Loading

log.critical stops the whole indexer, not just this call, see graphprotocol/graph-node#4902. Cases 2 to 4 degrade gracefully on the same kind of revert; 1 and 5 hard stop instead, on the assumption that revert is "impossible post launch." If that ever turns out wrong once, the blast radius is total outage instead of one off bad data. Worth making a deliberate call rather than an implicit one. Not blocking.