feat(node): serve fee RPCs from a background L1 fee snapshot

[spalladino]

Makes warm getPredictedMinFees/getCurrentMinFees calls zero-L1-request in-memory reads, so quote latency no longer scales with L1 RPC latency.

Context

Every getPredictedMinFees call paid two uncached eth_blockNumber round trips, and each new L1 block added up to ~8 sequential eth_call waves on the request path. With 100-150 ms of L1 RPC latency the quote RPC took ~800 ms — right when a wallet is preparing a transaction. Prediction fee configuration (manaTarget, proving cost) was also memoized for process lifetime, so a running node never saw governance updates to it.

Approach

• A background FeeSnapshotService (owned by the node factory, shared by the fee RPCs and the p2p mempool fee policy) refreshes once per L1 block, pinned to the archiver's atomically published L1 identity (new Archiver.getL1SyncSnapshot()), and publishes an immutable snapshot of complete precomputed quotes (current fee + all three ManaUsageEstimate prediction arrays) per candidate slot. Warm reads are pure memory lookups.
• A refresh batches its reads into two pinned Multicall3 eth_calls — tips + governance values + per-slot getManaMinFeeAt/canPruneAtTime/getL1FeesAt, then both checkpoint fee headers plus a getTips re-read that discards the refresh on cross-wave mismatch — followed by an exact-coverage top-up and an archiver identity re-check before the atomic publish. Falls back to parallel pinned individual calls where Multicall3 is not deployed.
• Read-time selection preserves both legacy anchor rules (current fee floors on the raw pending checkpoint slot; predictions on the pinned-block slot), evaluated across a small clock-drift window with element-wise max merging. Solidity getManaMinFeeAt stays canonical for the current fee; predictions reuse the existing TS fee math. Coverage misses trigger a shared single-flight refresh and never substitute a different slot's answer.
• Three independent fail-closed staleness bounds (refresh age, L1-head age, future-dated head), keep-last-good on refresh failure with backoff, and governance values read per refresh.
• The legacy computation is retained as an explicit-input (blockNumber, now) test oracle; an Anvil equivalence suite compares snapshot vs legacy on identical inputs across state/clock cases, and locks the within-slot fee constancy and checkpoint-slot invariant the design relies on. The RPC bench gates warm calls on the service counter (zero read-triggered refreshes) and includes a JSON-RPC latency-proxy fixture.

API changes

• getCurrentMinFees/getPredictedMinFees can now throw typed FeeSnapshot*Errors when the L1 view is stale or unavailable, instead of silently serving unboundedly stale data.
• RollupContract.getManaMinFeeAt third parameter changed from stateOverride to an options object { blockNumber?, stateOverride? }; getTips gains an optional { blockNumber }.

Fixes A-1416

Review fixes

A second-model review of the implementation surfaced three service-state defects, fixed in follow-up commits: publish() gated on block number and would wedge after an L1 height rollback (reorg or lagging fallback backend); staleness bounds were checked before the archiver identity, failing reads that one refresh would repair; and read-triggered refreshes bypassed the failure backoff, turning RPC traffic into L1 load during outages. Snapshots also now publish their actual materialized slot bounds, and fail-closed reads have counters.

Follow-ups (deliberately out of scope here)

• Plumb the service tuning/staleness bounds through AztecNodeConfig/env vars (defaults are currently derived from L1 constants and overridable only in code).
• Promote the FeeSnapshotStats counters and structured log fields to OpenTelemetry meters.
• Wire the latency proxy fixture into the bench for an end-to-end injected-latency scenario (the zero-fee-path-L1 property is currently gated on the service counter, which is authoritative but narrower).