fix(evm-rpc): keep finalizer queue contiguous so finalized head advances past probe window

[elina-chertova]

Problem

For chains whose finalized head advances in large jumps (L2 / rollup batch finality — e.g. zksync-mainnet, worldchain-mainnet, alpen-testnet, robinhood-mainnet), the hot-block data service stops advancing its reported finalized head. hotblocks_last_finalized_block freezes while hotblocks_last_block (unfinalized head) keeps advancing, and the data service logs block finalization lags behind and prevents cache purging on every new head. This fires the Portal_Hotblocks_Head_Metrics_Absent_* alert (increase(hotblocks_last_finalized_block[12h]) == 0).

Observed in production: zksync-mainnet reported finalized frozen at 71020159 while the chain's real finalized head (eth_getBlockByNumber("finalized"), agreed by two independent providers) was 71027108 — a ~6,900-block gap that only grows.

Root cause

Finalizer.visit() in evm/evm-rpc/src/data-source/finalizer.ts bounds its probe queue by overwriting the last slot once it exceeds ~50 entries:

if (this.queue.length > 50) {
    this.queue[this.queue.length - 1] = getBlockRef(block)   // drops the middle
} else {
    this.queue.push(getBlockRef(block))
}

This keeps only the oldest ~50 unfinalized refs plus the single latest block, leaving a gap in between. The finalizer probes from the front and can only advance the finalized head through contiguous refs. When the finality lag exceeds the window, it finalizes the ~50 contiguous refs, hits the gap, and can never reach the true finalized head — so the reported finalized head freezes.

Chains with smooth, incremental finality (Ethereum PoS, most L1s) slide within the window and are unaffected — matching the fact that only large-jump-finality networks froze.

Fix

Keep the queue contiguous — always push. The queue is now bounded by the finality lag (the range we must track anyway), and it holds tiny BlockRefs (number + hash), so memory is negligible relative to the full-block buffer the data service already keeps.

Test

Added evm/evm-rpc/src/data-source/finalizer.test.ts. The regression test feeds >50 unfinalized blocks in one batch with the finalized head beyond the old cap and asserts the finalizer reaches the true finalized head.

• pre-fix: expected 50 to be 120 (frozen at the 50-block cap)
• post-fix: passes; full evm-rpc suite green (154 passed, 27 skipped), tsc build and tsc --noEmit -p tsconfig.test.json clean.

Falsification

If, after deploying an evm-data-service image built from this commit, a large-jump-finality network's hotblocks_last_finalized_block still stops advancing while its RPC finalized tag advances, this fix is insufficient. (Note: running pods must be redeployed to a rebuilt image — the fix cannot take effect via restart alone, since the same code re-freezes once the lag re-exceeds the window.)

[elina-chertova]

Related to #499 (also in finalizer.ts), but a different cause: #499 removes the finalized-only output.put that stalls head ingestion (the blockNumber % 32 lag spike); this PR fixes the queue.length > 50 overwrite that drops middle refs and freezes the reported finalized head on large-jump-finality chains. #499 leaves the queue-cap logic untouched, so the two are complementary.