fix(p2p): gossip tx validation stalls behind tx pool finalization (A-1656)

[spalladino]

Fixes A-1656 (mainnet nodes logging `Gossip validation for tx took 20000+ms, approaching mcache eviction window of 8400ms`).

Root cause

Gossip tx validation waits on the tx pool's single serial queue (`canAddPendingTx` / `addPendingTxs`). `handleFinalizedBlock` occupied that queue as one item for the entire finalization of an epoch's worth of mined txs — hydrating each tx from LMDB, deserializing it, re-serializing it into the archive, and deleting it. Mainnet log analysis shows the slow-validation spikes cluster exactly on the once-per-epoch finalization tick (every 38.4 min, ~100–200 txs per tick); the worst spike (43s) coincides with the snapshot pause of archiver/world-state sync.

Reproduced in a unit test (tx_pool_v2_finalization.test.ts): finalizing 128 mined txs of ~1.1MB each stalls a concurrent canAddPendingTx for 3.9s on a local NVMe machine with archiving enabled — on 0.5–2 vCPU cloud nodes with network disks the observed 10–40s is consistent.

Chonk verification is ruled out: the batch verifier shows no pile-up (queue wait avg 6ms, max 172ms on next-net), and a prove-n-tps-real bench run from this PR pushed 891 real-proven txs at ~1 TPS through real node-side chonk verification with zero slow-validation warnings (tx validation ~170–440ms throughout).

Fixes

• Chunked finalization: handleFinalizedBlock now runs as chunk-sized serial-queue items (prepare → archive per chunk → delete per chunk → complete), so gossip pool operations interleave with finalization. Post-fix, the repro scenario stalls canAddPendingTx for 1ms (was 3.9s). Hardened per review: delete chunks re-check each tx is still mined at or before the cutoff, archiving is idempotent (a retried finalization can no longer create a duplicate FIFO index entry whose eviction corrupts the newer one), and concurrent finalizations are chained.
• Raw-buffer archiving: the pool already stores txs proof-stripped, which is exactly the archive format — the archive now copies buffers as-is instead of Tx.fromBuffer → withoutProof() → toBuffer() per tx. Total finalization time drops 3.9s → 395ms in the repro.
• Latent store deadlock: TxArchive.getHeadIndex/getTailIndex called .next() once on entriesAsync and abandoned the generator. Inside a write transaction the committed-state iterator is unbounded (WriteTransaction#iterate does not pass the limit down), so the abandoned generator never sent CLOSE_CURSOR, permanently leaking one of the store's maxReaders - 1 cursor-semaphore slots per call — deadlocking the whole store (and with it the pool serial queue) after 8 archive transactions. Now consumed via for await.
• Release image packaging (v5 line): standard-contracts/artifacts-historical/ was missing from the release image dockerignore whitelist, so every node started from a v5-line release image since fix(pxe): support contracts built with the v5.0.1 handshake registry #25032 (Jul 29) crash-looped with ERR_MODULE_NOT_FOUND — this broke all spartan network deploys from v5-next.

Instrumentation

• Per-stage timings for gossiped tx validation (deserialize, stage1_setup, stage1, pool_precheck, proof_verify, pool_add), attached to the slow-validation warn log and exported as aztec.p2p.gossip.tx_validation_stage_duration.
• Tx pool serial queue wait/execution histograms per operation plus queue length gauge (aztec.mempool.tx_pool_v2.queue_*).
• The peer BatchChonkVerifier now records the standard IVC verifier metrics (it previously recorded none).
• The spartan TPS benchmarks (n_tps, n_tps_prove) scrape and log all of the above at the end of a run.

Network-level validation (ci-network-bench runs on this PR)

With the fix, finalization queue items execute in ~10ms and no slow validations cluster on finalize ticks; canAddPendingTx/addPendingTxs wait ~10–19ms P95 across all TPS scenarios. The remaining serializer surfaced by the new metrics is handlePrunedBlocks (~57s on the queue during a reorg, 24 slow validations) — tracked separately as A-1668.