docs: correct the head-lag claim in the Serialized audit, add follow-ups

[Flotapponnier]

The correction

v1.0 of the audit claimed bench 001 references archive nodes and was therefore unaffected by providers disagreeing about on-chain time. That was wrong. It repeated benchmarks/aggregator-head-lag.yml instead of reading harnesses/aggregator-head-lag.

There is no archive-node reference in that harness. grep -rl "archive|eth_getBlockByNumber|getBlockTime|blockTimestamp" over its cmd/ returns nothing. The gauge behind the leaderboard is:

// head_lag_monitor.go:211  (Mobula)
onChainTime := time.UnixMilli(trade.Date)     // Mobula's own field
// head_lag_monitor.go:707  (Codex)
onChainTime := time.Unix(event.Timestamp, 0)  // Codex's own field

The spec claims otherwise in three places: methodology[7], the FAQ, and the per-chain explainers. Measured impact: for the same transaction hash, serialized.at minus mobula.date is −707 ms on Solana and +1000 ms on Base, so the published numbers partly measure clock placement rather than pipeline speed.

Plus head_lag_monitor.go:219 drops negative lags silently, which would discard 100% of Serialized's Base preconfirmed sample.

This is a documentation-versus-code defect on a live, publicly cited bench, independent of any provider. Not fixed here, only documented.

New results recorded

• Bench 067 now conclusive. The earlier inconclusive run used the wrong probe address; the harness pins canonical ones in registry.go. Re-run verbatim: listed 19, verified 5, 0 errors. Would rank last (CoinStats 127, Mobula 50, Zerion 42, Moralis 15).
• Negative capabilities probe-confirmed, not assumed: 404 on /v1/wallet/nfts, /v1/nft/collection, /v1/nfts, /v1/swap/quote, /v1/quote, /v1/route, /v1/bridge/quote.
• A real pricing defect. Serialized prices BONK at 6.33e-07 while Mobula, DexScreener (Orca, $306k liquidity) and GeckoTerminal all agree at ~3.31e-06. 5.2x low, $55.7M market cap against a real ~$290M. Root cause visible in their own /v1/token/pools: rank 1 is a pool with liquidity 4,339 and the deep Orca pool is absent from the list, while their own ranks 2 and 3 disagree with rank 1 by the same 5x.
• A cross-API price-accuracy bench is not proven. Two basket constructions, both recorded as failures so nobody repeats them. It cannot use another aggregator as reference; the reference has to be computed from on-chain reserves over an RPC we control.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CpArutAtXuBb1BVNUDXoYA