gas bench: over/under split + lag=2 secondary grade (methodology v2 seed)

[Flotapponnier]

Delivers the two red-team recommendations that don't require touching the primary ranking.

Over/under split (Phase C)

Emits gas_error_priority_over_gwei_histogram and gas_error_priority_under_gwei_histogram alongside the existing abs histogram. Same buckets. Only the branch the oracle actually erred on gets a sample.

Why: percentile trackers (PublicNode feeHistory) and inclusion-confidence oracles (Etherscan, MetaMask 'high') can post the same abs p99 gap but with opposite user consequences — over-prediction = overpay a few wei, under-prediction = tx stuck through the next spike. The leaderboard can now decompose these instead of collapsing them.

Lag=2 secondary grade (Phase D seed)

New dual buffer in the harness so every prediction is graded twice — once at the target block (primary, unchanged) and once at target+2 (new histogram gas_error_priority_lag2_gwei_histogram).

Why: grading against the very next block rewards whichever oracle scraped the mempool 100 ms before we did — a latency race, not accuracy. Block+2 matches actual wallet UX (sign, broadcast, propagate).

Verified locally

60s smoke run:

• over/under counters populated per (oracle, tier, chain) with the expected asymmetric distributions (Etherscan under_count > over_count on ETH-p50 during current low-vol period; PublicNode balanced)
• lag2 histogram counts increasing on all (oracle, tier, chain) pairs
• Build clean, vet clean, spec validation ✓

Primary p99 ranking is unchanged. Leaderboard columns can be added later once the histograms accumulate 24h of data.