Proposer silently skips its slot after every proposer-rotation wrap (slot ≡ 1 mod 512); no failure counted

Tái hiện mẫu slot trên thiết lập devnet-eth35 và so sánh GET /lean/v0/blocks/ với các slot đối chứng liền kề. Bắt đầu từ các đường dẫn build của proposer và interval-4 tạo ra log “Building block” và metric lean_block_building_failures_total. Hoàn tất khi các slot proposer bị bỏ lỡ được đếm và các miss mang tính xác định sau wrap hiển thị trong các metric.

Observed

On a 512-validator / 64-node devnet (image devnet5-leanvm-main, rev 691df39, 8 validators/node, 8 subnets), validator 1 has missed 100% of its proposal slots: 8/8 at first measurement (slots 1, 513, 1025, 1537, 2049, 2561, 3073, 3585 — every slot ≡ 1 mod 512), and still 5/5 when re-sampled four days and ~170 rotations later (slots 84481…86529 all absent, adjacent control slots present).

The node never logs Building block for these slots, and lean_block_building_failures_total stays 0 fleet-wide: the interval-4 build tick never fires, so nothing counts the miss.

Context

• The rotation-wrap block (slot ≡ 0 mod 512, proposer = validator 0) is imported by peers 7–8 s after its slot start at every observed wrap (7.3 s / 7.2 s / 7.8 s at slots 512 / 3072 / 3584).
• node_1 (host of validator 1, also an aggregator) is seconds behind real time when its slot arrives: its attestations for slot 512k publish 3–7 s late and its aggregation sessions run ~3 s.
• At 3 of 5 recently sampled wraps the following proposer (validator 2's node) also lost its block — net cost 1–2 blocks per 512-slot cycle, invisible to every failure metric.

Expected

A proposer that cannot build in time should at least increment a failure counter; a deterministic 100%-miss pattern tied to the proposer-rotation wrap should be visible in metrics.

Evidence

• Chain: devnet-eth35, genesis 2026-08-28 16:53:53Z (GENESIS_TIME 1787936033), hosts ethlambda-3/ethlambda-5.
• Missed-slot verification via GET /lean/v0/blocks/<slot> (404 for every s ≡ 1 mod 512 sampled; 200 for controls).
• Full measurement context: 3.9 h instrumented window + 4-day soak; per-node logs and Prometheus/Loki history retained on the hosts.