[consensus] Strict leader + latency-weighted reputation with 1 faulty validator at 4k TPS

[danielxiangzl]

Summary

• Combines daniel/latency-strict-leader-by-skip (failure_window=50x) + daniel/latency-weighted-leader-by-skip (latency-weighted heuristic)
• Both strict failure penalization AND latency-weighted proposer selection active
• 1 faulty validator, 4k TPS, no fullnodes, 10 min
• Prototype/experiment code — not for merge to main

Test plan

• Run forge land_blocking and compare against baseline (PR [forge] Latency baseline with 1 slow validator at 4k TPS #19330), strict-only (PR [consensus] Strict leader reputation with 1 faulty validator at 4k TPS #19336), and weighted-only (PR [consensus] Latency-weighted leader reputation + tighter classifier (combined) #19341)

🤖 Generated with Claude Code

[github-actions]

✅ Forge suite realistic_env_max_load success on 29a7aef36aeb8ef414b76c18e86cf58c82bac579

two traffics test: inner traffic : committed: 4000.01 txn/s, latency: 410.80 ms, (p50: 300 ms, p70: 300, p90: 600 ms, p99: 1200 ms), latency samples: 74760
two traffics test : committed: 100.00 txn/s, latency: 343.99 ms, (p50: 300 ms, p70: 300, p90: 400 ms, p99: 1000 ms), latency samples: 4260
Latency breakdown for phase 0: ["MempoolToBlockCreation: max: 0.151, avg: 0.123", "ConsensusProposalToOrdered: max: 0.157, avg: 0.153", "ConsensusOrderedToCommit: max: 0.016, avg: 0.015", "ConsensusProposalToCommit: max: 0.170, avg: 0.168"]
Max non-epoch-change gap was: 1 rounds at version 11694 (avg 0.00) [limit 4], 1.21s no progress at version 4334975 (avg 0.06s) [limit 15].
Max epoch-change gap was: 0 rounds at version 0 (avg 0.00) [limit 4], 0.00s no progress at version 0 (avg 0.00s) [limit 16].
Test Ok

• Grafana dashboard
• Humio Logs
• Axiom Logs
• Validator CPU Profile
• Fullnode CPU Profile
• Test runner output
• Test run is land-blocking

[danielxiangzl]

Closing in favor of the consolidated weighted-leader approach in #19343 (canonical) + #19341 (forge experiment). The fixed LatencyWeightedHeuristic supersedes both strict-only and strict+weighted variants — slow/failing validators are now penalized continuously through latency scaling rather than via the asymmetric failure-window classification. See discussion in #19343.