perf: optimize replay buffer metrics and bucket traversal

[allora2026]

Summary

• replace repeated full-buffer scans with incremental counters and UTF-8 byte accounting
• sort, deduplicate, index, and safely advance time-bucket traversal
• keep buffer gauges and empty waiters correct across checkpoint failures
• harden direct and timer-driven terminal failure callbacks against rejected promises

Stack

• Depends on feat: add replay stage observability #87
• This PR contains Phase 2 only; merge feat: add replay stage observability #87 first

Verification

• 83 tests passed
• typecheck passed
• lint passed
• format check passed
• build passed
• git diff check passed
• independent merge-blocker review: APPROVED