Benchmark: Warm-start MAB from greedy

[alltheseas]

Summary

Initialize MAB-UCB from greedy's solution instead of random, so it spends exploration budget discovering if alternatives are better in practice rather than rediscovering what greedy already knows.

Why

Greedy set-cover already produces good on-paper relay assignments (wins 23/26 profiles). MAB-UCB currently starts from scratch and needs many rounds just to converge on what greedy finds for free. Warm-starting would let MAB focus its exploration budget on testing whether greedy's choices actually deliver events — the real question.

What to do

1. Run greedy first to get an initial relay set
2. Initialize MAB arm counts using greedy's assignments as prior successes
3. Let MAB explore outward from that starting point
4. Benchmark against cold-start MAB-UCB across all 6 time windows

Effort

Medium — requires wiring greedy output into MAB initialization.

Reference

• IMPLEMENTATION-GUIDE.md: Improvement Opportunities (medium effort)