Warm-start A/B — does a 32² checkpoint transfer down to 16²?

What it tests: whether initialising a 16×16 / 4-agent run from a checkpoint trained on the larger 32×32 / 10-agent world (warm-started) reaches competence faster than training the same architecture from random weights (from scratch). Four matched runs — two scratch, two warm — same LPAC backbone, same explorer/relay roles, same reward; the only difference is the starting weights.