3 epochs (360-instance mix takes 3x longer per epoch)

3-regime heterogeneous training mix (tight/medium/loose due dates)

Add ATC as 6th arm + 5 epochs

Synthetic objectives + 5 arms (WSPT/Slack/FIFO/EDD/CR), weighted-tardiness reward

Bandits: switch arms to [WSPT, Slack, FIFO] for real differentiation

Skip comment/blank lines in Lawrence parser

Merge initial HF README

Trainer Space: download -> train -> push -> sleep

initial commit