A pass is one attempt inside healing a single node — depth. An episode is one full RHA cycle: sweep the whole tree, pick the upstream-most failing node, heal it (spending passes), recheck it and its dependents, record everything — breadth.
Passes bound how hard the system grinds on one node; episodes bound how many times it re-sweeps the artifact, so a cascade (heal A, which surfaces a failure in B, which surfaces C…) terminates instead of cycling forever. Tune episodes with artifact size — more nodes and deeper dependency chains legitimately need more sweeps. Tune passes with model weakness — a weaker model needs more tries per node.