Hands this node's in-flight episodes to recovery, in bulk, when the node shuts down.

A graceful stop (a deploy's SIGTERM) gives the VM a few seconds. Left to themselves, every running Cyclium.EpisodeTask would notice its shutdown EXIT at its next step boundary and write its own hand-off — a status update, a claim release, a journal step — a dozen episodes contending for a connection pool that is itself draining, with the ones still mid-step (usually an LLM call) getting no chance at all before the task supervisor kills them.

So the hand-off is done once, here, for all of them. This process sits in Cyclium.Supervisor after Cyclium.EpisodeSupervisor, which means it is terminated before the episode tasks are (a supervisor stops children in reverse start order). Its terminate/2 reads the running episodes off Cyclium.EpisodeProcessRegistry — every task registers itself once it holds its claim — and issues three statements for the lot:

1. UPDATE cyclium_episodes SET phase = 'interrupted' for every id still :running. The row stays running; Cyclium.Recovery.sweep/1 treats the phase as stale at once, and the expectation's recovery_policy decides.
2. One UPDATE failing this node's work claims on those episodes, so the recovering node can re-acquire them without waiting for the lease.
3. One INSERT of an :observation step per episode (%{"event" => "interrupted"}) — the marker a step feed renders as "interrupted, will resume shortly" — numbered from one MAX(step_no) query over all of them.

Then it broadcasts episode.step_journaled and episode.interrupted per episode (in-memory, no I/O) and sends each task :cyclium_drained, which the runner's loop takes at its next step boundary as "already handed off — stop, write nothing". A task killed mid-step never reaches that boundary, and does not need to: its row was already marked in step 1.

Each statement is best-effort: a failure is logged and the next one still runs, so a pool timeout on the journal insert cannot cost the phase stamp that recovery depends on.

interrupt/1 is the same hand-off for an explicit list, used by the runner for a single episode when it receives a shutdown EXIT with no drainer having run first (a host that embeds the runner without Cyclium.Supervisor).