prune

Custom implementations must be DeltaChannel-aware. DeltaChannel stores only a sentinel in channel_values for non-snapshot steps; reconstruction walks the parent chain via get_delta_channel_history, accumulating rows from checkpoint_writes until it reaches an ancestor whose channel_values contains a _DeltaSnapshot blob (written every snapshot_frequency updates).

A naive "keep_latest" that drops intermediate checkpoints and their writes can sever that chain: the surviving "latest" checkpoint is rarely a snapshot point itself, so its delta channels would silently reconstruct as empty (no error raised — get_delta_channel_history simply returns no seed). Safe options when the graph uses DeltaChannel:

• Walk back from each kept checkpoint and preserve every ancestor (plus its checkpoint_writes) up to the nearest one whose channel_values already contains a _DeltaSnapshot for every DeltaChannel-backed key.
• Force a fresh snapshot on the kept checkpoint before deleting ancestors — rewrite channel_values[k] = _DeltaSnapshot(value) for each delta channel k (resolving value via the existing ancestor walk first), then prune.
• Skip pruning threads whose graph uses DeltaChannel until one of the above is implemented.