Run output size can exceed 50 GB – investigate and reduce disk usage

[floo-dck]

Problem

A single AutoRecLab run can produce output directories of ~50 GB or more.
This is far too large for most machines and makes it impractical to store or
compare multiple experiments.

Expected behaviour

A single run should only consume storage proportional to the actual model
outputs (generated code, results, statistics). Storage should not silently
balloon due to implementation artefacts.

Likely causes to investigate

• Checkpoints: The checkpoint/ subfolder saves the full workspace state
  for every tree node. If the agent downloads or generates large datasets, every
  node checkpoint duplicates that data.
• Workspace accumulation: The workspace/ directory accumulates all files
  produced by the generated code (downloaded datasets, trained model weights,
  intermediate files) and is never pruned between nodes.
• keep_only_relevant_files = false (current default): All intermediate
  files are retained. Switching to true already deletes some artefacts, but
  apparently not enough.
• No size cap / warning: There is currently no mechanism to alert the user
  when the output directory exceeds a configurable threshold.