"""Synthetic filesystem-task data pipeline for MCPMark.

Generates benchmark-style filesystem tasks (test environment + description.md +
verify.py + meta.json) with deterministic, *recomputable* verifiers. An LLM is
used only to make file content/names realistic and diverse; it never authors the
verification logic. Every generated task is validated with a built-in oracle:
the pipeline solves it programmatically and asserts ``verify.py`` exits 0.
"""