Fix the safe step finder for a parallel pipeline

An agent sped up your data pipeline twice, and twice the faster version was wrong. The second time it ran a step in parallel that depended on the order of another. Now you decide which steps may run in parallel, and the agent may only optimize those.

An assistant wrote safe_parallel_steps(steps). It is in the editor. It has a bug. Find it and fix it.

Input and output:

• steps: the pipeline's steps, each written name:reads:writes, where reads and writes are comma separated keys of shared state. For example clean:raw:clean, or audit:: for a step that touches nothing shared.
• Return the names of the steps that are safe to run in parallel, each name once, sorted as plain strings.

Rules the tests check:

• Split each step on :. Ignore a step that does not have exactly 3 parts, or whose name is empty after trimming whitespace. An ignored step conflicts with nothing.
• Trim whitespace from the name and from each key. Ignore empty keys.
• Compare each step with every other step in the list, one by one, including another step with the same name. A step is safe only when, for every other step, none of these holds:
  • it reads a key the other step writes
  • it writes a key the other step reads
  • both steps write the same key
• Two steps that only read the same key do not conflict.
• A step never conflicts with itself. When two steps share a name, the name is in the result if either step is safe.