Clarify, make explicit, and add conformance test if CommandLineTool requires a baseCommand vs being a no-op

[whlavina]

Per the specification for CommandLineTool and its baseCommand:

If baseCommand is not provided or is an empty array, the first element of the command line produced after processing inputBinding or arguments must be used as the program to execute.

The wording and reference to the "first element" seems to imply there must be a command to execute, but this is not stated explicitly. There are use cases for a "no-op" command, such as when combined with InitialWorkDirRequirement or otherwise passing inputs to different outputs as a passthrough node in a graph.

In particular, it seems that cwltool allows a CommandLineTool without any baseCommand under some conditions (e.g. if there's a Docker image associated), while other runners like Rabix fail on such cases.

I recommend making the specification more explicit, and adding conformance tests.

As example, consider this naive conformance test, which "passes" in cwltool but fails in other runners:

#!/usr/bin/env cwl-runner

cwlVersion: v1.0
class: CommandLineTool
requirements:
    - class: DockerRequirement
      dockerPull: ubuntu
inputs: []
outputs: []

[tetron]

So what is happening is that it is running the CMD and/or ENTRYPOINT of the Docker image. This isn't documented explicitly and there's discussion that for consistency it shouldn't be allowed at all (#522). The same tool spec with no DockerRequirement would probably fail everywhere, including cwltool.

I agree that the specification should be explicit and there should be conformance tests.

No-op pass-through nodes that perform data reshaping are better implemented with an ExpressionTool.