This is something that we do need if we are going to keep using ClearML Pipelines, and we need it to be reliable and maintainable, so I don’t know whether it would be wise to cobble together a lower-level solution that has to be updated each time ClearML changes its serialisation code

Sorry if I was not clear, I do not mean for you ti do unstable low-level access, I meant that pipelines are Designed to be editable externally, they always deserialize themselves.
The only part that is missing is "export".
Do notice this will only work for pipelines from Tasks, is this a good fit for you?
BTW, what I'm thinking is:

pipe = Pipeline(...)
pipe.add_step(...)
task_id = pipe.build()
print(f"pipeline task ID = {task_id}")

wdyt?