`accumulate_from_start_of_forecast` does not emit step 0 in its output

[dnerini]

The Accumulate post-processor (accumulate_from_start_of_forecast) correctly accumulates fields from zero, but never emits a step 0 field. The accumulator is initialised to zeros implicitly on the first forecast step, so the initial condition (accumulation = 0 at t=0) is never included in the output state.

As a consequence, downstream consumers that need to de-accumulate the output (i.e. convert cumulative-from-start to per-step accumulations via differencing) require step 0 as a reference point. Without it they must either:

• synthesise a zero-valued step 0 themselves (requiring knowledge of this implementation detail), or
• silently drop the first accumulation window.

The expected behaviour should be thta when accumulate_from_start_of_forecast is active, a step 0 field with value 0 everywhere should be present in the output, consistent with the physical meaning that no accumulation has occurred at the initial condition.

One possible solution could be to emit a zero-valued field at step=0 for each variable in self.accumulations before the first process() call.

[floriankrb]

A third solution would be to implement code in anemoi-inference that would not accumulate from start, but that would provide accumulated-from-previous-step data, since this is what your consumer need.

I am not a big fan of creating a fake field for step 0 which would be accumulated from zero, because the period of accumulation of such field would be the interval [0,0], which could be considered as wrong metadata.