Built-in observation aggregators drop bounds-preserving specifications #560

Reproduction

Built-in observation aggregators declare a bounds_reserving attribute, while Updater.observation_spec() reads preserves_bounds. Consequently, even min, max, mean and median are treated as having unknown bound behavior: the updater logs a warning and returns an unbounded Array instead of retaining a BoundedArray specification.

The problem is visible directly:

from dm_control.composer.observation.observable import base
assert base.AGGREGATORS['mean'].preserves_bounds is True
# AttributeError: functools.partial has no attribute preserves_bounds

A bounded Generic observable reduced with the built-in mean loses its limits through the actual Updater. Returned values remain in bounds, but the resulting specification no longer rejects out-of-bounds inputs.

Expected behavior

Use the attribute name already documented and consumed by Updater. The four bounds-preserving reducers should retain scalar or per-element limits, while sum must remain unbounded. Preserve reduced values, shape/dtype inference, custom-aggregator behavior and unbounded inputs.

Reproduced on main at a04e3e4cf56c12117d2294bb090f9acec21e5c67, including a native MuJoCo observation sequence. No renderer, external services or model inference is needed.