Plugins cannot register NumPy functions wrapped with Wrap1to1Func (documented example fails) #415

Summary

The documented processing plugin example plugins/examples/datalab_custom_func.py, used by the Custom function tutorial and the plugin guide, no longer loads on develop. Its registration fails with:

ValueError: 'weighted_average_denoise' is not a valid computation function.

The plugin therefore adds no menu entry. Any plugin following the documented pattern — registering a NumPy function wrapped with sigima.proc.image.Wrap1to1Func (or sigima.proc.signal.Wrap1to1Func) through register_1_to_1 — fails in the same way.

This regression only exists on develop: no released version is affected (latest: v1.3.0).

When it happens

1. Copy plugins/examples/datalab_custom_func.py to a plugin directory.
2. Start DataLab or reload the plugins.
3. Its register_computations() raises the error above, and the "My custom filters" menu does not appear.

import sigima.proc.image as sipi
from datalab.gui.processor.base import ComputingFeature

ComputingFeature(
    "1_to_1",
    function=sipi.Wrap1to1Func(weighted_average_denoise),
    feature_id="org.datalab.examples.custom-filters:weighted-average-denoise",
    owner_plugin_id="org.datalab.examples.custom-filters",
)  # ValueError: 'weighted_average_denoise' is not a valid computation function.

Root cause

ComputingFeature.__post_init__ only accepts functions decorated with Sigima's @computation_function(). Commit ccb81ef ("feat: add plugin-owned processing lifecycle") moved the example from compute_1_to_1(wrapped_func), which does not check the function, to register_1_to_1(wrapped_func, ...), which does. No test loads the shipped examples in DataLab: test_plugin_examples_import only imports them.

Yet a Wrap1to1Func instance is a valid feature function: it is picklable (as required by the process pool), named after the wrapped function, and returns the expected object. Sigima documents it as usable "as a Sigima computation function and inside DataLab's infrastructure", and DataLab already detects the plugin origin of such wrappers (_detect_plugin_origin).

Proposed fix

• Registered 1-to-1 features accept the Wrap1to1Func instances of sigima.proc.signal and sigima.proc.image, besides decorated computation functions. The example, the tutorial and the migration guide stay unchanged: wrapping a NumPy function remains the simplest way to add a processing.
• Other callables remain refused, with an error message naming both accepted forms, so that mistakes are reported at registration rather than when processing in a separate process. A wrapper registered for another pattern than 1-to-1 is refused as well.
• A new application test loads every shipped example plugin in DataLab and runs the processing of the custom function example, also after reloading the plugins.
• The plugin guide states which functions processing plugins may register.

Out of scope

• Release notes: the regression was never released.
• Simplifying the single-file plugin generated by datalab-plugin create (Create single-file plugins with datalab-plugin create #413, Create single-file plugins with datalab-plugin create #414), which will register the wrapper directly once this is fixed.