[16.0][FIX] server_action_mass_edit: stop poisoning the registry on onchange

[DarioLodeiros]

Symptom

In production we have been seeing random KeyError: None traces every few minutes, fired from unrelated views (bank_rec_widget, knowledge.article, ...) and surfaced to the web client as RPC_ERROR:

File ".../odoo/models.py", line 6347, in _has_onchange
    for dep in self.pool.get_dependent_fields(field.base_field)
File ".../odoo/modules/registry.py", line 351, in get_dependent_fields
    if field not in self._field_triggers:
File ".../odoo/modules/registry.py", line 439, in _field_triggers
    dependencies = list(field.resolve_depends(self))
File ".../odoo/fields.py", line 775, in resolve_depends
    Model0 = registry[self.model_name]
File ".../odoo/modules/registry.py", line 186, in __getitem__
    return self.models[model_name]
KeyError: None

The trace never points at this module on the surface, but the only thing that has been planting Field instances with model_name = None into the registry is MassEditingWizard.onchange.

Root cause

MassEditingWizard.onchange injects dynamic Selection fields by assigning them straight into self._fields:

dynamic_fields["selection__" + line.field_id.name] = fields.Selection(
    [()], default="ignore"
)
self._fields.update(dynamic_fields)

These Field objects never go through __set_name__, so model_name and name stay at the default None. Any later access to registry._field_triggers — for example from _has_onchange while building a view for any other model — calls field.resolve_depends(self) → registry[None] → KeyError: None.

The for field in dynamic_fields: self._fields.pop(field) cleanup is also not wrapped in try/finally. If anything inside super().onchange() raises, the orphan Field instances are left in mass.editing.wizard._fields (a class-level dict, shared by every request on that worker). The worker stays poisoned for the rest of its lifetime: every request that touches _field_triggers then explodes randomly on completely unrelated views.

I could not find an open issue/PR covering this exact scenario in OCA/server-ux. PR #1203 deals with the view cache mechanism on 17.0+, which is unrelated.

Fix

• Call Field.__set_name__ on each dynamic field before inserting it into _fields, so model_name / name are populated and registry lookups resolve to mass.editing.wizard as expected.
• Wrap the super().onchange() call in try/finally, so the dynamic fields are always removed from _fields — even when the parent onchange raises.

Regression test

test_onchange_dynamic_fields_are_bound_and_cleaned asserts both invariants:

1. While super().onchange() is running, each dynamic Field has model_name == 'mass.editing.wizard' (without the fix the test fails with None != 'mass.editing.wizard': Dynamic field selection__bank_ids was injected with model_name=None).
2. _fields is restored to its original state on both the success path and a failure path that forces super().onchange() to raise.

Affected versions

The same pattern is present on 17.0, 18.0 and the 19.0 migration PRs. I will open companion PRs once this one is reviewed.