generate_mock_register.py invents rows for a schema the app already seeds, so demo data duplicates the catalogue

[rubenvdlinde]

hydra-gates/scripts/lib/generate_mock_register.py generates a mock object for
every schema it finds. It has no notion of "this schema is a catalogue the app
seeds itself", so for any app that ships seed rows in register.d and ships
demo data, installing demo data duplicates the catalogue.

Measured on dossiq

register.d/96-integrations.json seeds twelve dossiqIntegration rows with
stable slugs. The generator emitted three more:

key	seeded slug	generated slug
zgw	integration-zgw	dossiqintegration-dossiqintegration-1-1
stuf	integration-stuf	dossiqintegration-dossiqintegration-2-2
brp	integration-brp	dossiqintegration-dossiqintegration-3-3

ImportHandler upserts a seed object on (register, schema, slug), so the two
spellings never collapse. An instance with demo data installed shows two BRP
rows and two ZGW rows.

Why it costs more than a duplicate row

It presents as a product defect somewhere else entirely. On dossiq it broke two
E2E tests, and neither pointed here:

• a Playwright strict-mode violation, two rows matching /BRP/i
• status: error for StUF, from a generated row asserting Circuit open on Gemeente Zuid — a connection failure nobody measured, invented by the
  generator to fill a status field

That second one is the part worth fixing upstream rather than per app. A
generator that fills an enum by picking a value will eventually pick the one that
means "something is broken", and a demo instance then reports an outage that does
not exist.

Suggested shape

An opt-out the app can declare, so the exclusion lives next to the seed that
causes it. Either:

• a marker on the schema (x-hydra-mock: false, or reuse an existing annotation), or
• a generator input listing schemas to skip, defaulted from the schemas that
  register.d already seeds objects for, which the generator can see.

The second needs no per-app change at all: if a schema has seed objects in
register.d, the app has said what its rows are, and inventing more is never
right.

Worked around for now

dossiq#2304 removes the three rows and adds a mutation-checked guard test, so a
regeneration that puts them back fails loudly. That guard is per app, which is
why this is filed here.