#1445: derive the launch set from the registry — delete the hand list and the serialization point with it

[PaulFidika]

Closes pgw#1445.

LAUNCH_SET_NAMES was a serialization point: every ModelType PR had to edit the same
hand-maintained list, so two could never be in flight concurrently without one going
stale. Measured on pgw#1432 alone — origin/master moved out from under that branch
twice, 5 commits each time, and the list went stale on both. It went stale for four
lanes in a row: audio, LTX-2, pgw#1432, and master itself.

Paul's minimum-hand-specification ruling, and the same shape the master-green lane applied
to test_cli_entry_orders: a list is what went stale; a second list goes stale the same
way.

It also failed in the worst available place. The pinned list lives in the full tests
job, which is red for unrelated reasons — so a lane that added a ModelType, saw tests
red, and diffed the count against the known baseline concluded "not mine" and
enqueued. Only diffing failure names caught it, and nothing forced anyone to.

What is asserted is NOT circular — that is the whole design

export_document() builds its names as list(defaults_vocabularies())
(defaults_export.py:58). So checking the export against that mapping would test the
emitter against its own input and prove nothing. MODEL_TYPES and LORA_OVERLAYS are a
different source — the declaration tuples — so the export is checked against those.

What is deliberately NOT asserted: the exact sequence within the base block. That
order is defaults_vocabularies() insertion order and no independent source defines it —
the mapping's order genuinely differs from the tuple's today (the tuple runs
… Flux2Klein, Krea2, Anima, Ernie, QwenImage, ZImage …, the emitter runs
… flux2-klein, qwen-image, z-image, krea-2, anima, ernie …). Pinning it would recreate
the hand list this deletes.

The ordering property the docstring actually promises — base types first, then the
overlays — is checked, because that one is structural rather than incidental. Saying
which of the two is checkable is the honest half of the change.

Red/green verified on all four arms

By mutating the registry the test reads, not by inspection. Each names the specific
problem rather than printing a list diff:

arm	message
GREEN	passed
missing-from-export	['internvl-u'] are registered in MODEL_TYPES/LORA_OVERLAYS but the export document does not name them — they would ship with NO SCHEMA
extra-in-export	['ghost-family'] are named by the export document but are in neither MODEL_TYPES nor LORA_OVERLAYS
overlay-before-base	the export interleaves LoRA overlays with base types
empty-registry	the model-type registry collapsed to []

The vacuity arm is the one that matters most, and it is why the non-emptiness checks
come FIRST. Set equality against an empty registry is trivially true, so an import that
silently produced no types would have passed every other assertion in this test. It now
fails loudly on anchors that have been in the launch set since it existed.

Effect

Net -24 lines, and a new ModelType now has two registration sites instead of
three — the tuple and defaults_vocabularies(), both already fenced by pgw#1001. The 3D
lane's #996 and every future model-types PR stop conflicting here.

Verification

• 80 passed across tests/test_model_defaults_export.py + tests/test_model_defaults.py
• uv run mypy tests/test_model_defaults_export.py → clean; ruff → clean
• The test_eager_bridge_dtype_pgw1447.py mypy errors in a full-tree run are
  pre-existing: present at pristine origin/master with this change removed, and this
  diff touches exactly one file.