refactor(starter-models): split the 2428-line catalog into one module per architecture

[Pfannkuchensack]

Stacked on #30 → #29 → #28 → #27. Base is refactor/arch-conditioning-facet.

Fifth of the series from .ideas/Backend Modularization Plan.md. No behaviour changes.

185 definitions, the curated display order and twelve bundles lived in one 2428-line file, so every
new architecture appended entries to several regions of it. The definitions now live one module per
architecture.

Every definition moves verbatim, comments included — read it with git diff --color-moved. The
only text dropped is the # region markers, which grouped by model type inside the single file and
have no meaning once the files are grouped by architecture.

Two of the plan's premises did not survive contact with the data

There are no import cycles to break. Of 97 dependency edges, exactly five cross
architectures — krea-2 → qwen_image_vae, z-image → flux_vae (×2), sdxl-refiner → sdxl_fp16_vae_fix — and they form a DAG. Shared components go in common, which imports nothing
architecture-specific, and that is the whole of it. §3.7 proposed stable string ids specifically to
break cycles that do not exist.

Those ids would also have been an API change. StarterModel, StarterModelWithoutDependencies
and StarterModelBundle are served by GET /starter_models and appear in openapi.json, so turning
dependencies into a list of ids would have changed the schema and broken both frontends.

It stays a list of model references — which also keeps dangling dependencies structurally
impossible: an unknown name is an ImportError, not a runtime surprise. That is exactly the property
the ids were meant to reconstruct with "fail on unresolved id".

STARTER_MODELS keeps its curated order, explicitly

It follows neither base, nor model type, nor definition order: 179 entries in 47 contiguous runs by
base, with sdxl and flux each appearing in eight separate runs. The file it lives in says the
frontend does not re-sort it. It is product data.

An aggregator that concatenated the per-architecture modules would therefore have silently
reordered the model manager. So __init__ still holds that list — 179 names rather than 2428 lines
of definitions — and the risk this introduces, an entry defined but never listed, is what
test_every_entry_is_reachable covers.

This is the honest limitation of this PR: a new architecture still adds its entries to the curated
order. What it no longer does is add 200 lines of definitions to a shared file.

No consumer changed

__init__ re-exports every entry, so every existing import keeps working — including the ones that
reach for individual models: siglip in flux_redux, three image encoders in ip_adapter, one
external model in a test. The pre-existing test_starter_models.py passes untouched.

Bundle keys are preserved exactly, including that Wan has no base-keyed bundle at all — it ships
wan_t2v and wan_i2v as plain strings next to ten keyed by BaseModelType. Those keys are
serialised into the API response.

Proof of equality

The catalog was fingerprinted before the split — all 179 entries with every field, the resolved
dependency graph by source, the bundles and their key types — and compared after:

IDENTISCH: 179 Modelle, Reihenfolge, aufgeloester Abhaengigkeitsgraph und Bundles unveraendert

A set comparison on source alone would have passed while every dependency list silently emptied,
which is why the graph is part of the fingerprint.

Two findings the split surfaced

Neither caused by it, neither fixed here:

• esrgan_srx4 and flux_kontext are dead — defined, but listed nowhere, in no bundle, depended
  on by nothing, imported by nobody. flux_kontext in particular looks like a model someone meant to
  offer. Removing them is a content change, not a mechanical move; flagged for a follow-up decision.
• Four entries exist only as dependencies — nobody picks a CLIP image encoder from the model
  manager, it arrives with the IP-Adapter that needs it.

Both groups are named in the tests, so they stay decisions rather than oversights.

Verification

• pytest tests/backend/model_manager tests/backend/architectures tests/test_imports.py — 836 passed
• mypy --follow-imports=silent .../starter_models — clean across all 18 modules
• ruff@0.11.2 check . + format --check — clean
• openapi.json regenerated and compared normalized — identical.

🤖 Generated with Claude Code