refactor(architectures): drive latent-space facts from the registry

[Pfannkuchensack]

Stacked on #27. Base is refactor/arch-facet-registry; review that first. The diff shown here
is against it, not against main.

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

Why

The preview chain was never really keyed on architecture — the branch comments in the dispatch chain
this removes said so by hand: "Krea-2 decodes with the Qwen-Image VAE", "Z-Image uses
FLUX-compatible VAE", "Anima uses Wan 2.1 VAE". Fifteen architectures share eight latent
spaces, and the five most recent introduced no new preview data at all.

That one fact lived in six partial, drifting copies. This replaces four of them.

Reading the diff

Every projection matrix moves byte-identically, comments included — use
git diff --color-moved. step_callback.py goes from 396 lines to 60.

The two Wan dispatches collapse into one

step_callback.py picked the factors by sample.shape[-3] == 48, then forty lines later picked the
spatial scale by the same test. Two halves of one fact, in two places, free to drift. Resolving one
LatentSpace now settles both, because it carries spatial_compression.

LatentSpaceFacet.resolve() matches on channel count — literally what the old comment claimed it was
doing ("the latent channel count uniquely identifies the variant"). An architecture with a single
latent space short-circuits without touching the sample at all, so an unusually shaped tensor
cannot start raising IndexError for the other fourteen.

Deliberately not changed

• Ideogram 4 keeps its own preview loop. Its callback signature is (step, total, packed_latents)
  and it must unpatchify and denormalize first, so it cannot route through diffusion_step_callback
  without a fabricated PipelineIntermediateState. It now reads the same declaration instead of
  inlining the FLUX.2 factors and a hardcoded × 8. Its try/except fallback stays, and so does its
  missing is_canceled() check — that is a real bug (Ideogram generations do not stop promptly on
  cancel), but fixing it would change behaviour. Follow-up.
• invocation_context.py is untouched. The spec proposed collapsing flux_step_callback /
  flux2_step_callback into sd_step_callback, but they are already thin wrappers over the same
  function — the change would be pure docs churn in docs/src/generated/invocation-context.json.
• constants.LATENT_SCALE_FACTOR (blast radius: every latent node) and pid/decode.py (same
  shape, different units — packed latents at 128ch/16×, not VAE latents at 32ch/8×). Both follow-ups
  now that a correct source exists.
• denoise_latents.py keeps a third, divergent copy of the SDXL BGR rule. It reads the UNet's
  base where the other paths read each adapter's; the two disagree for an SD1 adapter on an SDXL
  UNet. Flagged with a NOTE rather than fixed — the natural home is a bgr_input field on
  UNetDownscaleFacet, together with untangling the two parallel T2I-Adapter paths.

max_unet_downscale

Duplicated verbatim across denoise_latents and T2IAdapterExt, comment and error string included.
It is a UNet property, not VAE geometry, so it gets its own optional facet rather than a field on
LatentSpaceFacet. The accessor — not require() — carries the error, reproducing the old message
exactly, including that BaseModelType is a str, Enum mixin and so renders as BaseModelType.Flux
rather than flux.

The three identical Wan 2.1 matrices

QWEN_IMAGE_, ANIMA_ and WAN_LATENT_RGB_FACTORS were byte-identical (verified programmatically),
as were their three biases. Now one WAN21_16, asserted by object identity so the duplication
cannot creep back.

Two things worth a reviewer's attention

The relocated projection tests now hardcode their expected pixel. The versions they replace
computed the expectation by summing the very matrix under test — a tautology that would have passed
against a corrupted matrix. It had already drifted: that test's docstring claimed column sums of
0.3677 / 0.4577 / 0.9101 where the real ones are 0.3887 / 0.8771 / 1.3152.

The registry fixture now empties Facet.FACET_TYPES too. The first REQUIRED facet exposed that
isolating only one of the two globals let validate() hold the dummy architectures to real facets.

Tests

• test_latent_space.py — a reference table read off the old elif chain, one row per architecture,
  plus first/last rows and column sums as exact fingerprints (IEEE-754 addition in fixed order, so
  platform-independent; the rows catch a reordering a sum is blind to). Plus the closed-set check and
  the relocated projection tests.
• test_step_callback.py — diffusion_step_callback end to end for all 15 bases and both Wan cases.
  The first automated coverage the spatial-scale path has ever had.
• test_unet_downscale.py — both values, and the verbatim error message for the other 13.

Verification

• pytest tests/backend/architectures tests/app/util tests/test_imports.py — 218 passed
• mypy invokeai/backend/architectures (strict) — clean
• ruff@0.11.2 check . + format --check — clean
• openapi.json regenerated and compared normalized — identical. No Pydantic class, invocation
  field or enum is touched, so schema.ts and invocation-context.json are unchanged too.

Not done here, and worth doing before merge: a manual generation on SD1, FLUX and Wan TI2V-5B to
see previews render at the right dimensions on hardware. The reference table covers the data; it
cannot cover the screen.

🤖 Generated with Claude Code