feat: add the augmentation registry

[Hendrik-code]

Stacked on #66. smauglab/registry.py, with no consumers yet — reviewable on its own before anything depends on it.

It answers four questions that today are answered by reading an if ladder: what augmentations exist per backend, which class a config key maps to, what parameters that class accepts, and which backends implement a given concept.

Accepted parameters come from inspect.signature, so there is no second schema to drift out of step with the constructors. Registering a class whose __init__ ends in **kwargs without declaring forwards_to is an error — signature-based validation would otherwise accept anything.

The module is stdlib-only and imports nothing from smauglab.transforms at module scope. That keeps the import graph acyclic (transform modules import register from here) and lets a CLI answer "what exists?" without paying for torch.

On pipeline order — the design question you asked about

Each augmentation's position is data the ladders held implicitly, so the registry has to hold it explicitly. Config key order cannot decide it: the two have never agreed, and honouring the file would silently reorder every pipeline the moment someone tidied a config. (A config will be able to opt into key order — pipeline.order — when config.py lands.)

Per your call, it's one PIPELINE_ORDER tuple per backend, not an order= integer on each @register:

Backend.GPU: (
    "RandomFlipTransformGPU",              # FlipTransform
    "RandomAffineGPU",                     # AffineTransform
    "RandomSynthSegGPU",                   # SynthSeg
    ...
)

The question a reviewer needs to answer is "does the new pipeline run things in the same sequence as the old one?" Against a single list that's a diff against the ladder it came from; against integers spread over eight files it isn't. Every GPU entry carries the ladder key it was derived from as a comment, so the correspondence is checkable line by line against gpu/transforms.py.

Registering a class absent from its backend's tuple is an error, so the table cannot fall out of date.

One entry needs a judgement call: RandomLaplaceGPU has no ladder key. The GPU Laplace kernel was reachable only by setting kernel_type="Laplace" on ScharrTransform, so its position is a new choice — placed next to its sibling. Worth a look.

isolated(order=...)

isolated() without an argument turns the position check off for the block, which is what most tests want — they register throwaway classes whose pipeline position isn't under test. With an explicit table, registration is checked against it, which is how the check itself gets covered.

Also in here

smauglab stops being a PEP 420 namespace package: every directory gets an __init__.py. The registry needs a deterministic import-time population point, and a namespace package lets a stray smauglab/ elsewhere on sys.path silently merge into this one. namespace_packages / explicit_package_bases come out of [tool.mypy] accordingly.

Testing

unit_tests/test_registry.py — 26 tests against synthetic transform classes, deliberately not the real ones (those get registered in the next PR; these are about the mechanism). Covers the registration invariants, signature-derived parameter validation, the "did you mean" suggestions, and the coverage matrix.

Full suite: 131 passed, 237 subtests. The 24 seeded config digests are unchanged — nothing imports the registry yet. ruff and mypy clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.