refactor: the registry builds the pipelines; delete the four if ladders

[Hendrik-code]

Stacked on #69. The largest PR in the stack, and the one that breaks configs.

The four if ladders go

smauglab/transforms/build.py does the dispatch once, for all three GPU pipeline modes and the CPU one:

file	before	after
gpu/transforms.py	714	31
gpu/transforms_list.py (two ladders)	974	117
cpu/transforms.py	412	80

Which class a config key maps to, which parameters it accepts, and where it sits in the pipeline all come from the registry now, so the four copies cannot drift again. The temporary aliases #66 left in contrast.py go with them.

Configs move to the new format in the same commit

The ladders and the old key names are the same thing, so they can't be separated. All 29 shipped configs were rewritten by migration/migrate.py, not by hand, and the originals are kept verbatim in unit_tests/fixtures/legacy_configs/ so the migrator has something to be tested against.

$ smauglab migrate --in-place smauglab/configs/*.json
smauglab/configs/transform_params.json -> smauglab/configs/transform_params.json
    pushed top-level retain_stats=True into the blocks that read it
    ConvTransform(kernel_type='Laplace') -> LaplaceConvTransform
    expanded FunctionTransform into 5 blocks
    moved mirror_axes -> CPU.MirrorTransform.allowed_axes

⚠️ This breaks every config file outside the repo. migration/ is how you bring one forward, and the error messages point at it.

How I know it changed nothing

Not asserted — checked. unit_tests/fixtures/legacy_effective_kwargs.json records the class and effective constructor kwargs of every transform the ladders built, for every config and both GPU modes. It was captured from the commit before they were deleted, by instrumenting each __init__. unit_tests/test_builder.py rebuilds each of those pipelines through the registry and compares.

All 27 match.

Three differences are normalised away, and the test file says why:

• the ladders built one class parameterised by kernel_type/func/invert_image; the builder builds the leaf;
• that shared class therefore bound kernel-specific parameters its leaves don't declare — a Scharr transform carried sigma, which only the blur reads;
• p_batch is new and defaults to kornia's own 1.0 either way.

The 21 pipelines with no "before"

Of 48 config×mode combinations, only 27 could be compared. The other 21 crashed in the old ladder:

choose_x_params = self.transform_params.get("RandomChooseXTransforms")
... choose_x_params.get("probability", 1.0)      # AttributeError when the block is absent

That's 21 of 29 configs — AugTransformsGPURandomOrder was unusable for all of them. They build now, because config.pipeline_options() always returns a dict. That's a fix, not a regression, but it does mean those configs have no baseline to check against.

Also

• pipeline.order is wired through build_transforms: registry order by default, config key order when a config asks for it. Tested both ways.
• .gitignore's repo-wide *.json would have swallowed the fixtures and the migration table, so each gets an explicit un-ignore.

Testing

unit_tests/test_builder.py (9 tests / 27 subtests) plus the existing suite: 172 passed, 5 skipped, 503 subtests. 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.