feat: add the sectioned config schema

[Hendrik-code]

Stacked on #68. smauglab/config.py, with no consumers yet — nothing reads it until the builder lands, and the shipped configs are untouched by this PR.

A config becomes a sectioned document — GPU, CPU, pipeline, and _-prefixed comments — where a key is a class name exactly and a parameter is a constructor argument exactly, both checked against the registry.

{
  "_comment": "ignored",
  "pipeline": { "mode": "sequential", "order": "registry" },
  "GPU": { "RandomFlipTransformGPU": { "flip_axis": [0], "p": 0.5 } },
  "CPU": { ... }
}

A flat config is rejected, not migrated

It used to be read as "GPU or CPU, whichever the keys look like". The two namespaces overlap enough that GaussianBlurTransform meant a different transform depending on which builder read it. That guess is gone.

Every problem at once

The old behaviour surfaced one problem per run — for a 30-key config, 30 edit-run cycles. Now:

config.json: 3 problem(s)
  - GPU.unknown GPU augmentation 'NotATransform'. Did you mean: ...
  - GPU.RandomFlipTransformGPU: unknown parameter 'nonsense'. Accepted: flip_axis, keepdim, p, p_batch, same_on_batch
  - GPU.RandomScharrGPU: unknown parameter 'probability'.
        'probability' -> p

That last one matters: probability → p is the commonest migration mistake and difflib cannot bridge it (the two strings score ~0.17), so it comes from an explicit hint table that a test keeps diagnostic-only.

pipeline.order — the opt-in you asked for

"pipeline": { "order": "config" }     // default: "registry"

registry is registry.PIPELINE_ORDER: fixed, and the same for every config. config takes the order the keys appear in the file instead.

It is not the default because it makes a pipeline sensitive to something people reasonably treat as cosmetic — reordering or reformatting a config would silently change what it does. Opting in makes that intent explicit and greppable. It's independent of pipeline.mode.

A layering change from the original PR

PipelineMode, OrderSource and validate_section live here rather than in the builder, where #56 put them. They describe a config document, so config.py depends only on the registry and the builder will import them from here. That's also what lets this PR land on its own rather than being welded to the builder.

pipeline.mode replaces encoding the arrangement in the trainer class — which baked the choice into the name of every run directory and could not be varied without a new class.

Testing

unit_tests/test_config_schema.py — 22 tests: sections, the flat-config rejection, cumulative problem reporting, the suggestion machinery, and pipeline.order in both directions.

Full suite: 163 passed, 5 skipped, 479 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.