feat(data)!: unify data.source and data.labels.source resolvers

[stanlrt]

Closes #127.

Summary

data.source and data.labels.source previously used different resolvers. Sample names like isic2018 worked for data.source (via _resolve_sample) but failed for data.labels.source (which only handled URL/local-path through get_source_path). This PR unifies the resolution path so both fields accept URL, local path, or sample name.

Changes

• samples.py: introduce SAMPLE_LABELS registry; _resolve_sample now writes a labels.csv into the cache dir for samples that ship ground-truth labels (imagenet_samples only, with real ImageNet class indices). New resolve_sample_labels_path() helper.
• data.py: get_source_path(source, kind="data" | "labels") falls through to sample lookup when the source matches a SAMPLE_SOURCES key. kind="data" returns the cache directory; kind="labels" returns the bundled labels.csv (raises if the sample has no labels). _load_labels calls it with kind="labels".
• configs/data/imagenet_samples.yaml: ship default labels block pointing at the sample-bundled CSV.
• configs/config.yaml: switch default data from isic2018 to imagenet_samples. Default raitap run no longer fires the "No ground-truth labels" warnings on metrics + robustness because the model (ViT-B/32, ImageNet-pretrained) now has matching ground-truth labels.
• docs/modules/data/configuration.md: clarify both fields accept sample names.
• Tests: cover the three resolver branches and the no-labels-for-this-sample error.

Notes

• Schema unchanged. data.source and data.labels.source remain separate fields with distinct meanings — source resolves to input data, labels.source to a labels file. The unification is at the resolver layer only.
• Built on top of Create the error/warning tracking logic #123 follow-up work; PR targets 123-create-the-errorwarning-tracking-logic rather than main.

[Copilot]

Pull request overview

Warning

The PR changes the default config (configs/config.yaml) from data: isic2018 to data: imagenet_samples, which changes out-of-the-box behavior. If this is intended as a breaking change, the PR title should include ! (e.g. feat(data)!: or feat(data/config)!:) or the description should explicitly call out BREAKING CHANGE.

This PR unifies resolution for data.source and data.labels.source so both accept URLs, local paths, and named demo samples, and adds sample-bundled ground-truth labels for imagenet_samples via a cached labels.csv.

Changes:

• Add a per-sample labels registry and materialize labels.csv into the demo-sample cache for samples that have trustworthy ground-truth labels.
• Extend get_source_path(..., kind="data"|"labels") to resolve named demo samples for both data and labels (with an error for samples that don’t ship labels).
• Update default configs/docs and add tests for the new resolver branches.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/raitap/data/samples.py	Adds SAMPLE_LABELS and writes labels.csv into the sample cache; exposes resolve_sample_labels_path().
src/raitap/data/data.py	Routes data.labels.source through get_source_path(kind="labels") and extends get_source_path to support sample names for labels.
src/raitap/data/tests/test_data.py	Adds tests covering sample-name resolution for data/labels and the no-labels error path.
src/raitap/configs/data/imagenet_samples.yaml	Configures a default labels block pointing to the sample-bundled CSV.
src/raitap/configs/config.yaml	Switches the default dataset from isic2018 to imagenet_samples.
docs/modules/data/configuration.md	Documents that both source and labels.source accept URLs/paths/sample names and explains caching behavior.

[Copilot]

Pull request overview

Copilot reviewed 6 out of 6 changed files in this pull request and generated 3 comments.