[jacoblambert]

Summary

Adds load_dataset(data_root) (t4_devkit/schema/v1/loader.py), a minimal loader for 1.0 datasets:

1. Reads manifest.json.
2. Reads the table files it lists, as raw dicts. A required table that isn't listed raises FileNotFoundError; a missing optional table is empty.
3. Builds and validates the records once, through the 1.0 registry (feat(schema): add T4Dataset 1.0 table names and registry #302).
4. Returns them as T4Tables: the manifest, the records per table, and get(table, token) lookup.

0.x datasets (no manifest.json) raise NotImplementedError for now; their conversion plugs in at step 2 in a later PR.

Shortcut fields

T4Tables fills the same convenience fields (init=False, not written to JSON) that 0.x T4Devkit fills when loading:

• sample.data (channel → key frame token), sample.ann_3ds, sample.ann_2ds
• sample_data.channel, sample_data.modality
• sensor.first_sd_token
• category_name on sample_annotation and object_ann

Differences from 0.x:

• sample.surface_anns is gone (surface_ann is merged into object_ann, feat(schema): add T4Dataset 1.0 label tables, merge surface_ann into object_ann #299).
• log.map_token is a stored field in 1.0, not filled at load time.
• Unset values are None, not "".
• first_sd_token is the earliest frame by timestamp, instead of relying on file order.

Why outside T4Devkit

A first attempt added a second load path inside T4Devkit, with "not supported for 1.0" guards in 11 helper methods. That puts scaffolding in the class users call. Instead, the load path is built and tested on its own (this PR, LiDAR and camera reading, 0.x conversion), and T4Devkit switches to it once, in a later PR.

Why raw dicts first, then records

That split is where the 0.x conversion plugs in: 0.x tables are converted to 1.0 dicts, then go through the same build and validation as 1.0 datasets.

No change to existing code; 0.x loading in T4Devkit is untouched until we finish the implementation and merge everything.

Tests

tests/schema/v1/test_loader.py:

• Every table of the 1.0 sample loads with its 1.0 class, and lookup by token works.
• Shortcut fields are filled from the links.
• A 0.x dataset raises NotImplementedError.
• A manifest without a required table (vehicle) raises FileNotFoundError; without an optional one (caption), that table is empty.

[ktro2828]

LGTM