dataset push: --category help lists 5 of 9 categories, and unknown-vs-empty category errors give two different lists (one leaks unsupported segmentation)

[saadqbal]

Severity: LOW (copy/consistency). Found during the #67 stress sweep.

(a) --category inline flag help is stale

tracebloc dataset push --help shows:

--category string task category: image_classification, tabular_classification, tabular_regression, time_series_forecasting, time_to_event_prediction (default "image_classification")

Only 5 categories listed, but the CLI supports 9 (object_detection, keypoint_detection, text_classification, masked_language_modeling are missing from this string). The long description and error messages correctly say 9.

(b) Two different category lists depending on how it's wrong

--category banana   # → "Supported: image_classification, object_detection, keypoint_detection, text_classification,
                    #     masked_language_modeling, tabular_classification, tabular_regression,
                    #     time_series_forecasting, time_to_event_prediction."   (9, sensible order)
--category ""       # → "value must be one of 'image_classification', 'object_detection', 'keypoint_detection',
                    #     'semantic_segmentation', 'instance_segmentation', 'text_classification', ...'"   (11!)

The empty-string path falls through to raw JSON-schema enum validation, which lists 11 values — including semantic_segmentation and instance_segmentation, which the CLI explicitly rejects elsewhere ("isn't supported by the CLI yet"). Different count, different order, and it advertises categories that don't work.

Expected

One canonical supported-category list, used in the flag help, the long help, and both error paths. Don't surface segmentation categories the CLI rejects.

Part of #67.