fix: radar_classification's run() dispatched to a dead debug function, not main()

[musicalplatypus]

This is part of a 4-PR stack: #34 (this PR) → #35 → #36 → #37. Recommended merge order is numeric; each later PR's diff currently includes the earlier ones' commits until they merge, since none have landed yet.

Summary

radar_classification/train.py's run() called main_debug() — a leftover notebook-parity harness (its own comment: "Following as close as possible steps from jupyter notebook to test if model learning plateau is coming from training loop") — instead of main(), which is fully wired to this codebase's shared training infrastructure.

Concretely, main_debug():

• Never calls quantization_wrapped_model — the "QuantTrain" pass produces a second, mislabeled float-training run instead of an actual quantized model.
• Never calls resume_from_checkpoint — checkpoint resume is silently a no-op.
• Hardcodes a fixed random seed and a plain SGD optimizer with no LR scheduler, unlike main()'s configurable setup shared with every other reference script.

This was a genuine regression risk: main() had zero execution history for radar before this fix. Verified end-to-end on a synthetic fixture with both --quantization 0 and --quantization 1 — the quantized pass now correctly shows QuantTrain phase log entries and exports a second (QDQ) model, which it silently never did before.

Note: at this PR's state, main() does not yet support --compile-model either (neither main() nor main_debug() called compile_model_if_enabled) — that's added in #35, on top of this fix making main() reachable at all.

Changes

• run() now dispatches to main, matching every other reference script's pattern
• New regression test pinning the dispatch target so this can't silently regress back to main_debug
• --compile-model support for radar is added in the follow-up PR (feat: wire compile_model_if_enabled into radar, image, and audio classification #35), which this PR's fix is a prerequisite for (compile wiring only exists in main(), which was previously unreachable)

CI

CI is red on all three OS jobs for this PR — but it's also red on main at this PR's base commit (6f2c11cc), from test_config_validation.py's radar_point_cloud_classification example config (added by PR #13) not being recognized by the task-type/model registries. This PR doesn't touch that config or those registries and doesn't change the failure set.

Test plan

• New regression test: tests/test_radar_entrypoint_uses_main.py
• Manual E2E on synthetic fixture, both quantization=0 and quantization=1, both completing cleanly with the expected artifacts
• Full existing test suite passes locally (2 pre-existing, unrelated CI failures — see CI section above)