Dead code sweep: write-only state, identical-branch logic, production-dead loaders #317

From /simplify codebase sweep (2026-06-10). All confirmed unreferenced in src/ (tests excepted where noted).

• transparency/explainers/base_explainer.py:85-86,133: AttributionOnlyExplainer.attributions is write-only; nothing reads it (all consumers use ExplanationResult.attributions). Pins the full attribution tensor on the explainer instance after the run -- the module otherwise fights exactly this with gc.collect()/empty_cache(). Delete the attribute and the now-empty __init__.
• models/backend.py:384-389: OnnxBackend._select_primary_output (torch-tensor variant) has zero callers; the call path went numpy-only (forward_numpy -> _select_primary_numpy_output). Delete.
• reporting/view_model.py:455-463 + :25-32: the elif key.startswith((...)) or key in _TECHNICAL_KEYS: branch and the else have identical bodies, so the condition and the _TECHNICAL_KEYS frozenset (no other users) are dead logic implying a classification that never happens. Collapse to else; delete the frozenset.
• data/data.py:873-881: _load_tabular_dir/_load_tabular_dir_numpy are only called from data/tests/test_data.py; production goes through _concat_tabular_numpy directly (pre-feat(data)!: path-based labels.csv resolution + recursive image walk #103 loader-refactor vestige, 8bc55e2). Delete; repoint tests at _concat_tabular_numpy/_list_tabular_recursive.
• tracking/mlflow_tracker.py:314: _backend_store_uri_configured is set in __init__ and read only by tests (its two siblings are used in production). Drop the flag; tests re-derive from config.