Smoke job repo: grouped tabular dataset + config

400 rows / 200 patients, two samples each, so grouped splitting is actually
exercised. Signal is modest and noisy on purpose — a separable toy would score
1.0 with a broken model and prove nothing.

configs/tabular_smoke.json

@@ -0,0 +1,52 @@
+{
+  "_notes": [
+    "Smoke test for the repo-backed job flow: tabular-only binary classification.",
+    "metadata_csv is relative to the repo root, which is the provider's working",
+    "directory when a job carries repo+ref. Deliberately small and short — this",
+    "exists to prove the path works end to end, not to produce a good model."
+  ],
+  "run_name": "tabular_smoke",
+  "num_classes": 2,
+  "split_identity_level": 1,
+  "eval_stage": "cd_fuse",
+  "save_predictions": true,
+  "seed": 1234,
+  "folds": 3,
+  "fold_seed": 100,
+  "output_root": "results",
+
+  "data": {
+    "module": "hypertower_core.profiles.generic",
+    "args": {
+      "metadata_csv": "data/smoke/labels.csv",
+      "id_column": "sample_id",
+      "label_col": "diagnosis",
+      "target_type": "classification",
+      "group_column": "patient_id",
+      "cat_cols": ["site"],
+      "exclude_cols": ["notes"]
+    }
+  },
+
+  "towers": [
+    {
+      "name": "cd",
+      "module": "hypertower_core.components.towers.clinical_tower",
+      "class": "ClinicalEncoder",
+      "data_source": "matrix",
+      "args": { "hidden_dim": 64 }
+    }
+  ],
+
+  "stages": [
+    { "name": "cd_warm", "type": "warm", "tower": "cd", "head_name": "cd_aux",
+      "level": "sample", "epochs": 8 },
+    { "name": "cd_aux",  "type": "head", "input": "cd", "train_with": "cd_fuse" },
+    { "name": "cd_fuse", "type": "fusion",
+      "module": "hypertower_core.components.bridges.mono_bridge",
+      "class": "MonoBridge", "inputs": ["cd"], "level": "sample",
+      "epochs": 10, "train_towers": true, "args": { "use_ln": false } }
+  ],
+
+  "training": { "lr": 1e-3, "batch_size": 32, "tune_binary_threshold": true }
+}