workbench 0.8.485__tar.gz → 0.8.486__tar.gz

{workbench-0.8.485 → workbench-0.8.486}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.485
+Version: 0.8.486
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License

{workbench-0.8.485 → workbench-0.8.486}/applications/aws_dashboard/Dockerfile

@@ -29,7 +29,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         -r /tmp/requirements.lock
 
 # Layer 2: Workbench only (changes often, ~20 MB)
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Copy app code and configuration (after pip installs to avoid cache invalidation)

{workbench-0.8.485 → workbench-0.8.486}/lambda_layers/build_deploy.sh

@@ -26,7 +26,7 @@ WORKBENCH_ROOT="$(cd "$SCRIPT_DIR/.." &> /dev/null && pwd)"
 AWS_ACCOUNT_ID="507740646243"
 REGION_LIST=("us-east-1" "us-west-2")
 PYTHON_VERSION="3.12"
-WORKBENCH_VERSION=0.8.484   # keep in lockstep with the image's ARG WORKBENCH_VERSION
+WORKBENCH_VERSION=0.8.485   # keep in lockstep with the image's ARG WORKBENCH_VERSION
 LAMBDA_ARCH="manylinux2014_x86_64"
 
 GREEN='\033[0;32m'

{workbench-0.8.485 → workbench-0.8.486}/sagemaker_images/base/inference/Dockerfile

@@ -33,7 +33,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Add the shared serve script

{workbench-0.8.485 → workbench-0.8.486}/sagemaker_images/base/training/Dockerfile

@@ -32,5 +32,5 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"

{workbench-0.8.485 → workbench-0.8.486}/sagemaker_images/ml_pipelines/Dockerfile

@@ -46,7 +46,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench (changes often, small layer)
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Layer 3: App code (tiny, changes rarely)

{workbench-0.8.485 → workbench-0.8.486}/sagemaker_images/pytorch_chem/inference/Dockerfile

@@ -43,7 +43,7 @@ COPY --from=builder /usr/local/lib/python3.12/site-packages /usr/local/lib/pytho
 COPY --from=builder /usr/local/bin /usr/local/bin
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"
 
 # Add the shared serve script

{workbench-0.8.485 → workbench-0.8.486}/sagemaker_images/pytorch_chem/training/Dockerfile

@@ -39,5 +39,5 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         -r /tmp/requirements.lock
 
 # Layer 2: Workbench — rebuilds per version bump, ~20 MB delta.
-ARG WORKBENCH_VERSION=0.8.484
+ARG WORKBENCH_VERSION=0.8.485
 RUN pip install --no-deps "workbench==${WORKBENCH_VERSION}"

{workbench-0.8.485 → workbench-0.8.486}/scripts/admin/inference_capture_check.py

@@ -19,7 +19,7 @@ CAPTURE_NAME = "ts_20200211"
 
 # Columns to compare
 COMPARE_COLS = ["prediction", "prediction_std", "confidence"]
-ID_COL = "udm_mol_bat_id"
+ID_COL = "compound_id"
 
 
 def main():

{workbench-0.8.485 → workbench-0.8.486}/scripts/benchmark_fp_proximity.py

@@ -57,7 +57,7 @@ def main():
 
     # Auto-detect id / target / fingerprint columns
     cols = list(df.columns)
-    id_candidates = [c for c in cols if c.lower() in ("id", "udm_mol_bat_id", "molregno", "compound_id")]
+    id_candidates = [c for c in cols if c.lower() in ("id", "molregno", "compound_id")]
     id_column = id_candidates[0] if id_candidates else cols[0]
     target_col = "logp" if "logp" in cols else next((c for c in cols if "logp" in c.lower()), None)
     if target_col is None:

{workbench-0.8.485 → workbench-0.8.486}/scripts/compare_dataframes.py

@@ -24,4 +24,4 @@ old_df = old_df[~old_df["smiles"].isin([old_error_smiles_1, old_error_smiles_2])
 new_df = new_df[~new_df["smiles"].isin([old_error_smiles_1, old_error_smiles_2])]
 
 # Compare the dataframes
-comparison = compare_dataframes(old_df, new_df, ["udm_mol_bat_id", "smiles"])
+comparison = compare_dataframes(old_df, new_df, ["compound_id", "smiles"])

{workbench-0.8.485 → workbench-0.8.486}/src/workbench/agent/guides/models/multi_task.md

@@ -23,18 +23,21 @@ validate_multi_task_data(df, target_columns=["logs", "logd"], id_column="id")
   several sources gets all its targets on one row. `merge_on_smiles=True` joins external
   data with no shared id namespace.
 - **`pull_multi_task_data`** — the same, sourced from FeatureSets rather than frames.
-  Each source takes a `target_info` (a list of columns or a `{src_col: output_col}`
-  rename map) and an optional `metadata` list. Metadata columns pass through as-named,
-  per source (NaN where that source has no row), but are not tasks. A column listed by
-  several sources collapses to one, holding the first non-NaN value per row.
+  Each source takes a `target_info`: a list of columns or a `{src_col: output_col}`
+  rename map. Only columns every source has survive as shared features, so a public
+  source without `project_code` drops it for every row. List such columns in
+  `metadata`: each is taken from whichever sources have it (NaN elsewhere), collapses
+  to the first non-NaN value per row, and logs conflicting values.
 
   ```python
   df = pull_multi_task_data(
       {
-          "logd_f1": {"target_info": {"udm_asy_res_value": "logd"}, "metadata": ["udm_asy_res_qualifier"]},
-          "pka_a1_f1": {"target_info": {"udm_asy_res_value": "pka_a1"}},
+          "logd_f1": {"target_info": {"assay_value": "logd"}},
+          "pka_a1_f1": {"target_info": {"assay_value": "pka_a1"}},
       },
-      id_column="udm_mol_bat_id",
+      {"logp_public": {"target_info": ["logp"], "src_id_col": "id"}},
+      id_column="compound_id",
+      metadata=["project_code"],
   )
   ```
 - **`validate_multi_task_data`** — null/duplicate ids, missing smiles, empty targets,

{workbench-0.8.485 → workbench-0.8.486}/src/workbench/utils/multi_task.py

@@ -195,7 +195,9 @@ def combine_multi_task_data(
     for df in dataframes:
         for col in df.columns:
             if col not in col_dtypes and not df[col].isna().all():
-                col_dtypes[col] = df[col].dtype
+                # These dtypes fill all-NA columns, so int/bool map to their nullable forms
+                dtype = df[col].dtype
+                col_dtypes[col] = {"i": "Int64", "u": "UInt64", "b": "boolean"}.get(dtype.kind, dtype)
 
     aligned_dfs = []
     for df, targets, passthrough in zip(dataframes, target_columns, passthrough_columns):
@@ -229,7 +231,12 @@ def combine_multi_task_data(
     agg_dict = {c: "mean" for c in numeric_cols}
     agg_dict.update({c: "max" for c in date_cols})
     agg_dict.update({c: "first" for c in other_cols})
-    result = result.groupby(merge_key, as_index=False).agg(agg_dict)
+    grouped = result.groupby(merge_key)
+    for c in [p for p in all_passthrough if p in other_cols]:
+        n_conflict = (grouped[c].nunique() > 1).sum()
+        if n_conflict:
+            log.warning(f"'{c}': {n_conflict} molecules have conflicting values; keeping the first")
+    result = grouped.agg(agg_dict).reset_index()
     log.info(f"Collapsing {n_before} rows -> {len(result)} ({len(dup_ids)} molecules appear in multiple sources)")
 
     # --- Step 3b: Drop rows with no supervision ---
@@ -280,6 +287,7 @@ def pull_multi_task_data(
     smiles_based_sources: dict = None,
     id_column: str = "id",
     date_col: str | None = None,
+    metadata: list[str] | None = None,
 ) -> pd.DataFrame:
     """Pull and combine multiple FeatureSets into a multi-task DataFrame.
 
@@ -291,7 +299,6 @@ def pull_multi_task_data(
     Each source config is a dict shaped like:
         {
             "target_info": {<src_col>: <output_col>, ...} | [<col>, ...]  # required; rename map or as-named list
-            "metadata":    [<col>, ...]                                  # optional; non-target columns to keep
             "src_id_col":  "<non_default_id_column>"                     # optional; FS id column, if not id_column
             "df":          <DataFrame>                                   # optional; use this frame, not the FS
         }
@@ -300,12 +307,6 @@ def pull_multi_task_data(
     per-source date column and the log line), so any source already in hand —
     a DataSource pull, a local file — can join the merge.
 
-    Metadata columns pass through as-named and merge like targets (NaN for rows
-    the source doesn't cover) but are not tasks: no target diagnostics, and they
-    don't keep an otherwise target-less row alive. A column listed by several
-    sources collapses to one, holding the first non-NaN value per row (in source
-    order).
-
     Args:
         id_based_sources: Mapping from FeatureSet name to source config.
             Primary task convention: the first entry is treated as primary by
@@ -314,7 +315,7 @@ def pull_multi_task_data(
             on canonical SMILES rather than id_column. Defaults to None / empty.
         id_column: Canonical id column name across the id-based sources.
             Defaults to "id".
-        date_col: Optional name of a per-source date column (e.g. "udm_asy_date")
+        date_col: Optional name of a per-source date column (e.g. "assay_date")
             shared across all sources. When set, each source's date is carried
             through the merge as a private per-source column and then collapsed
             post-merge as the row-wise max into the canonical `date_col`. This
@@ -330,6 +331,12 @@ def pull_multi_task_data(
             (no special date handling — `date_col` would flow through as a
             shared feature and collapse to an arbitrary "first" value across
             sources, which is unsafe for temporal splits).
+        metadata: Optional non-target columns to keep in the output, taken from
+            every source that has them. Unlike shared features, they survive a
+            source that lacks them (e.g. public data), which contributes NaN.
+            Each collapses to the first non-NaN value per row, in source order,
+            and conflicting values are logged. Raises if a column is in no
+            source. Defaults to None.
 
     Returns:
         Combined multi-task DataFrame from `combine_multi_task_data`, with all
@@ -340,6 +347,7 @@ def pull_multi_task_data(
     from workbench.api import FeatureSet
 
     smiles_based_sources = smiles_based_sources or {}
+    metadata = metadata or []
 
     def _pull_and_normalize(
         fs_name: str, fs_config: dict, date_required: bool = True
@@ -367,7 +375,7 @@ def pull_multi_task_data(
             target_cols = list(target_info)
 
         # Metadata rides the merge as passthrough columns
-        passthrough_cols: list[str] = list(fs_config.get("metadata", []))
+        passthrough_cols = [c for c in metadata if c in df.columns]
 
         # Carry per-source date through the merge as a private, source-specific
         # column. If left as a shared feature, the outer-join collapse would
@@ -426,6 +434,10 @@ def pull_multi_task_data(
         merged_targets.extend(target_cols)
         merged_passthrough.extend(p for p in passthrough_cols if p not in merged_passthrough)
 
+    missing_metadata = [c for c in metadata if c not in merged.columns]
+    if missing_metadata:
+        raise ValueError(f"Metadata columns not found in any source: {missing_metadata}")
+
     # Synthesize the canonical date as the row-wise max of all per-source
     # dates. Holding out on the latest date across any task ensures no future
     # information leaks into training through any auxiliary task.

{workbench-0.8.485 → workbench-0.8.486}/src/workbench.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.485
+Version: 0.8.486
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License

{workbench-0.8.485 → workbench-0.8.486}/src/workbench.egg-info/scm_version.json

@@ -1,7 +1,7 @@
 {
-  "tag": "0.8.485",
+  "tag": "0.8.486",
   "distance": 0,
-  "node": "g5329bcb756374720428a97d10a048fee6e63b65f",
+  "node": "g4b8704c0466eee669a576a30ca92ba9837bab1dc",
   "dirty": false,
   "branch": "HEAD",
   "node_date": "2026-09-28"

{workbench-0.8.485 → workbench-0.8.486}/tests/specific/multi_task_tests.py

@@ -8,6 +8,7 @@ temporal-date handling — in particular the id-strict / smiles-lenient rule for
 source that lacks the date column.
 """
 
+import logging
 import sys
 import types
 
@@ -180,21 +181,21 @@ def test_combine_smiles_merge():
 # validate_multi_task_data
 # ----------------------------------------------------------------------------
 def test_validate_clean_passes():
-    df = pd.DataFrame({"udm_mol_bat_id": ["1", "2", "3"], "smiles": ["CC", "CCC", "CCCC"], "t": [1.0, np.nan, 3.0]})
+    df = pd.DataFrame({"compound_id": ["1", "2", "3"], "smiles": ["CC", "CCC", "CCCC"], "t": [1.0, np.nan, 3.0]})
     # Should not raise
-    validate_multi_task_data(df, ["t"], id_column="udm_mol_bat_id")
+    validate_multi_task_data(df, ["t"], id_column="compound_id")
 
 
 def test_validate_duplicate_ids_raise():
-    df = pd.DataFrame({"udm_mol_bat_id": ["1", "1"], "smiles": ["CC", "CCC"], "t": [1.0, 2.0]})
+    df = pd.DataFrame({"compound_id": ["1", "1"], "smiles": ["CC", "CCC"], "t": [1.0, 2.0]})
     with pytest.raises(ValueError, match="duplicate"):
-        validate_multi_task_data(df, ["t"], id_column="udm_mol_bat_id")
+        validate_multi_task_data(df, ["t"], id_column="compound_id")
 
 
 def test_validate_missing_target_raises():
-    df = pd.DataFrame({"udm_mol_bat_id": ["1", "2"], "smiles": ["CC", "CCC"], "t": [1.0, 2.0]})
+    df = pd.DataFrame({"compound_id": ["1", "2"], "smiles": ["CC", "CCC"], "t": [1.0, 2.0]})
     with pytest.raises(ValueError, match="missing"):
-        validate_multi_task_data(df, ["t", "does_not_exist"], id_column="udm_mol_bat_id")
+        validate_multi_task_data(df, ["t", "does_not_exist"], id_column="compound_id")
 
 
 # ----------------------------------------------------------------------------
@@ -205,18 +206,18 @@ def test_pull_date_synthesis_row_wise_max(monkeypatch):
     frames = {
         "src_a": pd.DataFrame(
             {
-                "udm_mol_bat_id": ["1", "2"],
+                "compound_id": ["1", "2"],
                 "smiles": ["CC", "CCC"],
                 "ta": [10.0, 20.0],
-                "udm_asy_date": ["2024-01-01", "2025-12-01"],
+                "assay_date": ["2024-01-01", "2025-12-01"],
             }
         ),
         "src_b": pd.DataFrame(
             {
-                "udm_mol_bat_id": ["2", "3"],
+                "compound_id": ["2", "3"],
                 "smiles": ["CCC", "CCCC"],
                 "tb": [30.0, 40.0],
-                "udm_asy_date": ["2025-12-31", "2023-01-01"],
+                "assay_date": ["2025-12-31", "2023-01-01"],
             }
         ),
     }
@@ -224,14 +225,14 @@ def test_pull_date_synthesis_row_wise_max(monkeypatch):
 
     out = pull_multi_task_data(
         {"src_a": {"target_info": ["ta"]}, "src_b": {"target_info": ["tb"]}},
-        id_column="udm_mol_bat_id",
-        date_col="udm_asy_date",
+        id_column="compound_id",
+        date_col="assay_date",
     )
 
     # Synthesized date is a plain YYYY-MM-DD string (not datetime64) so it stores
     # verbatim via PandasToFeatures and round-trips through a tz-naive temporal split.
-    assert out["udm_asy_date"].dtype == object
-    dates = dict(zip(out["udm_mol_bat_id"], out["udm_asy_date"]))
+    assert out["assay_date"].dtype == object
+    dates = dict(zip(out["compound_id"], out["assay_date"]))
     # id 2 is in both sources -> max(2025-12-01, 2025-12-31) = 2025-12-31
     assert dates["2"] == "2025-12-31"
     assert dates["1"] == "2024-01-01"
@@ -243,16 +244,16 @@ def test_pull_date_synthesis_row_wise_max(monkeypatch):
 def test_pull_id_source_missing_date_raises(monkeypatch):
     """An ID-based (internal) source missing date_col fails loudly — likely a typo."""
     frames = {
-        "src_a": pd.DataFrame({"udm_mol_bat_id": ["1"], "smiles": ["CC"], "ta": [1.0], "udm_asy_date": ["2024-01-01"]}),
-        "src_no_date": pd.DataFrame({"udm_mol_bat_id": ["1"], "smiles": ["CC"], "tb": [2.0]}),
+        "src_a": pd.DataFrame({"compound_id": ["1"], "smiles": ["CC"], "ta": [1.0], "assay_date": ["2024-01-01"]}),
+        "src_no_date": pd.DataFrame({"compound_id": ["1"], "smiles": ["CC"], "tb": [2.0]}),
     }
     _install_fake_feature_set(monkeypatch, frames)
 
     with pytest.raises(ValueError, match="missing required date_col"):
         pull_multi_task_data(
             {"src_a": {"target_info": ["ta"]}, "src_no_date": {"target_info": ["tb"]}},
-            id_column="udm_mol_bat_id",
-            date_col="udm_asy_date",
+            id_column="compound_id",
+            date_col="assay_date",
         )
 
 
@@ -264,13 +265,13 @@ def test_pull_smiles_source_missing_date_is_training(monkeypatch):
     frames = {
         "ppb": pd.DataFrame(
             {
-                "udm_mol_bat_id": ["1", "2"],
+                "compound_id": ["1", "2"],
                 "smiles": ["CCO", "CCC"],
                 "ppb_human": [10.0, 20.0],
-                "udm_asy_date": ["2024-01-01", "2026-01-01"],  # id 2 is post-cutoff
+                "assay_date": ["2024-01-01", "2026-01-01"],  # id 2 is post-cutoff
             }
         ),
-        # public LogP: 'id' identifier, no udm_asy_date
+        # public LogP: 'id' identifier, no assay_date
         "logp_public": pd.DataFrame(
             {"id": ["a", "b", "c"], "smiles": ["CCO", "CCCC", "CCCCC"], "logp": [1.0, 2.0, 3.0]}
         ),
@@ -280,91 +281,84 @@ def test_pull_smiles_source_missing_date_is_training(monkeypatch):
     out = pull_multi_task_data(
         {"ppb": {"target_info": ["ppb_human"]}},
         {"logp_public": {"target_info": ["logp"], "src_id_col": "id"}},
-        id_column="udm_mol_bat_id",
-        date_col="udm_asy_date",
+        id_column="compound_id",
+        date_col="assay_date",
     )
 
     # Public-only molecules (CCCC, CCCCC) have no date -> NaT
     public_only = out[out["logp"].notna() & out["ppb_human"].isna()]
     assert len(public_only) == 2
-    assert public_only["udm_asy_date"].isna().all()
+    assert public_only["assay_date"].isna().all()
 
     # The CCO molecule overlaps the internal compound -> carries the internal date
     cco = out[out["smiles"].str.contains("O", na=False)].iloc[0]
     assert cco["ppb_human"] == 10.0 and cco["logp"] == 1.0
-    assert pd.to_datetime(cco["udm_asy_date"]) == pd.Timestamp("2024-01-01")
+    assert pd.to_datetime(cco["assay_date"]) == pd.Timestamp("2024-01-01")
 
     # Temporal split: only the post-cutoff internal compound is held out; the
     # date-less public rows fall to the training side (NaT is neither <= nor > cutoff).
-    _train, holdout = temporal_split(out, "udm_asy_date", end_date="2025-10-17")
-    assert set(holdout["udm_mol_bat_id"]) == {"2"}
+    _train, holdout = temporal_split(out, "assay_date", end_date="2025-10-17")
+    assert set(holdout["compound_id"]) == {"2"}
 
 
 # ----------------------------------------------------------------------------
 # pull_multi_task_data — metadata columns
 # ----------------------------------------------------------------------------
-def test_pull_metadata_columns(monkeypatch):
-    """Per-source metadata survives the merge without becoming a task."""
+def test_pull_metadata_coalesces_across_sources(monkeypatch, caplog):
+    """Metadata from every source collapses to one column: first non-NaN, in source order, conflicts logged."""
     frames = {
         "src_a": pd.DataFrame(
-            {"udm_mol_bat_id": ["1", "2"], "smiles": ["CC", "CCC"], "ta": [1.0, 2.0], "qual_a": ["=", ">"]}
+            {"compound_id": ["1", "2"], "smiles": ["CC", "CCC"], "ta": [1.0, 2.0], "qual": ["=", ">"], "n": [5, 6]}
         ),
         "src_b": pd.DataFrame(
-            {"udm_mol_bat_id": ["2", "3"], "smiles": ["CCC", "CCCC"], "tb": [3.0, 4.0], "qual_b": ["<", "="]}
+            {"compound_id": ["2", "3"], "smiles": ["CCC", "CCCC"], "tb": [3.0, 4.0], "qual": ["<", "="]}
         ),
     }
     _install_fake_feature_set(monkeypatch, frames)
+    # workbench logger has propagate=False, so attach caplog's handler directly.
+    monkeypatch.setattr(logging.getLogger("workbench"), "handlers", [caplog.handler])
 
     out = pull_multi_task_data(
-        {
-            "src_a": {"target_info": ["ta"], "metadata": ["qual_a"]},
-            "src_b": {"target_info": ["tb"], "metadata": ["qual_b"]},
-        },
-        id_column="udm_mol_bat_id",
+        {"src_a": {"target_info": ["ta"]}, "src_b": {"target_info": ["tb"]}},
+        id_column="compound_id",
+        metadata=["qual", "n"],
     )
 
-    rows = out.set_index("udm_mol_bat_id")
-    assert rows.loc["1", "qual_a"] == "=" and pd.isna(rows.loc["1", "qual_b"])
-    assert rows.loc["2", "qual_a"] == ">" and rows.loc["2", "qual_b"] == "<"
-    assert pd.isna(rows.loc["3", "qual_a"]) and rows.loc["3", "qual_b"] == "="
+    rows = out.set_index("compound_id")
+    assert list(rows["qual"]) == ["=", ">", "="]
+    # 'n' is only in src_a: it survives the merge (a shared feature would be dropped), NaN for id 3
+    assert rows.loc["1", "n"] == 5 and rows.loc["2", "n"] == 6 and pd.isna(rows.loc["3", "n"])
+    # id 2's sources disagree on 'qual', so the conflict is logged
+    assert "'qual': 1 molecules have conflicting values" in caplog.text
 
 
-def test_pull_metadata_shared_name_coalesces(monkeypatch):
-    """Metadata listed by several sources collapses to one column: first non-NaN, in source order."""
+def test_pull_metadata_survives_public_source(monkeypatch):
+    """A SMILES-based source lacking a metadata column keeps the internal values, NaN for public-only rows."""
+    pytest.importorskip("rdkit")  # the smiles pass standardizes SMILES
     frames = {
-        "src_a": pd.DataFrame(
-            {"udm_mol_bat_id": ["1", "2"], "smiles": ["CC", "CCC"], "ta": [1.0, 2.0], "qual": ["=", ">"], "n": [5, 6]}
-        ),
-        "src_b": pd.DataFrame(
-            {"udm_mol_bat_id": ["2", "3"], "smiles": ["CCC", "CCCC"], "tb": [3.0, 4.0], "qual": ["<", "="], "n": [9, 7]}
+        "internal": pd.DataFrame(
+            {"compound_id": ["1", "2"], "smiles": ["CCO", "CCC"], "ta": [1.0, 2.0], "project_code": ["P1", "P2"]}
         ),
+        "public": pd.DataFrame({"id": ["a", "b"], "smiles": ["CCO", "CCCC"], "logp": [0.5, 2.0]}),
     }
     _install_fake_feature_set(monkeypatch, frames)
 
     out = pull_multi_task_data(
-        {
-            "src_a": {"target_info": ["ta"], "metadata": ["qual", "n"]},
-            "src_b": {"target_info": ["tb"], "metadata": ["qual", "n"]},
-        },
-        id_column="udm_mol_bat_id",
+        {"internal": {"target_info": ["ta"]}},
+        {"public": {"target_info": ["logp"], "src_id_col": "id"}},
+        id_column="compound_id",
+        metadata=["project_code"],
     )
 
-    rows = out.set_index("udm_mol_bat_id")
-    assert list(rows["qual"]) == ["=", ">", "="]
-    # Numeric metadata is carried, not averaged: id 2 keeps src_a's 6, not mean(6, 9)
-    assert list(rows["n"]) == [5, 6, 7]
+    prj = dict(zip(out["smiles"], out["project_code"]))
+    assert prj["CCO"] == "P1" and prj["CCC"] == "P2"
+    assert pd.isna(prj["CCCC"])
 
 
-def test_pull_metadata_target_name_collision_raises(monkeypatch):
-    """A metadata column can't share a name with a target."""
-    frames = {
-        "src_a": pd.DataFrame({"udm_mol_bat_id": ["1"], "smiles": ["CC"], "ta": [1.0]}),
-        "src_b": pd.DataFrame({"udm_mol_bat_id": ["1"], "smiles": ["CC"], "tb": [2.0], "ta": [3.0]}),
-    }
+def test_pull_metadata_missing_everywhere_raises(monkeypatch):
+    """A metadata column found in no source is almost always a typo."""
+    frames = {"src_a": pd.DataFrame({"compound_id": ["1"], "smiles": ["CC"], "ta": [1.0]})}
     _install_fake_feature_set(monkeypatch, frames)
 
-    with pytest.raises(ValueError, match="Duplicate target/passthrough"):
-        pull_multi_task_data(
-            {"src_a": {"target_info": ["ta"]}, "src_b": {"target_info": ["tb"], "metadata": ["ta"]}},
-            id_column="udm_mol_bat_id",
-        )
+    with pytest.raises(ValueError, match="not found in any source"):
+        pull_multi_task_data({"src_a": {"target_info": ["ta"]}}, id_column="compound_id", metadata=["typo"])