diff --git a/changelog.d/full-scale-string-storage-canonicalization.fixed.md b/changelog.d/full-scale-string-storage-canonicalization.fixed.md
new file mode 100644
index 00000000..82530e96
--- /dev/null
+++ b/changelog.d/full-scale-string-storage-canonicalization.fixed.md
@@ -0,0 +1 @@
+Canonicalize pandas string storage at the ACS PUMS parse and ASEC raw-stage checkpoint load boundaries, and render dtype reprs in the spine-assembly refusal: pyarrow-carrying environments resolved fresh string casts to pyarrow storage, which failed shared-column dtype validation against checkpoint-restored python-storage channels with both sides printing as 'str'.
diff --git a/packages/microcosm-build/src/microcosm/build/us_runtime/acs_pums.py b/packages/microcosm-build/src/microcosm/build/us_runtime/acs_pums.py
index 81ebd047..dd0456fc 100644
--- a/packages/microcosm-build/src/microcosm/build/us_runtime/acs_pums.py
+++ b/packages/microcosm-build/src/microcosm/build/us_runtime/acs_pums.py
@@ -32,6 +32,7 @@
 import numpy as np
 import pandas as pd
 
+from microcosm.build.serialization_dtypes import canonicalize_frame_string_dtypes
 from microcosm.frame import US_SCHEMA, Frame, WeightKind, Weights
 from microcosm.frame.units import assign_us_unit_structure
 
@@ -309,6 +310,14 @@ def build_acs_pums_unit_frame(
             ),
         }
     )
+    # This parse is a serialization boundary: freshly cast string columns
+    # otherwise carry the environment-resolved pandas string storage, which
+    # diverges from checkpoint-restored channels when pyarrow is installed.
+    frame = canonicalize_frame_string_dtypes(
+        frame,
+        boundary="ACS PUMS source parse",
+        in_place=True,
+    )
     return frame, metadata
 
 
diff --git a/packages/microcosm-build/src/microcosm/build/us_runtime/asec_checkpoint.py b/packages/microcosm-build/src/microcosm/build/us_runtime/asec_checkpoint.py
index 9e7b21f8..0c6ebe0f 100644
--- a/packages/microcosm-build/src/microcosm/build/us_runtime/asec_checkpoint.py
+++ b/packages/microcosm-build/src/microcosm/build/us_runtime/asec_checkpoint.py
@@ -23,6 +23,7 @@
     FrameIdentity,
     frame_identity,
 )
+from microcosm.build.serialization_dtypes import canonicalize_frame_string_dtypes
 from microcosm.build.us_runtime.operator_boundary import (
     assert_operator_free_source_frame,
 )
@@ -187,7 +188,16 @@ def load_asec_raw_stage_checkpoint(
         )
     metadata["identity"] = stored_identity.to_payload()
     metadata["source_construction_identity"] = source_construction_identity.to_payload()
-    return loaded.frame, metadata
+    # Canonicalize only after every identity and binding check has passed on
+    # the restored representation: this load is a serialization boundary, and
+    # downstream spine assembly requires one physical string dtype per shared
+    # column regardless of the storage the checkpoint was written under.
+    frame = canonicalize_frame_string_dtypes(
+        loaded.frame,
+        boundary="ASEC raw-stage checkpoint load",
+        in_place=True,
+    )
+    return frame, metadata
 
 
 def _validate_outer_stage_binding(
diff --git a/packages/microcosm-build/src/microcosm/build/us_runtime/spine_assembly.py b/packages/microcosm-build/src/microcosm/build/us_runtime/spine_assembly.py
index 21a4c24d..ce6b6e29 100644
--- a/packages/microcosm-build/src/microcosm/build/us_runtime/spine_assembly.py
+++ b/packages/microcosm-build/src/microcosm/build/us_runtime/spine_assembly.py
@@ -296,8 +296,10 @@ def _validate_shared_column_dtypes(frames: Mapping[str, Frame]) -> None:
             if not all(dtype == values[0][1] for _, dtype in values[1:])
         }
         if mismatches:
+            # repr, not str: pandas string dtypes with different storages all
+            # str() as 'str', which renders the mismatch invisible.
             details = ", ".join(
-                f"{column}={[(channel, str(dtype)) for channel, dtype in values]}"
+                f"{column}={[(channel, repr(dtype)) for channel, dtype in values]}"
                 for column, values in sorted(mismatches.items())
             )
             raise ValueError(
diff --git a/packages/microcosm-build/tests/test_us_acs_pums.py b/packages/microcosm-build/tests/test_us_acs_pums.py
index 0c7fd22d..48b656d9 100644
--- a/packages/microcosm-build/tests/test_us_acs_pums.py
+++ b/packages/microcosm-build/tests/test_us_acs_pums.py
@@ -7,6 +7,7 @@
 import pandas as pd
 import pytest
 
+from microcosm.build.serialization_dtypes import CANONICAL_STRING_DTYPE
 from microcosm.build.us_runtime.acs_pums import (
     ACS_2024_1YR_SPINE,
     AcsPumsSource,
@@ -138,7 +139,12 @@ def _asec_shaped_frame() -> Frame:
             "person_marital_unit_id": np.asarray([601], dtype=np.int64),
             "source_year": np.asarray([2024], dtype=np.int64),
             "source_household_id": np.asarray([7], dtype=np.int64),
-            "source_person_id": pd.Series(["7-1"]).astype(str),
+            # Pinned to the canonical storage, exactly as checkpoint restore
+            # delivers the real ASEC channel. A bare astype(str) would follow
+            # the environment's storage default and mask the cross-channel
+            # divergence this fixture exists to model (pyarrow environments
+            # resolve fresh 'str' casts to pyarrow storage).
+            "source_person_id": pd.Series(["7-1"]).astype(CANONICAL_STRING_DTYPE),
             "source_row_id": np.asarray([0], dtype=np.int64),
             "A_AGE": np.asarray([55], dtype=np.int64),
         }
@@ -287,6 +293,31 @@ def test_built_acs_lineage_assembles_with_asec_without_measured_coercion(
     )
 
 
+def test_build_acs_pums_unit_frame_canonicalizes_string_storage(
+    tmp_path: Path,
+) -> None:
+    """The parse boundary must not leak environment-resolved string storage.
+
+    With pyarrow installed, pandas resolves fresh ``astype(str)`` casts to
+    pyarrow-backed storage while checkpoint-restored channels carry the
+    canonical python-backed dtype — the exact split that failed the first
+    full-scale spine assembly (both sides printed as 'str').
+    """
+
+    pytest.importorskip("microunit")
+    frame, _metadata = build_acs_pums_unit_frame(_source(tmp_path), chunksize=1)
+
+    for entity in frame.entities:
+        table = frame.table(entity)
+        offending = {
+            column: dtype
+            for column, dtype in table.dtypes.items()
+            if isinstance(dtype, pd.StringDtype) and dtype != CANONICAL_STRING_DTYPE
+        }
+        assert not offending, f"{entity} carries non-canonical string storage"
+    assert frame.table("person")["source_person_id"].dtype == CANONICAL_STRING_DTYPE
+
+
 def test_load_acs_pums_tables_rejects_duplicate_household_serialno(
     tmp_path: Path,
 ) -> None:
diff --git a/packages/microcosm-build/tests/test_us_asec_checkpoint.py b/packages/microcosm-build/tests/test_us_asec_checkpoint.py
index d0ce5d6e..4a604a3f 100644
--- a/packages/microcosm-build/tests/test_us_asec_checkpoint.py
+++ b/packages/microcosm-build/tests/test_us_asec_checkpoint.py
@@ -12,6 +12,7 @@
     OUTER_STAGE_CONTEXT_SCHEMA_VERSION,
     frame_identity,
 )
+from microcosm.build.serialization_dtypes import CANONICAL_STRING_DTYPE
 from microcosm.build.us_runtime import (
     ASEC_RAW_STAGE_ARTIFACT_KIND,
     ASEC_RAW_STAGE_OPERATOR_STATUS,
@@ -236,6 +237,13 @@ def test_loads_operator_untouched_raw_stage_checkpoint(tmp_path: Path) -> None:
     assert loaded_metadata["artifact_kind"] == ASEC_RAW_STAGE_ARTIFACT_KIND
     assert loaded_metadata["stage"] == ASEC_RAW_STAGE_STAGE
     assert loaded_metadata["operator_status"] == ASEC_RAW_STAGE_OPERATOR_STATUS
+    # The load is a declared string-storage canonicalization boundary: no
+    # entity may leak a non-canonical pandas string dtype downstream,
+    # whatever storage the checkpoint was written under.
+    for entity in frame.entities:
+        for column, dtype in frame.table(entity).dtypes.items():
+            if isinstance(dtype, pd.StringDtype):
+                assert dtype == CANONICAL_STRING_DTYPE, (entity, column)
 
 
 @pytest.mark.parametrize(
diff --git a/packages/microcosm-build/tests/test_us_spine_assembly.py b/packages/microcosm-build/tests/test_us_spine_assembly.py
index 1763e27f..947e91af 100644
--- a/packages/microcosm-build/tests/test_us_spine_assembly.py
+++ b/packages/microcosm-build/tests/test_us_spine_assembly.py
@@ -307,6 +307,52 @@ def test_assemble_spines__rejects_shared_dtype_mismatch() -> None:
         )
 
 
+def test_assemble_spines__string_storage_mismatch_is_named_in_the_error() -> None:
+    """Divergent pandas string storages must be visible in the refusal.
+
+    Both python- and pyarrow-backed 'str' dtypes print identically under
+    str(); the first full-scale run failed with an error whose two sides both
+    read 'str'. The refusal must render repr so the storages are legible.
+    """
+
+    pytest.importorskip("pyarrow")
+
+    def _with_probe(frame: Frame, storage: str, value: str) -> Frame:
+        tables = {entity: frame.table(entity).copy() for entity in frame.entities}
+        person = tables["person"]
+        person["stacked_probe_label"] = pd.Series(
+            [value] * len(person),
+            index=person.index,
+        ).astype(pd.StringDtype(storage=storage, na_value=np.nan))
+        return Frame(
+            tables,
+            US_SCHEMA,
+            {"household": frame.weights_for("household")},
+            frame.strata,
+        )
+
+    with pytest.raises(
+        ValueError, match="identical dtypes.*stacked_probe_label"
+    ) as excinfo:
+        assemble_spines(
+            {
+                "asec": _with_probe(_asec_frame(), "python", "x"),
+                "acs": _with_probe(_acs_frame(), "pyarrow", "y"),
+            },
+            household_mass_shares={"asec": 0.5, "acs": 0.5},
+        )
+    message = str(excinfo.value)
+    # The python-backed side names its storage explicitly; pandas omits the
+    # storage kwarg from the repr of the environment-default (pyarrow) side.
+    # Both reprs must appear and must differ — under str() both sides
+    # rendered as 'str' and the mismatch was unreadable.
+    asec_repr = repr(pd.StringDtype(storage="python", na_value=np.nan))
+    acs_repr = repr(pd.StringDtype(storage="pyarrow", na_value=np.nan))
+    assert asec_repr != acs_repr
+    assert asec_repr in message
+    assert acs_repr in message
+
+
 def test_assemble_spines__owns_support_provenance() -> None:
     asec = _asec_frame()
     tables = {entity: asec.table(entity).copy() for entity in asec.entities}