diff --git a/src/everos/core/persistence/lancedb/repository.py b/src/everos/core/persistence/lancedb/repository.py
index 6454c032d..3dab3c123 100644
--- a/src/everos/core/persistence/lancedb/repository.py
+++ b/src/everos/core/persistence/lancedb/repository.py
@@ -224,11 +224,12 @@ def _remove_empty_index_dirs(
 retry clears. ``LanceError(IO): Execution error: Spill has sent an error`` is
 DataFusion's sort / merge spill to the OS temp dir failing mid-query; lancedb
 raises it as a bare ``RuntimeError``. Seen only on the Windows soak box under
-nine concurrent clients (187 / 180 / 34 times over three runs), never on an
+nine concurrent clients (187 / 180 / 34 / 42 times over four runs), never on an
 idle box, and the same row projected fine on the next attempt — yet the worker
 filed every one as unrecoverable, so ~200 md files per run needed a manual
-``cascade fix``. Match the exact phrase: a generic IO error (disk full, file
-gone) must stay permanent."""
+``cascade fix``. The traceback frames put it in ``merge_insert`` (the write
+path, under :meth:`_locked`); the read path is covered as well. Match the
+exact phrase: a generic IO error (disk full, file gone) must stay permanent."""
 
 
 def _is_transient_execution_error(exc: BaseException) -> bool:
@@ -390,6 +391,22 @@ async def _locked(self, budget: float, op: str) -> AsyncIterator[None]:
                 f"{op} on table {self.table_name!r} exceeded its "
                 f"{budget:g}s write-lock deadline"
             ) from exc
+        except RuntimeError as exc:
+            # The soak's spill failures came out of ``merge_insert`` (this
+            # path), not the reads: worker -> upsert -> execute_merge_insert.
+            # The write is idempotent by id, so a retry is the right answer.
+            if not _is_transient_execution_error(exc):
+                raise
+            logger.warning(
+                "lancedb_transient_execution_error",
+                table=self.table_name,
+                op=op,
+                error=str(exc)[:200],
+            )
+            raise VectorStoreBusyError(
+                f"{op} on table {self.table_name!r} hit a transient lance "
+                f"execution error: {exc}"
+            ) from exc
         else:
             held = time.monotonic() - (acquired_at or started)
             if held >= _SLOW_HOLD_LOG_SECONDS:
diff --git a/tests/unit/test_core/test_persistence/test_lancedb/test_transient_execution_errors.py b/tests/unit/test_core/test_persistence/test_lancedb/test_transient_execution_errors.py
index 800241968..018929c80 100644
--- a/tests/unit/test_core/test_persistence/test_lancedb/test_transient_execution_errors.py
+++ b/tests/unit/test_core/test_persistence/test_lancedb/test_transient_execution_errors.py
@@ -41,6 +41,13 @@ async def test_spill_failure_inside_a_read_becomes_a_busy_error() -> None:
             raise RuntimeError(_SPILL)
 
 
+async def test_spill_failure_inside_a_write_becomes_a_busy_error() -> None:
+    """The soak's spill failures came out of merge_insert (the write path)."""
+    with pytest.raises(VectorStoreBusyError, match="transient lance execution"):
+        async with _Repo()._locked(1.0, "upsert"):
+            raise RuntimeError(_SPILL)
+
+
 async def test_other_runtime_errors_still_propagate_unchanged() -> None:
     with pytest.raises(RuntimeError, match="No space left"):
         async with _Repo()._deadline(1.0, "find_where"):