diff --git a/backend/tests/integration/conftest.py b/backend/tests/integration/conftest.py
index 8f3c24f1..0d65fe22 100644
--- a/backend/tests/integration/conftest.py
+++ b/backend/tests/integration/conftest.py
@@ -173,12 +173,75 @@ async def _truncate_all(engine) -> None:
             await conn.execute(text(f"TRUNCATE {quoted} RESTART IDENTITY CASCADE"))
 
 
+async def _drain_background_work(timeout: float = 30.0) -> int:
+    """Wait for whatever a test left running in the background (B-21). Never cancels.
+
+    Route handlers start work with `spawn_tracked`, and in-process `task_queue`
+    fallbacks with `asyncio.create_task`. Neither is awaited by the request, so they
+    outlived the test that started them — and on SQLite every session shares ONE
+    physical connection (`StaticPool`, above). A leftover task that rolls back on that
+    connection can undo the NEXT test's flushed-but-uncommitted insert; that test then
+    commits nothing, finds the object in its identity map without a query, and its
+    `UPDATE` matches zero rows — the `StaleDataError` `test_learnings_api` hit on
+    2026-09-23 and 2026-09-26, passing alone and on rerun.
+
+    **Waiting, not cancelling, and that was learned the hard way.** The first version
+    cancelled: a task cancelled in the middle of a statement makes SQLAlchemy invalidate
+    the connection, and under `StaticPool` the replacement is a NEW, EMPTY `:memory:`
+    database, so every later test failed with "no such table: users" (#439's first CI
+    run: 263 failed, 965 errors). Waiting lets each task finish its statement and its
+    transaction. A task still running after `timeout` is left alone and named in a
+    warning rather than cut mid-query. Returns how many were still running.
+    """
+    import asyncio
+    import logging
+
+    from app.core import task_queue
+    from app.core.background import _BACKGROUND_TASKS
+
+    pending = [
+        t for t in (*_BACKGROUND_TASKS, *task_queue._fallback_tasks.values()) if not t.done()
+    ]
+    if not pending:
+        return 0
+    _done, still = await asyncio.wait(pending, timeout=timeout)
+    if still:
+        logging.getLogger(__name__).warning(
+            "%d background task(s) still running after %.0fs, left alone: %s",
+            len(still),
+            timeout,
+            sorted(t.get_name() for t in still),
+        )
+    return len(still)
+
+
+async def _assert_schema_survived(engine) -> None:
+    """Fail at the cause, not a thousand tests later: the shared SQLite DB kept its tables."""
+    if not _ON_SQLITE:
+        return
+    from sqlalchemy import inspect as _inspect
+
+    async with engine.connect() as conn:
+        present = await conn.run_sync(lambda c: set(_inspect(c).get_table_names()))
+    missing = _CREATED_TABLES - present
+    if missing:
+        raise RuntimeError(
+            "the shared in-memory database lost its schema during this test's teardown "
+            f"({len(missing)} tables missing, e.g. {sorted(missing)[:3]}) — the StaticPool "
+            "connection was replaced, usually by a statement cancelled mid-flight"
+        )
+
+
 @pytest_asyncio.fixture()
 async def db_session(engine) -> AsyncGenerator[AsyncSession, None]:
     factory = async_sessionmaker(engine, class_=AsyncSession, expire_on_commit=False)
     async with factory() as session:
         yield session
+        # Before the rollback and the truncate: a task still running would otherwise
+        # execute against the next test's rows on the shared connection (B-21).
+        await _drain_background_work()
         await session.rollback()
+    await _assert_schema_survived(engine)
     await _truncate_all(engine)
 
 
diff --git a/backend/tests/integration/test_background_work_does_not_outlive_its_test.py b/backend/tests/integration/test_background_work_does_not_outlive_its_test.py
new file mode 100644
index 00000000..5d8bd476
--- /dev/null
+++ b/backend/tests/integration/test_background_work_does_not_outlive_its_test.py
@@ -0,0 +1,37 @@
+"""B-21: background work a test starts has finished before the next test begins.
+
+The two tests run in file order. The first starts a task that takes a moment and does
+not await it; the second asserts the `db_session` teardown waited for it — and did not
+cancel it, because cancelling mid-statement destroys the shared SQLite database
+(see `_drain_background_work`). Without the drain the task is still sleeping when the
+second test starts.
+"""
+
+from __future__ import annotations
+
+import asyncio
+
+import pytest
+
+from app.core.background import _BACKGROUND_TASKS, spawn_tracked
+
+_STARTED: list[asyncio.Task] = []
+_FINISHED: list[bool] = []
+
+
+@pytest.mark.asyncio
+async def test_a_test_that_leaves_work_running(db_session):
+    async def a_moment_of_work():
+        await asyncio.sleep(0.5)
+        _FINISHED.append(True)
+
+    _STARTED.append(spawn_tracked(a_moment_of_work(), name="b21-a-moment"))
+    assert not _FINISHED
+
+
+@pytest.mark.asyncio
+async def test_the_next_test_does_not_inherit_it(db_session):
+    assert _STARTED, "the first test did not run first"
+    assert _FINISHED == [True], "the previous test's work was still running"
+    assert not _STARTED[0].cancelled(), "the drain waits; it must not cancel"
+    assert not [t for t in _BACKGROUND_TASKS if not t.done()]