diff --git a/src/db.py b/src/db.py
index 6c7d5866..c543ce26 100644
--- a/src/db.py
+++ b/src/db.py
@@ -119,8 +119,11 @@ async def acquire_connection_with_retry(db: AsyncSession, context: str) -> None:
     in a Sentry span so wait time is visible in traces; on budget exhaustion the
     original error is reraised after capturing live pool stats to Sentry.
 
-    Retrying the same session is safe: no connection is bound until checkout
-    succeeds, so each attempt re-attempts the checkout cleanly.
+    Each attempt rolls the session back on a retryable failure before retrying:
+    a failed checkout can leave the autobegun transaction in a pending-rollback
+    state, which would make the next ``db.connection()`` raise instead of
+    re-checking-out cleanly. The rollback is pure Python-side state cleanup when
+    no connection was bound, so it is cheap and safe.
     """
     with sentry_sdk.start_span(op="db.pool.acquire", name=context):
         if not settings.DB.CONNECTION_RETRY_ENABLED:
@@ -139,7 +142,18 @@ async def acquire_connection_with_retry(db: AsyncSession, context: str) -> None:
             ):
                 with attempt:
                     attempts += 1
-                    await db.connection()
+                    try:
+                        await db.connection()
+                    except RETRYABLE_DB_CONNECTION_ERRORS:
+                        # Reset session state so the next attempt starts clean.
+                        try:
+                            await db.rollback()
+                        except Exception:
+                            logger.debug(
+                                "rollback after failed checkout failed",
+                                exc_info=True,
+                            )
+                        raise
         except RETRYABLE_DB_CONNECTION_ERRORS as e:
             _record_acquisition_outcome("exhausted")
             if settings.SENTRY.ENABLED:
diff --git a/src/telemetry/prometheus/metrics.py b/src/telemetry/prometheus/metrics.py
index d883816b..1a79c634 100644
--- a/src/telemetry/prometheus/metrics.py
+++ b/src/telemetry/prometheus/metrics.py
@@ -323,17 +323,25 @@ class DBPoolCollector:
         self.instance_type: str = instance_type
 
     def collect(self) -> Iterator[GaugeMetricFamily]:
-        # Lazy import to avoid an import cycle at module load (db imports config,
-        # telemetry is imported widely). Reads the async engine.pool directly.
-        from src.db import get_pool_stats
-
         namespace = settings.METRICS.NAMESPACE or ""
         gauge = GaugeMetricFamily(
             "db_pool_connections",
             "DB connections held by this instance, by pool state",
             labels=["namespace", "instance_type", "state"],
         )
-        for state, value in get_pool_stats().items():
+        # Fail soft: Prometheus aborts the entire scrape (dropping ALL metrics)
+        # if any collector raises, so never let a pool/import hiccup here sink
+        # the whole /metrics response.
+        try:
+            # Lazy import to avoid an import cycle at module load (db imports
+            # config, telemetry is imported widely). Reads engine.pool directly.
+            from src.db import get_pool_stats
+
+            stats = get_pool_stats()
+        except Exception:
+            logger.warning("Failed to collect DB pool stats", exc_info=True)
+            stats = {}
+        for state, value in stats.items():
             gauge.add_metric([namespace, self.instance_type, state], value)
         yield gauge
 
diff --git a/tests/test_db_resilience.py b/tests/test_db_resilience.py
index 0177373e..696a6c53 100644
--- a/tests/test_db_resilience.py
+++ b/tests/test_db_resilience.py
@@ -31,12 +31,16 @@ class _FlakyConnSession:
         self.fail_times: int = fail_times
         self.always_fail: bool = always_fail
         self.calls: int = 0
+        self.rollback_calls: int = 0
 
     async def connection(self) -> None:
         self.calls += 1
         if self.always_fail or self.calls <= self.fail_times:
             raise _make_operational_error()
 
+    async def rollback(self) -> None:
+        self.rollback_calls += 1
+
 
 def _acq_count(outcome: str) -> float:
     child = db_connection_acquisitions_counter.labels(
@@ -73,6 +77,8 @@ async def test_acquire_retries_then_succeeds_records_retried() -> None:
     session = _FlakyConnSession(fail_times=2)
     await acquire_connection_with_retry(session, "request:test")  # pyright: ignore[reportArgumentType]
     assert session.calls == 3  # two failures then success
+    # Session is reset after each failed checkout before the next attempt.
+    assert session.rollback_calls == 2
     assert _acq_count("retried") == before + 1