diff --git a/async_postgres/pg_connection.nim b/async_postgres/pg_connection.nim
index 6ef2d0e..9802c65 100644
--- a/async_postgres/pg_connection.nim
+++ b/async_postgres/pg_connection.nim
@@ -276,6 +276,14 @@ type
     wasClosed*: bool ## true if connection was closed instead of returned to pool
     handedToWaiter*: bool ## true if connection was given directly to a waiting acquirer
 
+  TracePoolCloseErrorData* = object
+    ## Data passed to the pool close-error hook. Fired when a pool-initiated
+    ## `conn.close()` raises — these errors are otherwise swallowed because
+    ## close runs from non-async cleanup paths and fire-and-forget tasks,
+    ## making leaks hard to observe without tracing.
+    conn*: PgConnection
+    err*: ref CatchableError
+
   PgTracer* = ref object
     ## Tracing hooks for async-postgres operations.
     ## Set only the callbacks you need; nil callbacks are skipped with zero overhead.
@@ -319,6 +327,7 @@ type
       proc(data: TracePoolReleaseStartData): TraceContext {.gcsafe, raises: [].}
     onPoolReleaseEnd*:
       proc(ctx: TraceContext, data: TracePoolReleaseEndData) {.gcsafe, raises: [].}
+    onPoolCloseError*: proc(data: TracePoolCloseErrorData) {.gcsafe, raises: [].}
 
 # Public API: read-only getters
 
diff --git a/async_postgres/pg_pool.nim b/async_postgres/pg_pool.nim
index c5acaec..0fe7b4e 100644
--- a/async_postgres/pg_pool.nim
+++ b/async_postgres/pg_pool.nim
@@ -162,6 +162,21 @@ proc metrics*(pool: PgPool): PoolMetrics =
   ## Cumulative pool metrics.
   pool.metrics
 
+proc reportCloseError(pool: PgPool, conn: PgConnection, err: ref CatchableError) =
+  ## Route a swallowed pool-initiated close error to the tracer. The pool
+  ## cannot propagate these errors to a caller (close runs from synchronous
+  ## cleanup paths and fire-and-forget tasks), so tracing is the only signal
+  ## operators have for leak detection.
+  if pool.config.tracer != nil and pool.config.tracer.onPoolCloseError != nil:
+    pool.config.tracer.onPoolCloseError(TracePoolCloseErrorData(conn: conn, err: err))
+
+proc tracedClose(pool: PgPool, conn: PgConnection) {.async.} =
+  ## Close `conn`, reporting any close error via `reportCloseError`.
+  try:
+    await conn.close()
+  except CatchableError as e:
+    pool.reportCloseError(conn, e)
+
 proc closeNoWait(pool: PgPool, conn: PgConnection) =
   ## Schedule connection close without waiting. For use in non-async contexts
   ## (e.g. `release()` is synchronous). The spawned task is tracked in
@@ -171,14 +186,12 @@ proc closeNoWait(pool: PgPool, conn: PgConnection) =
   ## Note on asyncdispatch: a close scheduled here may race with an inflight
   ## request future that the previous timeout could not cancel (see
   ## `invalidateOnTimeout`). That future will observe a closed fd and fail
-  ## quietly — we rely on `asyncSpawn` swallowing the resulting CatchableError
-  ## so it never surfaces. The connection is not reused either way.
+  ## quietly — `tracedClose` catches the error and routes it to the
+  ## `onPoolCloseError` tracer hook (nil when unconfigured). The connection
+  ## is not reused either way.
   pool.metrics.closeCount.inc
   proc doClose() {.async.} =
-    try:
-      await conn.close()
-    except CatchableError:
-      discard
+    await pool.tracedClose(conn)
 
   # Prune only once the seq grows past the threshold so the sweep is amortized
   # instead of O(n) on every call. Uses swap-remove (constant-time delete that
@@ -207,10 +220,7 @@ proc resetSession*(pool: PgPool, conn: PgConnection) {.async.} =
       discard await conn.simpleExec(pool.config.resetQuery)
       conn.clearStmtCache()
     except CatchableError:
-      try:
-        await conn.close()
-      except CatchableError:
-        discard
+      await pool.tracedClose(conn)
 
 proc maintenanceLoop(pool: PgPool) {.async.} =
   while not pool.closed:
@@ -227,20 +237,14 @@ proc maintenanceLoop(pool: PgPool) {.async.} =
       # Always close broken or in-transaction connections (unusable)
       if pc.conn.state != csReady or pc.conn.txStatus != tsIdle:
         pool.metrics.closeCount.inc
-        try:
-          await pc.conn.close()
-        except CatchableError:
-          discard
+        await pool.tracedClose(pc.conn)
         continue
 
       # Always close max-lifetime-exceeded connections (acquire rejects them anyway)
       if pool.config.maxLifetime > ZeroDuration and
           now - pc.conn.createdAt > pool.config.maxLifetime:
         pool.metrics.closeCount.inc
-        try:
-          await pc.conn.close()
-        except CatchableError:
-          discard
+        await pool.tracedClose(pc.conn)
         continue
 
       # Idle timeout respects minSize
@@ -249,10 +253,7 @@ proc maintenanceLoop(pool: PgPool) {.async.} =
         let totalCount = remaining.len + pool.idle.len + pool.active
         if totalCount >= pool.config.minSize:
           pool.metrics.closeCount.inc
-          try:
-            await pc.conn.close()
-          except CatchableError:
-            discard
+          await pool.tracedClose(pc.conn)
           continue
 
       remaining.addLast(pc)
@@ -305,10 +306,7 @@ proc newPool*(config: PoolConfig): Future[PgPool] {.async.} =
   except CatchableError as e:
     while pool.idle.len > 0:
       let pc = pool.idle.popFirst()
-      try:
-        await pc.conn.close()
-      except CatchableError:
-        discard
+      await pool.tracedClose(pc.conn)
     raise e
 
   pool.maintenanceTask = maintenanceLoop(pool)
@@ -379,18 +377,12 @@ proc acquireImpl(pool: PgPool): Future[AcquireResult] {.async.} =
     let pc = pool.idle.popFirst()
     if pc.conn.state != csReady:
       pool.metrics.closeCount.inc
-      try:
-        await pc.conn.close()
-      except CatchableError:
-        discard
+      await pool.tracedClose(pc.conn)
       continue
     if pool.config.maxLifetime > ZeroDuration and
         pool.cachedNow - pc.conn.createdAt > pool.config.maxLifetime:
       pool.metrics.closeCount.inc
-      try:
-        await pc.conn.close()
-      except CatchableError:
-        discard
+      await pool.tracedClose(pc.conn)
       continue
     # Health check: ping connections that have been idle too long
     if pool.config.healthCheckTimeout > ZeroDuration and
@@ -399,10 +391,7 @@ proc acquireImpl(pool: PgPool): Future[AcquireResult] {.async.} =
         await pc.conn.ping(pool.config.pingTimeout)
       except CatchableError:
         pool.metrics.closeCount.inc
-        try:
-          await pc.conn.close()
-        except CatchableError:
-          discard
+        await pool.tracedClose(pc.conn)
         continue
     pool.active.inc
     recordAcquire()
@@ -1061,10 +1050,7 @@ proc close*(pool: PgPool, timeout = ZeroDuration): Future[void] {.async.} =
   while pool.idle.len > 0:
     let pc = pool.idle.popFirst()
     pool.metrics.closeCount.inc
-    try:
-      await pc.conn.close()
-    except CatchableError:
-      discard
+    await pool.tracedClose(pc.conn)
 
   # Wait for any fire-and-forget closes spawned via closeNoWait so the server
   # observes Terminate and fds are released before this proc returns. A late
diff --git a/tests/test_tracing.nim b/tests/test_tracing.nim
index 31d805b..9306e0c 100644
--- a/tests/test_tracing.nim
+++ b/tests/test_tracing.nim
@@ -110,6 +110,10 @@ type
     wasClosed: bool
     handedToWaiter: bool
 
+  PoolCloseErrorRec = object
+    hasConn: bool
+    errMsg: string
+
   TraceLog = ref object
     connectStarts: seq[ConnectStartRec]
     connectEnds: seq[ConnectEndRec]
@@ -125,6 +129,7 @@ type
     poolAcquireEnds: seq[PoolAcquireEndRec]
     poolReleaseStarts: seq[PoolReleaseStartRec]
     poolReleaseEnds: seq[PoolReleaseEndRec]
+    poolCloseErrors: seq[PoolCloseErrorRec]
 
 proc newTraceLog(): TraceLog =
   TraceLog()
@@ -249,6 +254,13 @@ proc buildTracer(log: TraceLog): PgTracer =
       )
     )
 
+  tracer.onPoolCloseError = proc(data: TracePoolCloseErrorData) {.gcsafe, raises: [].} =
+    log.poolCloseErrors.add(
+      PoolCloseErrorRec(
+        hasConn: data.conn != nil, errMsg: (if data.err != nil: data.err.msg else: "")
+      )
+    )
+
   return tracer
 
 proc tracedConfig(tracer: PgTracer): ConnConfig =
@@ -637,6 +649,90 @@ suite "Tracing: pool acquire/release":
 
     waitFor t()
 
+suite "Tracing: pool close errors":
+  test "onPoolCloseError reports swallowed close failures":
+    proc t() {.async.} =
+      let log = newTraceLog()
+      let tracer = buildTracer(log)
+      var poolCfg = initPoolConfig(tracedConfig(tracer), minSize = 0, maxSize = 1)
+      poolCfg.tracer = tracer
+      let pool = await newPool(poolCfg)
+
+      let conn = await pool.acquire()
+
+      # Drive the hook through reportCloseError — the single chokepoint that
+      # tracedClose funnels errors through. Inducing a real conn.close() failure
+      # is impractical in tests: close() is idempotent and its only raising path
+      # is inside sendMsg, which it already swallows internally.
+      let err = newException(PgError, "simulated close failure")
+      pool.reportCloseError(conn, err)
+
+      doAssert log.poolCloseErrors.len == 1
+      doAssert log.poolCloseErrors[0].hasConn
+      doAssert log.poolCloseErrors[0].errMsg == "simulated close failure"
+
+      pool.release(conn)
+      await pool.close()
+
+    waitFor t()
+
+  test "nil onPoolCloseError hook is a no-op":
+    proc t() {.async.} =
+      # Build a tracer that sets OTHER hooks but leaves onPoolCloseError nil.
+      let tracer = PgTracer()
+      var poolCfg = initPoolConfig(plainConfig(), minSize = 0, maxSize = 1)
+      poolCfg.tracer = tracer
+      let pool = await newPool(poolCfg)
+
+      let conn = await pool.acquire()
+      # Must not raise even though the hook is nil.
+      pool.reportCloseError(conn, newException(PgError, "ignored"))
+
+      pool.release(conn)
+      await pool.close()
+
+    waitFor t()
+
+  test "tracedClose on healthy connection does not fire hook":
+    proc t() {.async.} =
+      let log = newTraceLog()
+      let tracer = buildTracer(log)
+      var poolCfg = initPoolConfig(tracedConfig(tracer), minSize = 0, maxSize = 1)
+      poolCfg.tracer = tracer
+      let pool = await newPool(poolCfg)
+
+      # Use a stand-alone connection so pool.active/idle counters stay
+      # consistent — tracedClose only touches the tracer, not pool state.
+      let conn = await connect(tracedConfig(tracer))
+      await pool.tracedClose(conn)
+
+      doAssert log.poolCloseErrors.len == 0
+
+      await pool.close()
+
+    waitFor t()
+
+  test "tracedClose on already-closed connection is a no-op":
+    proc t() {.async.} =
+      let log = newTraceLog()
+      let tracer = buildTracer(log)
+      var poolCfg = initPoolConfig(tracedConfig(tracer), minSize = 0, maxSize = 1)
+      poolCfg.tracer = tracer
+      let pool = await newPool(poolCfg)
+
+      # conn.close() is documented as idempotent. tracedClose against an
+      # already-closed conn exercises the real close path end-to-end and must
+      # neither raise nor fire the error hook.
+      let conn = await connect(tracedConfig(tracer))
+      await conn.close()
+      await pool.tracedClose(conn)
+
+      doAssert log.poolCloseErrors.len == 0
+
+      await pool.close()
+
+    waitFor t()
+
 suite "Tracing: queryEach":
   test "onQueryStart and onQueryEnd are called with rowCount":
     proc t() {.async.} =