diff --git a/CHANGELOG.md b/CHANGELOG.md
index 9628c18..dd5b5fa 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -13,6 +13,12 @@ gem but are what an operator runs against their own image.
 
 ## next / unreleased
 
+### HotCell::Client
+
+#### Added
+
+* The `perform.hot_cell` event carries `stderr`, the failure's captured stream, beside `signal`. A subscriber can log the diagnosis of a crash — `libgomp: Thread creation failed` — on the same line as its cause. Before, that text survived only in the exception's message, so a failure that was discarded rather than retried lost it. The field is already bounded by `Failure.sanitize`; it is text a tool wrote while processing a hostile file, so write it to a log field and interpolate it nowhere else.
+
 ### HotCell::Core
 
 #### Fixed
diff --git a/hotcell-client/lib/hot_cell/client.rb b/hotcell-client/lib/hot_cell/client.rb
index 27a07ff..b40364d 100644
--- a/hotcell-client/lib/hot_cell/client.rb
+++ b/hotcell-client/lib/hot_cell/client.rb
@@ -160,6 +160,11 @@ def verify_output(response, outputs)
       # the code filed every fsize kill as transient. `permanent` is the Failure's own answer, so no
       # subscriber re-derives it.
       #
+      # `stderr` is the diagnosis of a crash — `libgomp: Thread creation failed` — and the event is the only
+      # place a subscriber can log it: the exception's message carries it too, but only a retry logs that,
+      # and a discarded failure loses it. It is text a tool wrote while processing a hostile file, bounded
+      # by `Failure.sanitize` on the way in, and a subscriber should write it to a log field and nowhere else.
+      #
       # A subscriber's own duration minus perform_ms is transport plus queueing, and those want separate
       # metrics: a rising perform_ms means the work got more expensive, and a rising difference means the
       # cell is saturated.
@@ -171,6 +176,7 @@ def publish(event, cell, response, inputs, outputs)
         event[:code] = failure&.code
         event[:cause] = failure&.cause
         event[:signal] = failure&.signal
+        event[:stderr] = failure&.stderr
         event[:permanent] = failure&.permanent?
         event[:bytes_in] = byte_count(inputs)
         event[:bytes_out] = byte_count(outputs)
diff --git a/hotcell-client/test/classification_test.rb b/hotcell-client/test/classification_test.rb
index 32aa6dd..b24a9c8 100644
--- a/hotcell-client/test/classification_test.rb
+++ b/hotcell-client/test/classification_test.rb
@@ -134,6 +134,19 @@ def test_the_event_carries_the_cause_and_the_verdict_so_a_subscriber_can_classif
     assert_equal true, event[:permanent]
   end
 
+  # A crash's diagnosis is the tool's stderr — `libgomp: Thread creation failed` — and the event is the
+  # only place a subscriber can log it, since a discarded failure never reaches the retry line.
+  def test_the_event_carries_the_stderr_so_a_subscriber_can_log_the_diagnosis_of_a_crash
+    register_with failed(code: "killed", cause: "crashed", stderr: "libgomp: Thread creation failed\n")
+
+    event = events_for do
+      assert_raises(TemporarilyUnavailable) { Anything.perform_in_hotcell [], [], {} }
+    end.first.payload
+
+    assert_equal "crashed", event[:cause]
+    assert_equal "libgomp: Thread creation failed\n", event[:stderr]
+  end
+
   def test_the_event_carries_a_transient_verdict_for_a_deadline_kill
     register_with failed(code: "killed", cause: "deadline")
 
@@ -154,6 +167,7 @@ def test_the_event_carries_no_code_on_success
     assert_nil event[:code]
     assert_nil event[:cause]
     assert_nil event[:signal]
+    assert_nil event[:stderr]
     assert_nil event[:permanent]
     assert_equal 12, event[:perform_ms]
     assert_equal({ perform_ms: 12, operation_ms: 9 }, event[:timing])
@@ -245,9 +259,9 @@ def register_with(response, **options)
       end
     end
 
-    def failed(code:, cause: nil, signal: nil, error_class: nil, message: nil)
+    def failed(code:, cause: nil, signal: nil, error_class: nil, message: nil, stderr: nil)
       HotCell::Response.failed HotCell::Failure.new(code: code, cause: cause, signal: signal,
-                                                    error_class: error_class, message: message),
+                                                    error_class: error_class, message: message, stderr: stderr),
                                timing: { perform_ms: 1 }
     end
 end