Recording fails when you bombard the tracee with SIGCONT

If you apply this patch:

diff --git a/src/test/thread_exit_signal.run b/src/test/thread_exit_signal.run
index f39c19ee6..d022b8b1f 100644
--- a/src/test/thread_exit_signal.run
+++ b/src/test/thread_exit_signal.run
@@ -6,11 +6,13 @@ until grep -q ready record.out; do
   sleep 0
 done
 
+PID=$(pidof $TESTNAME-$nonce)
 for i in $(seq 1 100); do
   sleep 0.001
-  kill -CHLD $(pidof $TESTNAME-$nonce) >& /dev/null
+  kill -CHLD $PID >& /dev/null
+  kill -CONT $PID >& /dev/null
 done
-kill -USR2 $(pidof $TESTNAME-$nonce) >& /dev/null
+kill -USR2 $PID >& /dev/null
 
 # Wait for 'record' to actually terminate. Otherwise we might start
 # replaying before the trace file has been completely written.

you get all sorts of failures. It appears that when a PTRACE_SINGLESTEP races with a SIGCONT the kernel reports PTRACE_EVENT_STOP with a SIGTRAP and those can be reported in all sorts of places.

Start by reproducing the failures with the proposed changes to src/test/thread_exit_signal.run, which repeatedly sends SIGCONT during recording. Trace how PTRACE_SINGLESTEP, SIGCONT, PTRACE_EVENT_STOP, and SIGTRAP are reported, then verify that the test completes reliably without recording failures.