t/stress: add a test that usually triggers a bug using 950-curse.

It looks like we're updating the stamp for t/countall while another
task is replacing the file, which suggests a race condition in our
state management database.

t/.gitignore

@@ -1,2 +1,3 @@
 /broken
 /shellfile
+/stress.log

t/950-curse/all.do

@@ -1,4 +1,4 @@
-rm -f *.n[012]
+rm -f *.n[012] countall
 
 . ./check-1.sh
 redo-ifchange 1.n0 2.n0 3.n0

t/all.do

@@ -29,3 +29,6 @@ sed 's/\.do$//' | {
         redo "$d"
     done
 }
+
+# if all that worked run a repeated stress test to look for races
+redo stress

t/clean.do

@@ -2,5 +2,5 @@
 sed 's/\.do$//' |
 xargs redo
 
-rm -f broken shellfile *~ .*~
+rm -f broken shellfile *~ .*~ stress.log
 rm -rf 'space home dir'

t/stress.do

@@ -0,0 +1,10 @@
+rm -f $1.log
+exec >$1.log
+
+# This test twiddles the same files over and over, and seems to trigger race conditions
+# in redo if run repeatedly with a large redo -j.
+for d in $(seq 25); do
+    echo "stress test: cycle $d" >&2
+    ./flush-cache 2>&1
+    redo 950-curse/all 2>&1 || { rv=$?; echo "stress test: log is $1.log" >&2; exit $rv; }
+done