claudetool: improve codereview

Do a bunch of un-vibecoding and bug fixing.
Unfortunately, lots left.
Get rid of vet; gopls check covers it.

Add testing infrastructure and a bunch of fixtures.
diff --git a/claudetool/testdata/skipped_to_failing.txtar b/claudetool/testdata/skipped_to_failing.txtar
new file mode 100644
index 0000000..50348d2
--- /dev/null
+++ b/claudetool/testdata/skipped_to_failing.txtar
@@ -0,0 +1,33 @@
+Going from a skipped test to a failing test is a regression
+
+-- p_test.go --
+package p
+
+import "testing"
+
+func TestP(t *testing.T) {
+	t.SkipNow()
+}
+
+-- .commit --
+Initial commit with skipped test
+
+-- p_test.go --
+package p
+
+import "testing"
+
+func TestP(t *testing.T) {
+	t.FailNow()
+}
+
+-- .commit --
+Change skipped test to failing test
+
+-- .run_test --
+Test regressions detected between initial commit (3635353dfefb2fa652728318567209ccfc4aba42) and HEAD:
+
+1: sketch.dev.TestP: Was skipped, now failing
+
+
+Please fix before proceeding.