This is a test showing SF bug 422177. It won't trigger until I check in

another change (to test_import.py, which simply imports the new file).  I'm
checking this piece in now, though, to make it easier to distribute a patch
for x-platform checking.

Lib/test/double_const.py

+from test_support import TestFailed
+
+# A test for SF bug 422177:  manifest float constants varied way too much in
+# precision depending on whether Python was loading a module for the first
+# time, or reloading it from a precompiled .pyc.  The "expected" failure
+# mode is that when test_import imports this after all .pyc files have been
+# erased, it passes, but when test_import imports this from
+# double_const.pyc, it fails.  This indicates a woeful loss of precision in
+# the marshal format for doubles.  It's also possible that repr() doesn't
+# produce enough digits to get reasonable precision for this box.
+
+PI    = 3.14159265358979324
+TWOPI = 6.28318530717958648
+
+PI_str    = "3.14159265358979324"
+TWOPI_str = "6.28318530717958648"
+
+# Verify that the double x is within a few bits of eval(x_str).
+def check_ok(x, x_str):
+    assert x > 0.0
+    x2 = eval(x_str)
+    assert x2 > 0.0
+    diff = abs(x - x2)
+    # If diff is no larger than 3 ULP (wrt x2), then diff/8 is no larger
+    # than 0.375 ULP, so adding diff/8 to x2 should have no effect.
+    if x2 + (diff / 8.) != x2:
+        raise TestFailed("Manifest const %s lost too much precision " % x_str)
+
+check_ok(PI, PI_str)
+check_ok(TWOPI, TWOPI_str)