If you allocate a block with pymalloc, it will be carved out of a larger allocation. If you forget to free that memory, Valgrind doesn’t notice because it is only tracking the larger block that pymalloc allocated (which is properly referenced by pymalloc internals).

By bypassing the pymalloc code, each of the allocations is tracked separately by Valgrind and the leak is evident. An alternative approach to the problem would have been to annotate the pymalloc code with Valgrind’s memory pools client API so that it knows what Python thinks is going on inside the larger blocks it allocates. This is a fair bit more work, and I don’t know if it is worth the effort (it’d probably be useful for debugging pymalloc itself though).