Measured 2026-10-03 (frankuser)

Found while adding str unpacking: def f(w): first, *rest = w failed under HEAP_DEBUG, and so did the same shape over a list parameter, a list literal, a tuple and a call result -- eight of eight shapes inside a def, none at module level. Present at pin v452 and at 79b1a0bca7.

The mechanism is general, so it is worth stating for the next reader: an argument that is a FRESH call result is released by the call site after the call, on the assumption that the callee only read it. A pylib routine that RETURNS ITS ARGUMENT (the mark_* family: pylist_mark_list/tuple/set/ frozenset, pyvar_mark_list) therefore hands back an object the caller is about to free, whenever its argument is fresh. The set comprehension met the same family earlier (bug-n-a-set-comprehension-over-releases-its-own-list). The other mark_* consumers were checked by measurement, inside a def with fresh arguments: set()/frozenset()/tuple() of a call result, of a generator and of a range, set and tuple literals, frozenset of a set literal, list of a set -- byte-identical to CPython under HEAP_DEBUG and flat under the census (25 vs 8 live at 300/3000). The starred unpack was the one site that wrapped a fresh slice.

The variant-source arm now also lists a str, range, deque or user iterable before slicing (a slice of a str is a str, and a range had no slice), and pylen_v answers a range and a deque, which len(x) on an unannotated parameter and the unpack's count both use.

Test: test/test_nilpy_a_starred_unpack_target_inside_a_function_is_not_freed.npy -- byte-identical to CPython under HEAP_DEBUG; census: 56 live after 3000 passes, keep control trips.