Give a script none for an object that no longer exists, instead of crashing - #553

A dedicated server on build-337 crashed during a CTF game on the Custom Edition map coldsnap_20th_xl_rc5, with about 12 players in:

EXCEPTION assert in c:\halo\SOURCE\memory\data.c,#412: object index #449 (0xe43001c1) is unused or changed (release build)
segmentation fault (signal 11, code 1) at 0x8

Symbolising the release binary gives object_get_and_verify_type ← scenario_trigger_volume_test_object ← scenario_trigger_volume_test_object_evaluate ← hs_thread_main. The map's script ran volume_test_object on a Warthog that had already been deleted. A script keeps an object's index after the object is gone, in a global or an object list, and nothing between the script and object_get checked it. In a release build datum_get reports the stale index and returns NULL, and the read through that NULL faults.

The fix

The script runtime now turns such a value into none, which is already what a script gets for a named object that doesn't exist.

• hs_cast checks every global read and every function's or script's result, before casting, so no list is made of a deleted object.
• hs_arguments_evaluate checks the arguments again just before the call, because a later argument's static script can sleep after an earlier one was evaluated.
• Object lists: a list that has been collected counts as gone too, and walking a list passes over deleted objects. list_count still counts every entry, as it always has.
• unit_kill, unit_kill_silent and scenery_get_animation_time crashed on none, so they now do nothing for it. The other script functions with object parameters were searched for a check of none, and those without one were read.

Nothing the machines send each other changes, so the network version stays. Scripts need not agree between machines: only the host runs them in co-op, and in other games the host decides what they do to the game.

Testing

• New harness test tools/harness/tests/test_script_stale_objects.py, on the real runtime and data arrays, with negative controls: 42 pass.
• Locally on Windows, the full harness has the same 631 failures as testing (Linux-only tests) and no new ones. The four changed source files compile for Windows; CI builds the other platforms.
• Not run on the real map. The crash depends on what happens to that Warthog in play, so it hasn't been reproduced in a game, only in the harness.

🤖 Generated with Claude Code