MDK2 grenade transition: predicates and evidence limits — 2026-11-04

The retained run does not establish why the first grenade instruction was absent. It establishes visible dummy breakup/disappearance and no instruction in the sampled frames. The specific unverified transition predicate is l1_r1.TutFuncs["state 4"].tasks >= 2, credited by death-animation callbacks, with the tutorial dispatcher still active. No runtime defect or required extra player action before state 5 is proved. No repair is justified by this evidence.

This offline review follows the published observation note at d6f10e18c. Analysis baseline is c66edc193102d3bde2588661833501a37e7faa6e; original execution source is 1329b1ba6eb00144ff77d4cd92a8513d4fc953ad, canonical module cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee. The predecessor's accepted-to-execution-source.patch remains in its run. New analysis, reference hashes, relevant binary diff and owned new-file patches are in scratch/runs/20261104-mdk2-grenade-transition-cause/ (ignored).

Original media was read at test/binaries/candidates/mdk2-demo/installed/mdk2demo/data/scripts.zip, SHA-256 cfdae6e7d02b1823a59d0b85aa8eaca2f057dbb27e7b86d9a7199fc069f642d8. ZIP method 10 was decoded only in memory with the existing tools/mpq.js explode; decoded level1.lua SHA-256 is 7d7de7b115b6fb1deb6290791c222e59c2726c607324095d3ba4c6136d9c08fe. No media was changed or copied, and no game, browser or sandbox operation ran.

The exact script chain is:

Original level1.lua lines Required behavior
2217–2240 State 4 initializes tasks=0, statetasks=2; both dummies receive r1DummyOnDamage and r1DummyOnAnimLoop.
4554–4571 Each damage callback increments gob.hitcount. At hitcount >= 10 and not gob.dead, set dead=1 and call mdkDecoyDeeeeeeeeeeeestroy. This function does not increment tutorial tasks.
4544–4550, 3132–3137 r1DummyOnAnimLoop(gob, animid) calls TaskInc("state 4") only when animid == ANIM_DIE. TaskInc adds one to that state's counter. There is no per-dummy deduplication here; the literal test is a total of at least two credits, normally one per dummy.
4859–4876, 4883–4900 Initialization installs Level.TutorialOnUpdate. On an update, Save.intutorial must be truthy, current tasks must reach statetasks, and not l1_r1.holdstate must hold. The dispatcher calls EndState, increments numeric state, then calls the new state's InitState in the same update. scriptrunning is not a dispatcher gate.
2254–2258 State 4's EndState plays ANIM_DIE on arrows 06/07 and assigns statename="state 5". It does not wait for arrow animations to finish.
2265–2285 State 5 installs the explosion event, starts a script: preload l1_tut8.wav, display strings.l1tut_dr21, wait for mdkDialogIsDone, then clear scriptrunning and credit state 5. It also creates OBJ_GRENADEITEM at l1_r1grengen and spawns arrow 09.

There is no facing, distance, jump, sniper toggle, extra shot, pickup, or grenade explosion condition between state 4's completed counter and that first state 5 panel. The <4 center-arrow distance test belongs to state 2 (2178–2182), before the dummy stage, and is not checked again here.

Pickup is a later native action: state 5 requires two tasks (2263–2264). One comes from its completed dialogue script. r1SGOnPickedUp, with id==OBJ_GRENADEITEM, removes four items, kills arrow 09, clears the grenade reference and starts a one-second timer (4584–4590); timer state 2 credits state 5 (3012–3013). Only then can state 6 display strings.l1tut_dr22, the throwing instruction (2295–2308). Pickup cannot explain a missing first state 5 panel. No retained pickup callback or inventory observation proves that action.

The reviewed existing PNGs review/targets-settle-35488.png and review/grenade-center-37360.png show absent dummies, arrows visible in the former and absent in the latter, health 100, and no dialogue panel. Their JSON companions contain renderer/module/input observations, no Lua state or animation callback trace. The 643 recorded commands include 591 step requests totaling 37,360 steps and 22 trusted DOM key/button events. After the initial ArrowUp hold (commands 565–567), the gameplay controls are mouse aiming and two left-button firing holds; the final "center" action (639) is mouse motion to (440,500), not walking to the room center or grenade pickup. There are no post-shooting movement commands. No Space/sniper-toggle event is recorded.

Thus neither ten callbacks per target, two ANIM_DIE callbacks, the task total, Save.intutorial, holdstate, state 5 initialization nor its dialogue script cursor was measured. Advancing another 1,872 steps does not measure any of them, and sampled screenshots cannot exclude a panel appearing between samples. Arrow loss is consistent with state 4 ending but is not unique proof: the native KillTutorial path also kills those arrows/dummies, clears Save.intutorial, removes OnUpdate, and stops a running script (2854–2884, 2973–2979). Its sniper-damage trigger at snipe1_01.OnDamage (3114–3123) was not demonstrated in this run. Neither cancellation nor a missed death callback should be promoted from a possibility to the cause.

One concrete next ordinary-control attempt, if separately authorized: release fire after both breakups, then use a short released ArrowUp pulse toward the front of the room while keeping the floor and grenade dispenser in view; stop on a visible grenade/arrow or pickup and record the instruction before and after. This tests the unperformed approach/pickup route for the later throwing instruction; it is not a claimed fix for absent state 5 initialization. Do not re-shoot empty positions or equate another disappearance with a task credit.

Validation: the offline checker passed 16 original-line assertions, hashed 21 retained references, and checked command/step counts and final input release. Its initial relative module-path typo was corrected before the passing run. No production source changed, so no production build or gameplay regression was run. No tutorial completion, grenade use, save or runtime repair is claimed. No box was assigned or accessed in this task. The predecessor's last runtime handoff was bx_kbzzavgf, RUNNING/API idle, expiry 2026-11-04T03:06:56.554Z; the coordinator subsequently recorded STOPPED at 03:00:40.620Z on the messageboard. That is reported history, not a fresh status query. Lifecycle ownership remains with the coordinator. This task ends here.