Native button delivery, 17 November 2026

Evidence: scratch/runs/20261117-disciples2-native-button-delivery/.

The authored real-ticks slow USER pump reproduces the missing-DOWN mechanism. The renderer handler enqueues movement, hit-test and DOWN. The old guard reports CONSUMED immediately on that handler's return. Independent release then purges the still-undelivered tagged DOWN and queues UP. The guest's ordinary wndproc sets its public title to UP WITHOUT DOWN. Movement coalescing is not needed for this reproduction. No original endpoint was reused.

The repair keeps host application separate from native consumption. Production CLI tickets remain DELIVERY_PENDING after the handler returns and acknowledge CONSUMED only when the exact resulting edge returns through check_input. Receipts explicitly identify native-input-return, retain host appliedAt, and include the actual message metadata. This boundary does not assert guest DispatchMessage/wndproc execution or a game consequence. The Disciples controller refuses a host-handler receipt.

Release continues to cancel a pending edge and purge its queued effects before issuing independent UP. It never preserves DOWN for delivery after UP. The immutable wall/monotonic deadline, nonce tombstones and original release references are unchanged. The post-handler public USER context is pinned until native delivery, so target, focus, modal, generation or retirement changes cancel the edge. The pre-handler reviewed context remains mandatory.

The authored normal and slow PE fixtures use the ordinary public message pump, renderer, native module and actual Node guest Worker helper. Their input consumer is the main VM. An ordinary GetTickCount loop under existing --real-ticks makes the slow pump genuinely wait; Sleep alone would measure scheduler batch time. Both positive fixtures show DOWN before independent UP and the guest caption UP after DOWN. A separately armed process owns release. Early release cancels without consumption and without a later DOWN. Production guard plus actual renderer-queue units cover expiry, focus/target changes, retirement, owner replacement, inspection/effect UNKNOWN, missing edges, lookalike events and nonce replay. Current input-owner, reviewed-context, independent-release and per-slot metadata tests pass.

The broader test-input-consumption-state.js extraction fixture fails with cliObservation is not defined on both HEAD and the delivery patch. Its baseline failure is retained and is not counted as a pass. Early experimental cross-thread PE shapes that did not establish guest handling are retained as failed experiments and are not qualification.

The canonical flock build and focused current native input/context tests passed. Material commit 91285ebf1 and correction b954d09da were published before the single original. The correction removed an accidentally included concurrent VB snapshot dependency, then restored that exact line as its owner's uncommitted work. The runtime uses Git b954d09daff03a15659b2a96afa823fe0d1af5be, no owned patches and no foreign WIP. The sealed source SHA-256 is 30e697d2a04e85a24c8c663e03392ee6e54dd1c50850b1de1eafa977655e9e99; the independently compiled module is 1360defeb5faa36d2007b4c4c3b5651e8c11e0a9fd80c4b07661b220e09c75b1. The same holder qualified normal native DOWN at 50 ms and slow DOWN at 294 ms, both followed by independent UP and the guest's UP after DOWN. Early release produced CANCELLED, no consumedAt/nativeMessage, UP without DOWN and no late DOWN. Native, release, watchdog and supervisor actual waits completed zero with both EOFs. These fixtures use a main-VM input consumer and an actual guest Node Worker helper; they do not qualify a Worker-owned window procedure.

The one fresh original verified all 128 unchanged media files before and after. No f.sg, imported save, config, EXE, clock, hold or deadline change was used. Its independently prearmed native/watchdog limits remained 3600/3620 seconds with a 300-second Quit reserve. The personally viewed complete main-menu frame authorized one Single Player click: actual native DOWN then UP, confirmation 87.845 ms and independent UP 134.180 ms. A later full frame showed the Single Player menu. One newly reviewed New Quest click confirmed in 37.407 ms and released in 120.969 ms. original/quest-choice.png then visibly showed the selected Search for Timmoria scenario, map size 48×48. This is real original progress beyond the previous missing-DOWN menu blocker, not map gameplay.

The freshly reviewed forward-arrow click at canvas (947,682) was CANCELLED before its handler ran: handlerApplied:false, consumed:false, no native message, reason reviewed native scene changed. Independent release still acknowledged. The controller latched a terminal failure; no later live request, capture, rebind, replay or second original followed. Offline replay of the published validator over the retained public snapshot deterministically reports pointer outside reviewed target: its 800×600 native target rectangle is compared directly with the 1024×712 canvas click. The CLI passes the parsed canvas coordinates to this validator; the renderer maps them later. This is a demonstrated coordinate-space blocker. The live wrapper discards the detailed validation error, so no claim is made that its generic reason proves this was the only mismatch at that instant. Fix review-space conversion using the actual public viewport transform and preserve target containment; do not bypass it or guess native coordinates for another original.

Owned host cancellation closed the original. Native and watchdog actual waits were zero with both EOFs; this is not native Quit. The holder's module hash was verified immediately before normal closure; holder and independent holder wait also completed normally. No map/hero/town/turn action or native Quit is claimed. Only selected complete screenshots remain, with hashes for omitted observations. No box was created: ID null, state NOT_CREATED, expiry null, stop not applicable.

The concurrent VB authored dialog fixture exposed an additional unsupported synchronous native BUTTON path: successful dialog_route_mouse_screen can return without a queued check_input edge. The new guard truthfully closes it UNKNOWN rather than inventing consumption. The VB owner received the guard and runner ownership for a separately witnessed synchronous native-return boundary; that work is excluded from this sealed runtime and is not qualified here.