Stronghold consumed Custom Game press: guest execution remains unobserved
The retained run proves host dequeue of a balanced native (543,292) press, followed by an unchanged ChooseType capture. It does not establish execution of the game's WndProc, button hit/press evaluator, or Custom Game callback. No concrete production defect can be inferred from these records. Production and all route/Start guards remain unchanged. This task supplies a tested, bounded read-only observer for those missing boundaries; it does not claim a causal fix or world entry.
Evidence: scratch/runs/20261104-stronghold-consumed-press-cause/. Predecessor: hover recognizer and actual press, commit 7dbf713d61ae53d78bdfded14806e402d47c7860. Prior native callback/source audit remains valid.
What the actual records establish
I personally viewed the existing pre-press review/route-native-boundary-3.png, SHA-256 8af95d757560b25af4972fc5dd76b38625174669797b686cae1f9102080afa70, and final recovered/closed/route-native-boundary-5.png, SHA-256 0d721dbb0a829c3ba5c9eda1645821fd49afd1c49e68ffa44fa35e514f060554, under the predecessor run. Both show ChooseType, Custom Game and its description. The right towers are visibly hovered. The final picture is not the preset shell.
The retained console logs WM_NCHITTEST and WM_LBUTTONDOWN at 08:56:50.502/.503Z, and WM_LBUTTONUP at 08:56:50.859Z, all HWND 0x10001. Down has wParam1; up has wParam0. Queued native coordinates are (543,292), lParam 19137055. The reviewed pre-press boundary is 10688 steps; final 11008 has empty host queue, null active/last input and mask0. The prior audit verifies 172 ACKs / 11008 steps. The first unexpected scene at 08:56:51.164Z is terminal. No later input, preset, Start or retry qualifies this result.
These are actual run records. Replaying the PNGs with modeled ACKs reproduces the controller's terminal rejection, but adds no guest execution evidence.
Where host consumption stops proving progress
The production h.check_input implementation is byte-identical between retained runtime Git 1188e9bf57ef35bd3363fb88b7ca0da442ce696d and this checkout. It calls renderer.takeInput, sets host _lastInputEvent and renderer _activeInputEvent, logs [input], and returns a packed message. An empty poll clears those host fields. Therefore the observed console lines are emitted before the guest receives or dispatches its MSG.
The retained/current WAT message, dispatch and tracing fragments are identical. PeekMessageA can retain hardware input in pending_input_packed, route it to its owner, admit it through the guest's filters or move an excluded event into the posted queue. GetMessageA has its own delivery path. Neither host queue drainage nor a frozen-step ACK says the application subsequently called DispatchMessageA with that event. DispatchMessage's x86 path constructs a WndProc frame and redirects EIP; that is another boundary after its API entry. The production Worker broker captures forwarded API arguments inside the calling Worker because the page instance's registers are not authoritative. This observer follows that same rule; page-register sampling would be insufficient.
Additional original EXE evidence
All VAs below refer to the original preferred-base EXE, not allocated emulator regions. audit.py verifies its full SHA and 15 exact instruction assertions, and regenerates bounded disassembly from existing media.
- Window-class registration at
0x467228supplies WndProc0x4a9810. Its WM_LBUTTONDOWN branch0x4ab9fapasses the lParam coordinates and operation1 to0x4687f0; the WM_LBUTTONUP table target0x4aba46passes operation2. That method stores buffered coordinates at mouse-object offsets+0x1f4/+0x1f6and sets/clears bit0 at+0x1fa. This is guest mouse state, distinct from the renderer's physical mask. No retained dump records it during this press. - Button evaluation
0x4e9620first tests record words+0x36/+0x38, then other record state. It calls rectangle hit-test0x468880at0x4e971e;0x4e9723tests its EAX result and skips on zero. The hit-test uses guest mouse-object coordinates and strict rectangle comparisons. The visible hover does not prove those conditions remained true during press evaluation. Dynamic button dimensions, guest edge state and evaluator outcomes were not retained. - As previously audited, ChooseType object
0x692c38owns descriptor0x5bd098; button2 is record0x5bd138, artwork527. Press callback0x430e10reads its first stack argument. Its ID2 branch at0x430e46pushes scene0x14and calls transition manager0x46e050at0x430e71. The distinct hover callback is0x430f20. A Custom Game label proves neither entry into0x430e10nor argument2. - The manager stores the requested scene at
this+0x18(0xd13770for object0xd13758) at0x46e068, while its current scene isthis+0x0c(0xd13764). Thus even a request is a separate observable from a committed/displayed scene. Neither value was captured in this run. No demo restriction, failed readiness condition, swallowed dispatch or insufficient waiting period is established as the cause.
Tested observer and next boundary
observer.cjs attaches synchronous wrappers to the existing Worker host.log API-entry hook and host.log_block(eip,esp) trace hook. It must be attached to the import object before instantiation, with an exports getter bound to that same instance and translate: self.memUtils.g2w from the existing production memory utility. The pure JS translator reads packed mappings without calling the WAT translator; a WAT translation miss can touch the sentinel or raise a guest exception, so it is deliberately never invoked by this observer. It remains inactive until explicit start() at a qualified frozen boundary. It validates the externally preflighted EXE SHA, preferred base and original instruction signatures. Records include Worker slot, live guest thread ID, stack arguments, EAX/ECX/ESI, current/requested scenes, hover ID, raw button record state and guest mouse fields. API-entry rows copy the actual DispatchMessageA MSG synchronously. Block rows distinguish WndProc, guest mouse update, evaluator, hit-test return, press callback, ID2 branch and scene request. They are observations, not replacements for game calls.
For a separately authorized original run, attach the helper in each relevant Worker's existing instantiation instrumentation; retain all original import forwarding. Before the same normal reviewed gesture, use the existing diagnostic set_api_log(1) and set_trace_esp controls on the owning instances. The smallest first-boundary configuration is callback range 0x430e10..0x430f15, together with DispatchMessage API rows and sample() at existing pre/post ACK boundaries. This directly answers the currently missing callback ID/request question. To collect every supported upstream block in that same one-gesture window, the inclusive range is 0x430e10..0x4e9723. That wider diagnostic range forwards many unrecorded blocks through the original broker and can substantially perturb timing; it is not qualified as a cheap gameplay configuration. Neither configuration was executed here. Do not infer timing or absence from an incomplete trace.
Keep the current initial ChooseType, exact specific hover, personal review/fresh recapture, native mapping, ownership, complete ACK/drain, first-failure terminal and native Start guards unchanged. Observer installation is not a substitute for those prerequisites. End recording at the existing terminal boundary; do not add settling input or a retry. Export each instance's report() while still frozen. Storage is capped at 1024 rows by default, with explicit drop/error reporting; trace arming and Worker coverage remain independently unqualified in reports. An absent callback row alone is not proof that it never ran. detach() restores only wrappers it still owns. No guest setter, guest invocation, forced scene, memory write, counter reset or input occurs inside the observer.
observer.test.cjs passes against original EXE bytes, actual production GuestRpc.createWorkerImports/createMainBroker, actual renderer down/up and the exact retained/current host dequeue body. It verifies synchronous Worker stack capture, original broker forwarding, quiet-log semantics, fully drained host input producing zero guest-boundary proof, callback/hover separation, hit-result zero/one separation, wrong pin/base/signature, unmapped stack, overflow, immutable report, complete memory nonmutation, use of the real pure JS translator with the WAT translator forbidden, and detachment. These are explicitly fixture invocations, with no WASM instantiation or guest execution. The predecessor actual-capture postpress regression also passes, with its ACKs explicitly modeled.
Reproduce offline with python3 scratch/runs/20261104-stronghold-consumed-press-cause/audit.py and node scratch/runs/20261104-stronghold-consumed-press-cause/observer.test.cjs. Only this note is committed; the ignored helper/tests and recovery patches are retained in the run directory. No production/test-tree code changed, so no production code commit or full-build gate applies.
Provenance and stop
source-references.json, exact-git-sha.txt, relevant binary diffs, owned new-file patches and artifact-manifest.json preserve exact inputs. Git remains the source archive. Existing original EXE path is test/binaries/candidates/stronghold-crusader-demo/installed/App_Executables/Stronghold Crusader multiplayer demo.exe, SHA-256 9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f. Existing modules under scratch/runs/20261104-stronghold-spear-delivery-recruit/recovered-modules/ are verified in place: normal cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee, compat 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3. Full original media/runtime path pins remain in the predecessor's retained-module-media-path-sha.json. No media/module copies or source archives were made.
Unresolved boundary: which instance/thread dispatched the original down/up, whether its WndProc updated the guest mouse buffer, and whether button2 reached callback 0x430e10 with argument2 and requested scene20. The historical captures cannot recover these missing observations. No speculative production fix is justified.
No box access/allocation/resume/lifecycle, browser, heavy game CLI, guest execution/input, subworker/successor, HeroesII or deployment occurred. Per owner evidence, bx_v34fdr58 was STOPPED at 09:04:18Z, snapshot 09:04:18.162Z; former expiry 09:18:46.794Z. This task made no live query or state change. Lifecycle custody remains with the coordinator. Single assignment complete; stop after own-file publication.