Stronghold setup render completion: original contract, offline 2026-11-04

The repeated bounded-arrival failures are observations inside original drawing, not evidence of completed setup. The current helper recognizes two exact partial images; each newly encountered partial image fails closed. Replacing that list with a general fade recognizer is not justified by the retained evidence: the observer records scene ownership and setup return, but omits the original fade amount, direction, timing and draw/presentation completion. No helper or test implementation changes are made. This investigation supplies the original cause and the specific missing next observation, rather than another whitelist.

Entry reference: d5d2e95aaf2d7b05a36e28a9f1e727d3322410c8, predecessor report, original-game notes, and the retained runs below. Evidence: scratch/runs/20261104-stronghold-setup-render-completion/. All new work is offline: static original-EXE disassembly, retained PNG/report inspection, pure arithmetic checks, mock transport tests and the required build. No browser, original executable/game runtime, box, heavy runtime lock, worker, deployment, asset/clock/configuration change or new gameplay run.

Original native contract

Addresses below are original guest VAs, verified against the existing executable: test/binaries/candidates/stronghold-crusader-demo/installed/App_Executables/Stronghold Crusader multiplayer demo.exe, 2,699,264 bytes, SHA-256 9024c533095f2227b0105b7e8a9e4cf28ec4fb206b165629787b2c97cf99f13f. The audit pins 23 instruction byte sequences; disassembly is retained in the run.

The transition also has an input-feedback consequence. While advancing an active fade, 0x429321 sets [0xd138b0]=1. Widget drawing at 0x4e9ee0–0x4e9ef3 uses this to clear hover feedback [0x995674]. The main loop resets [0xd138b0] at 0x576c56; a zero sampled just after that reset does not prove fade completion. Preset hover 0x42d850 needs mode 99 and nonzero hover feedback to set description [0x69332c]. Activation 0x42d880 independently needs mode 99 and no pending preset [0x69165c]; choice 1 queues the pending action and clears native press edge [0x1bd26a4]. Updater 0x42db40 checks the native edge before applying through 0x42e490. Draw calls the updater at 0x429b1b only under the mode-99 gate. Scene 20, callback return, a drained host input queue, visible text or a description alone therefore supplies no input permission. Fade completion is a visual transition condition, not proof of those independent activation/application preconditions.

What the actual captures prove

Both PNGs were personally viewed. Both show partial/translucent setup panels over the shield/swords background. The later report's Human/Lord, Deathmatch, 10000/10000 and CrusaderDemo labels are clearer; this does not establish a monotonic fade sequence across different native histories.

Retained run PNG / report under its browser/ Final native boundaries / PC Remaining to 50M
20261104-stronghold-quiet-rpc-original-gameplay route-native-boundary-8.png / preset-attachment-80.json 40,092,338 / 0x44f7d8 9,907,662
20261104-stronghold-setup-transition-onward route-native-boundary-10.png / preset-attachment-133.json 39,098,384 / 0x45550a 10,901,616

PNG SHA-256 respectively: f48150b6dea1489b67978949eadbd702b8865985648ce57a62e12297778a652d, 6e73425ddafb362c96280e2f174e7737ded8b3d8b101d0c2c74a5b48bda93d0c. Decoded report-content SHA-256 respectively: 6b96a5eff33886a7952cb7f3814e8943707415d8f39912966e99eafcf1563107, bca28c6137ab143ffb841ce873754c94a6c887c505e12b7c89c47bb8a0642495. Reports are read through the existing retained/gzip mapping; nothing is copied.

The audit binds each capture to its report's own epoch, owner slot 0/endpoint 1, step count, SHA, monotonic continuous boundary history and preceding ACK. Both have setup-return followed by callback-return and custom-path completion; in the onward run those occur at 22,783,070 and 22,789,897 boundaries. Captures have scene/current request 20, mode 99, manager4 0xffffffff, saved 0; manager8 is 142643/144135. Their final ACKs have no queued/active/last input and mask zero. These are useful ownership/release facts. They contain none of the fade/time fields above, no complete draw return and no associated presentation fence. The final PCs are inside different pixel loops: 0x44f7d8 in RGB565 lookup rendering and 0x45550a in sprite/RLE copy (mov ax,[esi]). Neither is a completed setup-frame witness. A host ACK/capture can expose ongoing drawing.

Both full retained reports are already command-failed, with fault null and pending empty. Terminal status still prevents another command. They must not be turned into live prefixes or used as permission to resume. Original known scene recognition retains precedence; the two partial hashes grant only setup-transition-observation, never setup readiness or preset/Start input.

Budget result and next observable condition

It is undetermined whether the unchanged 50M-boundary/360-second continuous history can reach complete setup. The retained record establishes neither success within that bound nor impossibility. The isolated initialized 800×600 render fragment's 1,441,804-boundary cost is not a bound on the actual remaining composite draws, refresh work, timestamp gaps or input readiness. Six or seven such fragment costs fitting the remaining space proves nothing about actual completion. Frozen 16-ms host steps and ACK counts do not reveal the two native timeGetTime samples per draw. No remaining-headroom argument can replace those samples. The clock must remain original.

The concrete next condition, in a separately authorized original attempt, is a same-owner, same-epoch, continuously counted observation of actual draw entry, fade update, timestamp refresh and draw return. Read only [0x692e18], [0x690e54], [0x6921b8], [0x175f9fc], [0xd138b0] at those points, together with scene/request/mode and existing input/pending state. Record entry and late timeGetTime results so the discarded-residual mechanism can be confirmed or ruled out. Observe the initialized refresh clearing, subsequent negative fade updates, then fade=0/direction=0 followed by the complete draw return 0x42a20e. Associate the resulting normal presentation and real captured pixels with that same draw; the precise presentation fence still needs mapping. A zero input-suppression flag before drawing or a fade value sampled midway through drawing is insufficient. Preserve capture/ACK ordering, no queued input, original clock/assets/config, first fault and all existing caps. If that condition does not occur before the cap, stop with the actual last phase; do not restart counting, increase caps or assume one more frame will finish.

Only those paired native observations and independently reviewed real partial and complete pixels could justify a bounded transition observation predicate. It would still end at observation, with unrelated/mutated scenes, stale or foreign ownership, terminal histories, wrong temporal order and exhausted budgets rejecting. Preset readiness/application would need its own existing review and native checks. This task provides no runtime authorization.

Validation and preservation

audit.cjs checks the actual two partial captures against the original unknown classifier and current observation-only helper; 14 one-pixel mutations and three unrelated actual scenes reject. Twenty-four report mutations cover owner/epoch, fault/pending/terminal status, mismatched pixels/steps/epoch, nonmonotonic history and missing setup return; two actual terminal enqueue attempts reject (43 negative controls total). These are evidence-audit controls, not a proposed general recognizer. Pure arithmetic checks model the pinned signed multiply/shift: 1,000 alternating 16/32-ms gaps retain fade=32; 32 gaps of 40 ms reach zero. Those are synthetic mathematical controls, not invented native traces. Initial audit byte-pin typos were corrected against disassembly; the final audit passes and the earlier failure log is retained.

test/test-stronghold-local-runtime.js passes, including its composed entry, independent spawn prohibition, zero Chrome spawns, private transport/flock, terminal/ownership/input/cap/readiness and retained partial-image regressions. flock /tmp/wine-build.lock bash tools/build.sh passes. No native initializer or game-execution test was run. Exact Git SHAs, relevant binary diff and this new file's patch are in the run; existing media/modules are referenced by path and SHA only. No source snapshot, tar, media copy or module copy was made. Box ID/state/expiry: NA. Coordinator retains runtime shutdown ownership.