Shadow Man original Beenden: first-zero causality, 2026-10-23

Historical source last updated, 2026-10-23. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Verdict: the original capture stops inside a synchronous send while an already-existing outer modal completion is pending. The accepted synthetic trace exercises a different nesting order. The immediate zero writer remains unobserved. The original remains rejected by its unchanged required-block/modal oracle. This analysis establishes the earliest concrete differences and the exact unresolved native boundary; it does not certify a normal return, guest exit, repaired game, or successful original capture.

No packaging/native acceptance review or execution was repeated. The source candidate is historical b41a7c5f4b752d1f0bdab1141a11d4fed4dc6f9a; exact Git blobs, complete file hashes and byte counts are in source-identities.json (historical artifact: shadow-man-original-first-zero-cause-20261023/source-identities.json). These are candidate files, never substituted main files. Strict read-only ls-remote still returned main 32c0b9bcf4cbaeee3102a5944ca905ca9298169a, recorded in remote-context.receipt.json (historical artifact: shadow-man-original-first-zero-cause-20261023/remote-context.receipt.json).

Earliest differences, before final Beenden. The earliest retained abnormal original transition is Configuration sequence 1, batch 5585, HWND 0x10001: [sync] ABANDONED ... wParam=0x400 ... at 0x07500218 after 64 rounds (yield_reason=15). OK sequence 2, batch 14859, HWND 0x10006, repeats this for command 1, after creating launcher 0x1001c. Both host commands return zero status. Their callback completion is explicitly absent, not implied by the host response. Raw Configuration dispatch (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/configuration-dispatch.json), OK dispatch (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/ok-dispatch.json), and session summary (historical artifact: shadow-man-original-first-zero-cause-20261023/original/session-summary.json) retain these warnings. Candidate dialog runtime (historical artifact: shadow-man-original-first-zero-cause-20261023/09c3a-dialog-runtime.wat), lines 227–288, bounds recursive execution, emits that warning, leaves the synchronous child and restores caller registers even on abandonment; it does not restore all modal globals. CLI run.js (historical artifact: shadow-man-original-first-zero-cause-20261023/actual-source/test/run.js), lines 3652–3666, decodes the warning. Thus there is an actual earlier abandoned-callback history, but no captured causal chain from it to the final zero writer.

At the first aligned diagnostic boundary, pre-arm, the original and synthetic already have materially different state:

Retained field Original actual CLI Accepted synthetic
Selected action seq 3, batch 24720, cmd 3, HWND 0x1001c seq 1, batch 50, cmd 3, HWND 0x550
Prior actions Config and OK returned:true, both abandoned in raw logs none
EIP / ESP 0x07500218 / 0x074fff50 0x404100 / 0x0740f000
pump / shared pump 0x1001c / 0x1001c 0 / 0
modal return address 0x4015d4 0
ended / shared ended / depth / event count 0 / 0 / 0 / 0 0 / 0 / 0 / 0
bytes at 0x4015d4 c3 a1 c0 dd 40 00 ... (RET first) 68 00 00 00 00 ... (fixture callback opening a modal)

Sources: original null-capture.jsonl (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/null-capture.jsonl), lines 1–3; synthetic capture.capture.jsonl (historical artifact: shadow-man-original-first-zero-cause-20261023/synthetic/capture.capture.jsonl), lines 1–6; synthetic observations (historical artifact: shadow-man-original-first-zero-cause-20261023/synthetic/capture.observations.jsonl), initialized-before-action, explicitly identifies callback 0x4015d4, synthetic code and DlgProc/template. In the original, native dlg_ret_addr identifies that same VA as the existing modal's continuation; it is not evidence that the injected Beenden callback itself starts there. Candidate USER handler source (historical artifact: shadow-man-original-first-zero-cause-20261023/09a-handlers1-user.wat), lines 2201–2217, saves this return address and retains a separate modal frame. Same module identity does not mean same guest bytes or same callback nesting.

Original pre-arm → sender → zero, through the exact hook. Final intent (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/final-intent.json) and state snapshots before (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/launcher-before.json), Configuration (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/configuration.json), after OK (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/launcher-after-ok.json) establish observed selection; the selected pre-arm receipt binds actual PID 28583, action sequence 3, resolved live HWND 0x1001c. Old launcher/configuration records remain visible in the snapshots, so those alone cannot certify valid modal-frame ownership. CLI run.js lines 8423–8472 resolves the actual dialog and wraps we.send_message(dlg,0x111,3,0); adapter (historical artifact: shadow-man-original-first-zero-cause-20261023/actual-source/tools/shadow-man-null-capture.cjs) lines 97–112 checks actual instance/PID/action/selection and arms immediately before that send. Preload (historical artifact: shadow-man-original-first-zero-cause-20261023/actual-source/tools/shadow-man-null-preload.cjs) lines 30–32 and 53–59 fsyncs records and binds log_eip to this instance. Sender-pre is therefore positive evidence that arming occurred before the actual native send, rather than after the missed block.

The original sender-pre has saved EIP 0x07500218, saved ESP 0x074fff50, token 13/root 2/depth 1/eventCount 1. First-zero has EIP 0, the same ESP/token/root/depth, count 2, ended/sharedEnded both 1, and unchanged pump 0x1001c and return address 0x4015d4. The residual word at ESP−20 is 0x07500210, consistent with a consumed synchronous return frame, but residual stack bytes do not prove that word was executed. The record's savedEip/savedEsp zeros are hardcoded arguments of the zero event, not a saved zero caller (candidate exports (historical artifact: shadow-man-original-first-zero-cause-20261023/13-exports.wat), lines 10–32 and 102–106).

The only armed native diagnostic events are sender-pre and first-zero: count 1 → 2. No modal-consume/complete event was silently filtered by the address selector: these have negative tags handled independently of the positive block address (adapter lines 20–46). No overflow/record quota failure occurred. The block observer is a block-entry observer, not an instruction-by-instruction history. Candidate exports lines 2611–2615, 2544–2575 and 347–363 enable range tracing and disable chaining/tier shortcuts via debug guards. The evidence does not demonstrate a dropped observer event, wrong selected process or late arm. It establishes absence of the required instrumented block entry and modal frame events before the stop, without proving every guest instruction that ran.

Why the two event orders can differ. Candidate exports lines 4659–4676 calls $wnd_send_message; dialog runtime (historical artifact: shadow-man-original-first-zero-cause-20261023/09c3a-dialog-runtime.wat) lines 176–208 saves caller EIP/ESP, enters the synchronous child, emits sender-pre before building its frame, pushes $sync_msg_ret_thunk and enters the resolved WndProc. Candidate dispatch (historical artifact: shadow-man-original-first-zero-cause-20261023/09b-dispatch.wat) lines 1332–1337 implements CACA0005 by explicitly setting EIP=0 to stop this nested run. Only after that does the sender normally emit before-restore and restore the outer caller. The armed observer instead exits the host process at that first zero; adapter lines 38–46 and preload line 32 explain OS exit 86 and why sender-before-restore is absent. Beenden dispatch (historical artifact: shadow-man-original-first-zero-cause-20261023/original/original-evidence/beenden-dispatch.json) consequently reports a socket hangup, not a completed guest callback result.

By contrast, CACA0004 (dispatch lines 961–1059) owns modal frame completion: it consumes sharedEnded, destroys/restores the modal frame, emits consume and complete, restores EIP from dlg_ret_addr, and clears ended. EndDialog (USER source lines 1434–1460) records ended/result and defers teardown to this pump; it does not itself jump to 0x4015d4. The original's ended/sharedEnded 1 with pump/return unchanged is consistent with a pending outer modal, whereas synthetic first-zero has pump/return/ended/sharedEnded all zero after its callback-created modal completed. The synthetic can complete a modal inside its synchronous callback before reaching its callback-return zero; that is not the original's already-active-modal arrangement.

This is the concrete causal limit on the required ordering: if original Beenden returns normally to CACA0005, its nested stop precedes resumption of the outer CACA0004 and thus precedes that pump's return to 0x4015d4. Requiring those outer events before that nested zero cannot establish original modal completion. The accepted synthetic path does not resolve this mismatch. The trace nevertheless does not show which native branch actually wrote the original zero, so the conditional cannot be promoted to a proven normal-return verdict.

Proven versus unresolved. There is a preexisting active modal, a concrete abandoned Config/OK history, and a selected synchronous send ending at an unproven zero before the captured modal frame was consumed. That is stronger than merely listing missing events. There is no evidence of observer event loss. A genuinely unexpected control transfer to zero, or an abandoned/stale modal frame contributing to it, remains possible: neither callback entry $wp, the zero-producing predecessor/marker, nor full modal-frame provenance is retained. Token/root/depth equality and balanced ESP do not identify the zero writer. Conversely, absence of the outer continuation before a synchronous stop is not proof that Beenden bypassed its intended guest handler or that the game can never complete the modal. No production repair is justified yet; no oracle was weakened.

ONE minimal next capture requirement, for a coordinator-assigned task only: retain a single bounded native provenance record for the immediate transfer that produces the first zero, in the same selected original actual-CLI path, while keeping the existing first-zero stop and required oracle. It must identify writer site/kind, pre-transfer EIP and ESP, post-transfer EIP/ESP, active thunk guest address/metadata marker when applicable (bind actual sync_msg_ret_thunk and dlg_loop_thunk values), and consumed return word/address for a guest RET, with token/root/depth and the existing modal boundary fields. Record before mutation/zero stop so a native CACA0005 assignment is distinguishable from guest RET-to-zero or another writer; a stale dbg_prev_eip alone is insufficient because thunk dispatch precedes ordinary block tracing. This one provenance observation answers the current fork: intended nested sentinel with outer completion still pending versus an unexpected zero transfer. It does not claim to validate every abandoned modal frame. No continuation past zero, general harness expansion, guest run or instrumentation implementation is authorized by this report.