technical review of Shadow Man synthetic native handler capture

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.

ACCEPT_SCOPED_SYNTHETIC_FULL_MODULE_HANDLER_CAPTURE at 2de825962dc45ac541ee9e0849bff2273eb5f4c8. The retained records and implementation support execution of the actual shipping sender, nested interpreter, constructed x86 callback, DialogBoxIndirectParamA, guest DLGPROC/EndDialog, CACA0004 modal completion and CACA0005 synchronous return, with the actual preload stopping durably at first zero. This accepts a hidden, zero-control synthetic dialog in a headless host through an extracted runner action. It does not qualify the original game or the complete CLI.

The subject is the qualification report and raw evidence at the reviewed commit. All subject evidence below is under its adjacent shadow-man-memory-recovered-handler-qualification-20261023/ directory. The read-only local donor is /home/user/wine-assembly-shadow-man-memory-recovered-handler-qualification-oct23. Coordinator verification of its 72 retained/committed seals is reused, not replayed. This review independently checks decisive raw envelopes and their structure rather than treating native-results.json or verify-native.py as an oracle. Neither subject verifier nor native fixture was executed here.

Exact inputs and independent checks

review.py (historical artifact: shadow-man-native-handler-independent-accept-20261023/review.py) is a new offline structural oracle; its passing results (historical artifact: shadow-man-native-handler-independent-accept-20261023/review-results.json) bind the subject commit, source pins, branch and PIDs. It independently derives both guest instruction streams and relative calls from instruction lengths and absolute targets, checks used export signatures against the accepted inventory, and uses the accepted 269-record 9a3bc9149d58209c layout. GUEST_BASE 0x12000 and THUNK_BASE 0x07112000 give guest thunk base 0x07500000. API IDs are 2875 and 131; four eight-byte descriptors identify CACA0005, CACA0004 and the two real APIs. The template is the disclosed hidden 40×30-unit DLGTEMPLATE with zero controls, menu, class and title. Heap/thread initialization, translation, window table, decoder, interpreter and dialog machinery are actual module exports/handlers.

The oracle initially assumed a void sender return and correctly stopped on the accepted signature mismatch. It was corrected to the shipping (i32,i32,i32,i32) -> i32 signature; the original failed receipt is retained alongside the passing offline-review-correct-signature.receipt.json. No subject evidence or oracle condition was weakened. A resealed copy with modal ESP increased by four is rejected, demonstrating structural checking beyond envelope hashes. The tamper exists in memory only; accepted raw records remain untouched.

The decisive trace is capture.capture.jsonl (native PID 1989364, child exit86):

Boundary Independently checked raw facts
pre-arm selected sequence1, batch50, command3, HWND0x550; prior count/depth0; exact initial registers
sender-pre saved EIP0x404100, ESP0x0740f000; event1, sync depth1
decoded-block EIP0x4015d4, ESP0x0740efec; stack [0x07500000,0x550,0x111,3,0]; code matches constructed callback
modal-frame-consume EIP0x07500008, ESP0x0740efd4; sentinel0xdcbaabcd at ESP−4; return0x4015f2 and five prior-pump zero words; outer sync thunk at ESP+24; HWND65537 ended, shared-ended already consumed; event2
modal-frame-complete ESP advances24 to0x0740efec; EIP0x4015f2, EAX0x77; previous empty pump/shared pump restored; event3
first-zero CACA0005 returns to EIP0, ESP0x0740f000, EAX0x77; event4; stop=true, continuation forbidden
terminal terminal:true, reason:first-zero, exit:86

Child token3/root2 and sync depth1 persist across all four boundary events. The real guest DLGPROC reads its HWND from [esp+4], calls EndDialog(HWND,0x77), returns1 with ret 0x10; the outer guest stores the modal result then returns with ret 0x10. The capture stops inside nested run, before sender restoration; absence of sender-before-restore is expected. The normal control independently covers that later restoration. Ordinary CACA0005 first-zero remains ordinary-or-unknown; zero-writer unproven. No null instruction or zero writer is established by this trace.

Authenticity, action binding and controls

Source inspection, preserved in source-excerpts.json (historical artifact: shadow-man-native-handler-independent-accept-20261023/source-excerpts.json), connects the sender's child allocation/frame construction and real $run to the modal consume/complete events and CACA0005 setting EIP0. EndDialog sets the actual dialog result/ended flags; the continuation restores the saved pump and advances the API frame by24. The host does not supply those events or results. The fixture writes initial guest code, template, thunk descriptors and register inputs; it does not script a sender, allocator, interpreter, callback executor or native event emitter, and never replaces/proxies instance exports.

All three retained *.runner-branch.txt files equal a fresh extraction of the whole pinned shipping dlg-cmd branch: convert its opening } else if to if and append its closing brace. SHA256 is a231512463281994fd1ebe646dba0a62d325000c35d304344b8b70a89e1fd4c5. A one-iteration fixture loop supplies continue semantics. The synthetic renderer record supplies the initial HWND; the unchanged branch resolves it and calls actual we.send_message through the supported selected-send hook. This exercises the resolved-dialog route, not every branch of that action.

The fixture deliberately sets process.argv to the pinned runner and explicit module argument before requiring the real preload. That satisfies a disclosed fixture contract; it is not evidence of launching that CLI. Actual command receipts name fixture.cjs, and observations identify both fixture and runner. The preload enforces PID, runner realpath/hash, exact module bytes, one compile, one bound Module/Instance and no start section; the selected adapter checks instance identity, PID, runner, action, counted sequence, command and resolved HWND. No alternative instance/export facade is passed to it. Constructor and streaming restrictions remain active. Binding is within this fixture process, not an adversarial attestation that argv proves an executable launch.

The actual retain function writes then calls real fsyncSync. The fixture's fsync wrapper invokes the saved real fsync first, then records completion using the separately saved real function. Adapter source orders retained pre-arm receipt before diagnostic mutation and before invoking send. The seven completed fsync observations are one pre-arm receipt, five envelopes and the terminal; the first three completions precede the first dialog host event. Pre-arm count0 and sender-pre event1 further bind the ordering. A failed retain cannot arm or send. The wrapper does not replace capture bytes or fake the fsync result.

The controls discriminate this limited claim:

Host facade and saturated capture budget

The finite facade is adequate for this immediate-EndDialog hidden dialog. Actual successful imports are beep process/thread lifetime acknowledgements, log notifications, dialog_loaded, get_window_rect, invalidate, log_api_exit, service_load_library, set_window_class and destroy_window. The accepted void notification signatures match their no-return implementations. The rectangle import has a void signature and writes four i32 values [0,0,100,80] at its bounded WASM output pointer, not at guest frame/register locations. The beep acknowledgements check nonzero epoch/thread identities; their returned values are dropped by shipping initialization (see beep references (historical artifact: shadow-man-native-handler-independent-accept-20261023/beep-source-references.json)). Neither a beep request nor audio execution is serviced.

The sender's i32.and eagerly calls service_load_library even when yield reason is not5. The facade asserts reason!=5 and returns0, which truthfully means no load serviced. An actual pending DLL load fails; it cannot become a dummy successful DLL import. Other undeclared contracts throw with signature/args. Unused clock/client-size/show-window helpers do not establish renderer/window visibility correctness; show_window's constant acknowledgement would be insufficient for a visible-dialog claim. It was not called in these runs.

All four earlier failures remain explicit in raw normal-*-failure records: beep_process_begin, get_window_rect, invalidate and service_load_library; each exited1 and was reaped. Their fixes supplied the above finite contracts inside the fixture, without changing native exports or success predicates. The verifier's earlier negative-order failure reflected the later real exit fsync; this review independently checks rejection followed by that fsync.

This review imported only the subject's unchanged Python supervisor, redirected its paths to a new owned boundary/ directory, and seeded an explicitly synthetic counter receipt at524288. No actual task counter was reset. Zero-output receipt (historical artifact: shadow-man-native-handler-independent-accept-20261023/boundary/zero-output.receipt.json) passes0 at524288. One-byte receipt (historical artifact: shadow-man-native-handler-independent-accept-20261023/boundary/one-byte.receipt.json) rejects125 with aggregate capture overflow, aggregate524289 and retained0. Both children/groups were fully reaped. This is a fresh concrete boundary probe, not a native fixture rerun or a claim that the negative produced capture data.

Remaining original-game prerequisite

The next capture requires a separately authorized runtime with the complete original Shadow Man fixture and actual runner dependency closure. Preserve this exact diagnostic candidate/module rather than rebuilding from current main. Observe a fresh Configuration/OK/launcher route in the actual game process, including returned prior-action receipts and the visible current state. Count the actual dequeued actions and select the final Beenden command3 intent; do not reuse synthetic sequence1/batch50/HWND0x550 or historical HWND/batch. Bind actual runner/PID, named module hash and single instance, resolve the live dialog at that action, fsync pre-arm context before arm/send, require the 0x4015d4 block and terminal first-zero records, use exclusive output and an external watchdog for synchronous spins. If that run lacks a required record, retain the exact missing-record blocker; synthetic evidence cannot fill it.

Full CLI startup, original game, browser, UI correctness, settings persistence, reopen, Play and gameplay remain unqualified. Configuration/OK returning does not by itself establish any of those. A proven zero writer would require additional origin evidence; ordinary continuation zero is insufficient.