Shadow Man Beenden: modal ownership boundary established; zero writer unobserved
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.
Offline diagnosis of retained run 20261023-shadow-man-settings-observer-repair-runtime
(published b47634c719a289e2f1ab6e5067c0d8c1995c012a). No production repair. The earliest proven incomplete execution boundary is the first synchronous
WM_COMMAND callback abandoned after opening Configuration. The final failure is
proven only as a sampled EIP=0 after ordinary Beenden/EndDialog, with last decoded
block 0x004015d4. Neither an indirect NULL call, a NULL RET, nor an escaped
continuation sentinel is established. The harness message is a classification,
not an instruction trace.
A specific, source-supported candidate is a split between restored caller ESP/EIP and surviving modal-pump globals/stack frame. The final ESP fits that candidate particularly well, but key registers and the final opcode are absent. Treat it as the priority capture target, not an established cause or repair.
Proven observations and their limits
The retained log (historical artifact: shadow-man-null-exit-diagnosis-20261023/retained/session1-fixed-session.log)
contains these API entries (guest addresses as recorded, without assigning
unobserved x86 instructions):
| Entry | HWND/template | API ESP | saved guest return |
|---|---|---|---|
| DialogBoxParamA #172 | template 0x7a, original launcher | 0x074fff50 | 0x00402fd5 |
| EndDialog #173 | 0x10001 | 0x074fff30 | 0x00402937 |
| DialogBoxParamA #176 | template 0x7c, Configuration | 0x074fff20 | 0x00401504 |
| EndDialog #226 | 0x10006, Configuration OK | 0x074fff28 | 0x004022f6 |
| DialogBoxParamA #227 | template 0x7b, new launcher | 0x074fff18 | 0x004015d4 |
| EndDialog #228 | 0x1001c, ordinary Beenden | 0x074fff30 | 0x0040295e |
The first two command injections each end with ABANDONED wndproc, 64 rounds,
EIP 0x07500218 and yield_reason 15: original launcher WM_COMMAND/1024, then
Configuration WM_COMMAND/1. At the second boundary the log already contains
EndDialog(Configuration) and creation of the new launcher. These warnings prove
incomplete synchronous delivery; they do not prove which memory word later
caused zero execution. The after-OK snapshot still has both old dialogs visible
alongside the new launcher. Visibility alone is not guest frame liveness.
Beenden #228 precedes the input receipt for command 3 because dlg-cmd calls
send_message before logging its receipt (test/run.js:8461 in the exact
runtime; locate by the branch text if line numbers change). The first zero sample
is at the same batch 2578996. It reports dbg_prev_eip=0x004015d4,
ESP=0x074fff6c, [0x074fff68]=0, [0x074fff6c]=0 and
[0x074fff70]=0x0040363e. The last word is only a retained stack value; it has not
been identified as an executed return address. No guest ExitProcess was observed. OS -9 is external watchdog termination, not a guest exit code.
The observer JSONL's three DialogBoxParamA records labelled return, each
EAX=1, are host API-handler-exit hooks. The handler redirects to the initial
DLGPROC before the modal guest call finishes. These records do not prove normal
DialogBox/callback completion. See the retained JSONL and donor observer
h.log_api_exit wrapper.
Exact source boundary
src/09c3a-dialog-runtime.wat:33$wnd_send_message_innersaves caller EIP/ESP/GPRs and yield state, pushes a 20-byte synchronous callback frame, and drives bounded recursive$runcalls. At lines 248–299 the 64-round path logs abandonment, checks invocation identity, leaves its sync depth, sets completion false and restores saved caller registers/yields. It does not roll back guest memory or modal-pump globals. This is the earliest proven incomplete callback ownership boundary in this session.src/09a-handlers1-user.wat:2078$handle_DialogBoxParamAsaves the previous modal globals into its consumed argument slots at API ESP+4..20, records its return address globally, and keeps a 24-byte API frame under the initial callback frame. The current pump HWND/proc/return address change immediately.src/09a-handlers1-user.wat:1418$handle_EndDialogmarks the matching active modal ended and yields; it does not itself complete the modal API's return.src/09b-dispatch.wat:961CACA0004 consumes any callback sentinel, and on modal completion reads the previous pump from current ESP+4..20, sets EIP to the captureddlg_ret_addr, and advances ESP by 24. Correct ESP ownership matters independently of correct invocation-token checks.src/09b-dispatch.wat:1331CACA0005 deliberately sets EIP=0 to stop the synchronous nested run. Normally the sender restores caller EIP afterward. Thus the mere presence of a zero writer in this source is not proof that it escaped in this run.src/13-exports.wat:281handles thunk-zone entries before updatingdbg_prev_eipat line 311.src/04-cache.watalso updates it on decoded-block transfers. It identifies decoded-block context, not necessarily the last instruction or last WAT continuation that wrote EIP.test/run.jssamples EIP after work and prints the NULL label if zero without a host exit request. It does not distinguish CALL/RET/JMP/continuation writers.
Discriminating offline control, not guest reproduction
GIT_NO_LAZY_FETCH=1 python3 docs/re-notes/shadow-man-null-exit-diagnosis-20261023/check.py .
Without the optional . argument the retained evidence checks are standalone. Exit 0: exact trace order/frames/stack/real cfg verified, six corrupted-evidence
controls rejected, nine runtime source pins verified. The script is an evidence
validator and conditional frame arithmetic, not an emulator or x86 reproducer.
For the new launcher's recorded API frame F=0x074fff18, coherent completion would consume that frame and produce ESP=0x074fff30. If instead the abandoned sender has restored the original launcher's F=0x074fff50, the same CACA0004 arithmetic produces ESP=0x074fff68. A hypothetical subsequent plain RET would then produce the observed 0x074fff6c; the coherent control would produce 0x074fff34. This is an exact numeric fit to the displaced-frame hypothesis. It is not proof: the retained run does not expose the pre-send/restore ESP, CACA0004 entry ESP, or bytes/terminator of the block at 0x004015d4. No RET opcode is asserted at that address. The zeros sampled at 0x074fff68/6c are after the failure and have no recorded write provenance. Other stack adjustments or zero writers have not been excluded. Six rejecting controls establish validator sensitivity, not production correctness.
Exact missing observation and minimal next action
No further runtime action is authorized by this task. Do not use the partial 479-file save-vfs export as input, repeat restoration/helper construction, or change the configuration bytes to make a test pass.
The smallest useful next observation is to stop at 0x004015d4 before it runs
after Beenden, retaining all GPRs including ESP/EBP, 0x80 stack bytes around ESP,
and 0x40 code bytes there from the actual loaded executable. Existing flags
--trace-at=0x004015d4 --trace-at-limit=1 --trace-at-mem=esp-0x20:128,0x004015d4:64
provide a starting point (confirm the existing argument parser on that revision). This point is a recorded API continuation/last decoded block, not an assumed
faulting opcode. Existing break/small-run control should stop after its first
zero transfer; do not let another long run overwrite the failing frame. If the breakpoint is not reached, retain that fact and capture the thunk writer.
To confirm or reject the ownership candidate, the same tightly bounded diagnostic
needs tuples at the sender's pre-call and restore boundaries and CACA0004's
completion boundary: EIP, ESP, saved old EIP/ESP, dlg_pump_hwnd, dlg_ret_addr,
SHARED_DLG_PUMP_HWND/ENDED, invocation token/root/depth, current thunk metadata,
ESP+0..24 and the original/new frame words. Capture the actual write that first
sets EIP=0: decoded opcode address/bytes and target source word for CALL/RET/JMP,
or the CACA continuation tag and restored caller state. These internal modal
fields are not all exposed by the existing runner; a bounded diagnostic hook
would be necessary if the existing breakpoint/register trace leaves ambiguity. Do not invent fixed allocated WAT addresses to read them.
A captured resumed ESP=0x074fff50 with active frame 0x074fff18, followed by CACA0004 completion at that wrong ESP and a decoded zero transfer, would promote the candidate to a concrete defect. A correct modal ESP would reject this candidate and focus attention on the recorded block's actual return/target or an escaped synchronous sentinel. Only then design a focused positive/rejecting native test and narrow production repair. No speculative source change here.