NFS3 setup: foreign-thread window dispatch explains the zero-EIP call

Two emulator ownership errors are demonstrated and repaired. The native device chooser and setup-to-race route remain unqualified: after the first repair the probe reaches ordinary teardown, exposing the second error. Both repairs pass focused regressions and a locked build. Final native-browser qualification was held when shared /home free space fell below the task's 5 GB floor. No runtime box was allocated (ID/state/expiry/stop receipt: null / NOT ALLOCATED / null / null). No game input, renderer choice, guest patch, DLL alias or state forcing was used.

Evidence: scratch/runs/20261108-nfs3-setup-zero-eip-cause/. Git is the source archive: base-commit.txt, nc-only-source.patch and relevant-source.patch identify the source variants; fetched modules and browser response hashes pin what actually executed. media-verification.json freshly verifies the existing 133 media files and manifest without copying them. The released predecessor is nfs3-loadlibrary-missing-vendor-20261108.md.

First wrong context: NC work retrieved by the wrong UI thread

The clean predecessor's null call was real, but was not evidence of loader or message-order corruption. Run-specific read-only import observation on the executing Workers supplies the earlier boundary (bootstrap-trace2):

  1. Main Worker slot 0 installs MFC procedure 0x41f6cb for window 0x10001 through its ordinary CBT/window setup path.
  2. The native renderer probe starts at 0x877770, stack 0x7efffff8. Its GetCurrentThreadId returns 2. Its own native window 0x18001 uses procedure 0x873430 and completes WM_NCCREATE, WM_CREATE, show, activation, focus and sizing calls through its native default procedure.
  3. Secondary Worker slot 1, sequence 23, calls GetMessageA with all filters zero. Its next TranslateMessage/DispatchMessage frame contains {hwnd:0x10001, message:0x85, wParam:1, lParam:0}. Sequence 25's dispatch redirects EIP to the main window's 0x41f6cb on thread 2.
  4. Original x86 0x41f6cb calls its permanent MFC handle-map lookup at 0x41f648. Thread 2's MFC TLS slot 3 is initially zero and is initialized naturally there; it has no permanent wrapper for thread 1's window. The lookup's zero result becomes the object argument to 0x41f3b4.
  5. 0x41f3f3 reads that argument into EDI. The call at 0x41f43b reads [vtable+0xa0] with EDI/vtable zero, pushes return 0x41f441, and targets zero. The observed registers and stack match the original x86 instructions.

The first demonstrated emulator mismatch is retrieval of foreign NC work, not the final indirect call. $nc_flags_scan walked shared NC_FLAGS without checking the window owner. The existing ordinary PAINT_FLAGS scanner already has the equivalent owner restriction. The repair filters NC work by WND_THREAD/current_thread_id, preserving foreign flags without clearing, painting, or dispatching them. The existing owner-zero fallback remains.

GetMessage with a NULL window retrieves messages for windows belonging to the calling thread. Creating Windows in Threads assigns each window and its message queue to its creating thread. These contracts explain why entering a foreign MFC window procedure on the probe's stack/TLS is invalid. No x86 CALL/RET decoding, export address, thread-entry address or TLS read/write mismatch was demonstrated in this chain.

test-nc-flags-message-wake.js fails before the repair with 65537 !== 98305 at actual PeekMessage retrieval. Afterward, actual PeekMessage/GetMessage retrieve both NC messages for thread 2, preserve thread 1's pending bits, and allow thread 1 to consume its own messages later. The shared paint-work audit remains zero. This is not a null-call bypass or a WM_NCPAINT suppression.

Next wrong context: DestroyWindow adopts a foreign main window

The clean unmodified-Worker bootstrap-clean run executes the NC-only module b1f06ca6d34568dc80d71ae3782952dad023dca5bccb63347d7df34709d8caef. The observed bootstrap-trace3 run of that same module supplies the boundary: thread 2 now dispatches its own window's 0x83, 0x464, 0x85, 0x468 and 0x466 messages, passing the original failure. The last message triggers the probe's ordinary DestroyWindow(0x18001) at guest return 0x8745a1. Its own native procedure successfully handles WM_DESTROY and WM_NCDESTROY.

DestroyWindow's exit hook then observes EIP redirected to 0x41f6cb instead of returning to its API caller. The same MFC null call follows with message 7 (WM_SETFOCUS), EIP zero, return 0x41f441, and EDI zero. This call is synchronous inside DestroyWindow, not another NC message stolen by GetMessage.

$destroy_main_window_lifecycle treated the per-instance main_hwnd as process-wide: its adjacent-HWND shortcut and general successor scan could both choose a foreign-thread top-level. The DestroyWindow focus continuation then entered that window's procedure on the destroying thread. Both successor selection paths now require the calling thread's ownership. A thread destroying its last probe window retains no foreign main/focus callback target.

SetFocus requires a window attached to the calling thread's queue; the window/thread ownership contract also applies during teardown. No AttachThreadInput path was observed here. The change does not attempt a broader focus or destruction redesign.

test-destroy-main-window-recreation.js fails before this repair with 98306 !== 0. It now covers both adjacent and nonadjacent foreign successor candidates, plus the existing same-thread replacement, quit-marker, synchronous focus callback, stdcall return and nonvolatile-register preservation cases. An intermediate parenthesis error was caught during test compilation and fixed; its failing log is retained separately from the passing final test/build.

Qualification boundary

The initial observer attempt (bootstrap-trace) could not decode names directly from shared memory; its records are unqualified. bootstrap-trace2 and bootstrap-trace3 copy only the diagnostic name bytes before decoding, have no observer errors, and use console output from the executing Worker to avoid the predecessor's observer-routing failure. Entry/exit hooks preserve their original imports and arguments/results. Redirecting/yielding exit-hook EAX values are not claimed as completed API returns.

The expanded final state collector in clean/trace3 could not serialize its renderer-window object, so those JSON files retain fetched module/host hashes and errors but no state snapshot. Their console traces and PNGs remain useful; no complete state/registry acceptance is inferred from them. The collector is now bounded to plain scalar window fields for the next run. The earlier trace2 state snapshot is intact.

Before and NC-only screenshots were personally reviewed: the chooser is still blank. The inherited native catalogue, initial Software selection and storage/reload route is reused; authentic missing-d3d retry to native d3da and THRASH lookups execute again. Final registry-transfer acceptance remains pending. Secondary-thread teardown is a later checkpoint than the original NC failure, not chooser acceptance.

After both repairs the final production module is SHA-256 7ec49cbbae4cb177909fd8fd9d3599713ee5f2a5ffb7f7e8c890848a540052bf (2,182,452 bytes). The two ownership regressions, NC/input-order, paint-work counts, synchronous window ownership, dialog tab-stop focus and Win16 destruction neighbors pass. The locked full build passes. The optional BUTTON BS_NOTIFY neighbor fails at its preexisting button-down focus expectation; an in-memory Git-source baseline check reproduces the identical failure against both original source fragments at 0df0f1667, without reverting shared files or copying a checkout. That separate preexisting focus-notification issue is not repaired here.

The next dependency is final native chooser qualification with both repairs once /home again has more than 5 GB free. The task reported the shared disk drop (about 5.35 GB to 4.46 GB) on messageboard.txt and started no new browser or box afterward. All its earlier browsers/servers closed in finally blocks. Only after local chooser and registry-transfer success may a fresh UUID-idempotent box be allocated with the independent 25-minute watchdog, ordinary native D3D choice/OK and strict reviewed input/ACK/UNKNOWN guards. No such remote step has occurred. Actual D3D context/backend/fallback, game acceleration/steering, race motion, three-renderer timing, full27, audio and parity remain open.