NFS3: missing vendor DLL failure and native suffix retry

The general loader repair lets the original d3dtest.dll retry d3da and resolve the authentic renderer exports. Native setup then hits a separate null virtual call while creating its chooser. Direct3D selection, chosen-game state transfer and race control remain unqualified. No runtime box was allocated: box ID, expiry and stop receipt are null, state NOT ALLOCATED.

Evidence: scratch/runs/20261108-nfs3-loadlibrary-missing-vendor/. Source recovery uses the recorded Git SHA and relevant binary diff; media is referenced by existing path and SHA-256. No source/media archive was created.

Loader contract and regression

Microsoft's LoadLibraryA documentation specifies NULL on failure and GetLastError for the extended error. Its naming rules append .dll to an extensionless name; a trailing period suppresses that suffix. System error codes define 126 as ERROR_MOD_NOT_FOUND.

Previously, after native-file lookup failed, an arbitrary bare .dll name returned the EXE base. It now returns zero/error 126 unless a recognized system module implements the request. Existing normalization, native mapping yields, cached retains, locks, reference lifetimes, dependencies and real built-in identities retain their existing paths. d3d9.dll remains an implemented module with its own identity. The optional absent uxtheme.dll case now also sets 126. There is no d3d alias, guest patch or renderer injection.

The actual-handler policy fixture fails against the original source with d3d: actual 4194304, expected 0. The repaired fixture checks unrelated missing vendor names, case/path/stem/trailing-period normalization, error 126, no synthetic reference mutation, successful built-in LastError preservation and available qualified vendor native loading. The system-path neighbor also checks A/W/Ex entry points. Its previously empty harness now explicitly mounts the system files whose qualified availability the existing execution policy requires.

Focused missing-dependency, built-in policy, system-path, synchronous-send, startup rollback, import-graph and real PE lifecycle regressions pass. The last includes both main/cooperative and real Worker thread cases, nested lookup, unload, failed attach, cached sibling handles and join notifications. The locked full build passes. The callback census records a narrow lifetime review; broader retained-context ownership remains pending.

Production native setup and next failure

The clean local Chromium run executes the published native catalogue installer, first setup and second setup after reload using the existing registrations and 133-file fixture. All 133 media files (22,070,305 bytes), plus the manifest, were freshly hash-verified. Native catalogue writes and the utility's own first Software Renderer / softtri choice transfer through ordinary shared storage. The chosen-game registration remains free of startup renderer writes, and the standing Glide defaults are unchanged.

The first exploratory run crossed the build's publication time and fetched both old and repaired modules; it is not the clean qualification. bootstrap-clean fetches only the repaired 2,182,404-byte production module, SHA-256 f5cfb1c6dea588eb538831d42069c699f6b7a6306cb530b8eb0abb437af754a7. The real guest main Worker and secondary Worker are recorded in its logs/state. Served JS/HTML response hashes and the fetched module bytes are retained.

Second setup now records:

LoadLibraryA("d3dtest.dll") -> module 0x0086f000
GetProcAddress(0x0086f000, "D3DTEST_GetCards")
LoadLibraryA("d3d")
LoadLibraryA("d3da")
GetProcAddress(0x008ae000, "_THRASH_about@0")
... authentic _THRASH_* exports on the same native module ...

Original d3dtest.dll at runtime 0x008703db tests the first LoadLibrary result; only zero executes its own %sa formatting and second call at 0x008703fc. This is the original suffix retry, not a host replacement. D3DTEST_IsHAL repeats the same route. Entry traces and module-handle arguments prove native retry/export resolution; a yielding LoadLibrary entry alone is not a final return-value observation.

The chooser window appears but remains blank. Secondary thread 1 exits at EIP zero with prev_eip=0x41f430, prev2_eip=0x41f41a, ESP=0x7effff38, EBP=0x7effff94, EDI=0, ECX=0, ESI=0x85. Retained guest stack words start with 0x41f441, 0x85, 1, 0. Original executable disassembly at 0x41f430 pushes lParam, loads the vtable from [EDI], sets ECX from EDI, pushes wParam and message, then calls [EAX+0xa0] at 0x41f43b. The return address 0x41f441 matches the stack. Guest memory at zero and the corresponding vtable slot are zero. These observations identify a null object/virtual dispatch during WM_NCPAINT (0x85); they do not establish why the object lookup is null.

Additional run-specific Worker import observers were exploratory diagnostics, not accepted return-value instrumentation. The first response override broke Worker isolation headers and fell back to cooperative execution; subsequent message routing/collector attempts did not produce usable return records. Their logs and hook/source hashes are retained separately. Acceptance above uses the unmodified production Worker in bootstrap-clean, the executed native retry branch and native handle arguments, not those observer attempts. The unrouted observer browser was explicitly closed and its stuck collector stopped.

Next work is the native chooser's window-object/message dispatch during creation. No click or game input was sent after this failure. The remote-runtime gate did not pass, so no watchdog/allocation/stop action was needed. All task browsers and loopback servers close in finally. Physical GPU remains unknown; no actual D3DIM/context/fallback acceptance, acceleration, steering, road-relative motion, FPS, parity, three-renderer benchmark, full27 or audio claim is made.