NFS3: live Win9x device enumeration and native bootstrap

Implemented the missing Config Manager registry tree. Original production Worker 3DSetup.exe now opens HKDD\Config Manager\Enum, enumerates the virtual display and monitor, reads their HardWareKey values, and reaches the end of that enumeration. The chooser and Direct3D race control remain unqualified. No remote box was allocated (ID, expiry and stop receipt null; state NOT ALLOCATED). All local probe browsers close in finally.

Evidence: scratch/runs/20261108-nfs3-win9x-device-enumeration/. No input, renderer injection, clock/RNG override, guest patch, replacement DLL, source archive or media archive was used. Physical GPU remains unknown. Three-renderer benchmarks, full27, audio, FPS and parity remain open.

Platform contract

Microsoft's KB250568 describes Config Manager as the RAM hardware tree, recreated at startup from the current device configuration (Microsoft-authored article, retained mirror). The current compatibility implementation in Wine's own wineboot source creates Config Manager\Enum independently of individual device creation. Wine's PnP implementation creates eight-digit hexadecimal child keys with HardWareKey strings pointing to device instances and a DWORD Problem value. Microsoft documents root-enumerated devices and vendor namespaces. The Microsoft KB describes Win9x; the current Wine implementation and current Microsoft driver documentation support the representation, not a claim that this emulator boots a real Win98 PnP driver stack.

The emulator already exposes one fixed primary Wine-Assembly Display and one Default Monitor through EnumDisplayDevicesA/W. The registry now represents those same two virtual devices, using explicitly emulator-owned ROOT\WINEASSEMBLYDISPLAY\0000 and ROOT\WINEASSEMBLYMONITOR\0000 instances under HKLM\Enum. Each has DeviceDesc and Class; its live HKDD record links through HardWareKey and reports Problem=0. These are virtual instance names, not fabricated PCI/physical GPU IDs, driver filenames or D3D capabilities. No PCI device, network adapter or other unimplemented bus device is invented. This fixed inventory matches the current fixed display service; hotplug and a full Configuration Manager API implementation are outside this change.

Config Manager records use a RAM backing map through the ordinary registry service. Case-insensitive paths, relative opens, value lookup, key/value iteration, metadata, writes and deletion use the same registry machinery as other keys. Export/import excludes this live subtree, and preexisting persisted copies cannot introduce stale physical devices. Clear/reseed reconstructs it. Closed/unknown open/create/query/set/enum/close handles fail; enumeration requires space for the terminator and reports ERROR_MORE_DATA without truncating output. Absent keys and values retain their normal failures, and enumeration ends with ERROR_NO_MORE_ITEMS. There is no unconditional successful open or empty-tree shortcut.

Ordinary native setup and next dependency

The retained predecessor trace failed at opening Config Manager, after the utility wrote an empty description. With the service implemented, a production Chromium Worker run passes that point, then reports “Could not open 3D Data registry key.” The bundled Setup/english/SETUP.INS explicitly launches 3ddata.exe before 3dsetup.exe. Executing that original catalogue installer through the production launcher writes the EA catalogue using native RegCreateKey/RegSetValue; its real catalogue includes vendor/card IDs as catalogue data, not installed physical hardware. It exits with code zero.

Local-only app registrations now expose that bootstrap, the original setup, and the original game using its chosen renderer. Setup receives only the DisplayName installation metadata specified by SETUP.INS; it receives no selected renderer or renderer capability values. All use the existing 133-file manifest and ordinary shared storage. Existing Glide entries are unchanged. The chosen-renderer game entry has no startup registry writes.

On first native setup with no description, the utility itself selects and writes Software Renderer / softtri, then exits zero. This is its installer/default path, not a user-selected Direct3D result. A second ordinary launch (also repeated after a full page reload, preserving native registry state) reads that native choice, loads the shipped d3dtest.dll and resolves D3DTEST_GetCards.

Detailed ordinary API tracing then records:

LoadLibraryA("d3dtest.dll") -> native handle 0x0086f000
GetProcAddress(0x0086f000, "D3DTEST_GetCards")
LoadLibraryA("d3d")
GetProcAddress(0x00400000, "_THRASH_about@0")
... 33 other _THRASH_* requests against the EXE handle ...
Program exited (worker), EIP/diagnostic registers zero

The last line is the host's terminal diagnostic, not a successful native ExitProcess or accepted chooser. No further input was attempted. The shipped tree contains d3da.dll, not d3d.dll. Static inspection of the actual requested module explains the bounded next dependency: d3dtest.dll original VA 0x100013d9 calls LoadLibrary; 0x100013e2 takes the nonzero branch directly to export lookup. Only a zero result formats %sa (at original VA 0x1002cf08) and retries, naturally requesting d3da. The current general LoadLibrary handler normalizes the first request to d3d.dll, finds no file/static module, but its bare-.dll fallback returns the EXE image base. This suppresses the utility's authentic suffix fallback. The trace demonstrates the wrong nonzero handle path; it does not establish that every later symptom has this sole cause.

Next work is a general missing-vendor-DLL LoadLibrary failure contract, with regression coverage that preserves genuinely implemented built-in system DLLs (including the existing d3d9.dll API-backed case), followed by the utility's own retry of the shipped d3da.dll. Do not alias d3d to d3da, inject a chosen renderer, or patch the guest. This is separate loader behavior; it was not changed speculatively in the registry fix. Until that ordinary path qualifies, no runtime box or original-game inputs are justified.

Validation and recovery

test/test-storage-registry.js records a failing-before Config Manager open and passing-after live enumeration, relative A/W opens, linked values, absent values/keys, buffer canaries, closed handles and snapshot isolation. Existing registry/INI/COM/state-transfer checks also pass. The display enumeration test now compares actual API device descriptions against the registry records. Focused DirectX metadata and snapshot-diff tests pass. The focused registration check preserves Glide and rejects renderer seeds on the native route. The locked full build passes; the final run's build/test logs are retained. The broad debug-app suite was not claimed: the predecessor documented unrelated stale Uplink assertions.

Native catalogue and first-setup writes demonstrably transfer between production Workers and across reload through normal shared storage. This does not prove successful setup-to-D3D-game transfer: the native chooser is still blocked as above. Screenshots show the intervening error or the desktop after terminal exit, not a working chooser or race.

All 133 standing media files (22,070,305 bytes) were freshly SHA-256 verified; the manifest and paths are recorded without copying media. Source recovery uses Git SHA plus relevant binary diffs, including relevant untracked code as patches. The fetched production WASM is 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. The detailed run additionally hashes served JS/HTML responses. Browser: Chrome 151.0.7922.108; isolated real guest main Worker, 1000×800 viewport. No D3DIM context/module/fallback acceptance, acceleration, steering or road displacement is claimed. No remote allocation, watchdog, resume, deletion or public deployment occurred; no task box remains billing.