Lords2 Win16 GetTempDrive
Task: NEW-GAME-LORDS2-DEMO-20261015. Evidence: scratch/runs/20261107-lords2-tempdrive/. Follows the absolute __WINFLAGS repair, published as f832a18a6; that implementation and its machine results must remain unchanged.
The retained original reached KERNEL.92 with one zero WORD argument at 0057:4E9C, called thunk 001F:00DC, and trapped before returning to 0057:4EA3. This is an unimplemented emulator API, after the preceding repair allowed its actual GlobalAlloc(2,65536) to succeed. Installation completion and gameplay have not yet been established.
Microsoft's Windows 3.0 Programmer's Reference, entry 4-216, describes a BYTE suggested drive letter, with zero selecting the current disk. This is a character, unlike GetDriveType's zero-based drive index. The Wine export specification establishes the actual Pascal WORD stack slot and full DX:AX return. Its Win16 implementation documents the extra register contract: AL is the uppercase drive letter, AH is a colon, and DX is one. It selects the temporary path's drive and ignores the suggestion, using C when querying that path fails. We target this established Win9x-compatible reference behavior through the emulator's existing VFS temporary-path helper; the older SDK description and the reference selection policy differ.
Aranna-relative-map-load released runtime/source/build at 02:09:47Z. Before taking the lane at 02:10:39Z, I independently checked its retained process IDs, ports, profiles and descendants, plus all four locks; handoff-cleanup.json records that check. The source change adds ordinal 92 dispatch and one handler in src/09e-win16-api.wat. It asks the existing host temporary-path helper for its path in invocation-local Win16 scratch space, uppercases the initial letter, sets AX to the letter and colon and DX to one, and removes the two-byte Pascal argument. It preserves the earlier __WINFLAGS implementation and existing machine-result expressions.
The generated test/test-win16-tempdrive.js NE consumer traps against both retained predecessor modules. With the repair, both new standard and compatibility modules pass: four ordinary imported far calls with zero, uppercase, lowercase and arbitrary suggestions; independently sampled SP after every call; AX/DX; preservation of BX/CX/SI/DI/BP; default C: temporary path while the current directory is D:; alternate E:/Z: temp paths; and the reference fallback after an empty path result. The alternate paths are synthetic test VFS providers, not guest configuration changes. Existing test-win16-winflags-import, NE-loader, huge-pointer-import and arena-bounds tests pass. Both modules also pass the run-local passive observer controls, including complete 512 MiB memory and CPU equality with observation enabled, disabled and saturated. Ordinary input range controls and flock /tmp/wine-build.lock bash tools/build.sh pass. focused-tests.json lists the checks. Repair d1f0ad7d3 was committed and published with strict SSH pull/push before the original attempt.
The one original ran locally from 02:12:47Z to normal cleanup at 02:17:04Z, under the 35-minute total bound. Both observed instantiated modules were the tested standard module, SHA-256 f18229189ae8e50fe938baaa6f2fe2821ee60c297acf7116f5cfeb8529915eaa, 2,181,162 bytes. tested-modules.json records the compatibility identity too. All 181 original media files and the browser manifest matched their inventories before and after. The HTTP observer recorded 202 exact body matches, two normal dev-server index injections reconstructed exactly, and eight retained absent build-info/CARDS probes. It does not establish that each guest file read caused an HTTP request. Git commit plus the recorded relevant binary diffs and generated-test patch preserve source provenance; no source tree or media archive was made.
I personally inspected the startup menu and each successive system-test screen before sending its ordinary offered continuation. Seven balanced canvas clicks and a normal quit produced eight acknowledgments, with no unknown/partial operation or replay. The displayed memory, Windows version and CPU results were left as reported. Audio was not independently verified. The original called GetTempDrive(0) at the retained call site and now returned to 0057:4EA3 with AX=3A43 (C:), DX=1, and a six-byte stack delta including the far return. original-causal-result.json preserves the entry and return contexts; the full passive traces are losslessly compressed in after-continue.json.gz and after-destination.json.gz, with their uncompressed hashes.
This advances installation to a new visible Destination Directory prompt offering C:\SIERRA\L2Demo (after-continue.png). Accepting that default reaches Copy Files, which displays: "Setup cannot find the file, C:\LANGUAGE.INF, on the current disk" (after-destination.png). No Retry was selected. Read-only live VFS inspection at that point (next-blocker-live-vfs.json) showed current directory C:\, a resident c:\language.inf with 3,313 bytes, and a newly created c:\sierra\l2demo\language.inf with zero bytes. The original media inventory contains LANGUAGE.INF, 3,313 bytes, SHA-256 2a758d7248807e9566bb24d461426dc3732704a3fe20c4a624d31a9c216345d2. Thus absent staged media is not an adequate explanation for this next copy/I/O discrepancy. The first underlying API/CPU mismatch in that path has not been isolated; the captured ordinary C-runtime DOS open/read/close blocks and profile-query contexts are evidence for the next task, not proof of a new emulator defect. No extra repair was made. Installation completion and gameplay/world movement remain unverified.
Probe and Chrome exited zero. Independent cleanup checked all four owned PIDs, ports 36929/32799, the owned browser profile and its descendants absent, and all four locks free. Runtime/source/build ownership was explicitly released at 02:17:27Z; the successor took it at 02:17:48Z. No boxes were used or stopped. The initial board-watch tool session separately ended with status 143, cause unknown; the replacement watcher was explicitly cleaned up. No guest/media/config/save edit, machine-value adjustment, deployment, archive, automated risk flag, retry or second original occurred. Next work is a bounded investigation of the first failed copy/I/O result for the already-present C:\LANGUAGE.INF.