Lords II WinExec executable search

Evidence: scratch/runs/20261107-lords2-winexec-search/. Follows the published original installation after PROGMAN repair, e562c1fd5, whose actual terminal and 04:21:07Z release were checked. Work is limited to the emulator and generated ordinary launcher fixtures; the original guest is unchanged.

The predecessor's retained command is setup32.exe C:\SIERRA.INF A, issued by the Win16 task loaded as C:\WINDOWS\TEMP\_SETUP.EXE, while the VFS current directory is C:\. Its helper exists in Windows Temp and Sierra, with SHA-256 b62ffd44fa2c8a0ac5873320f6d6c86827d6e78bb0af3e605cdf7b8f50fa8218; the current-directory candidate is absent. This is application-directory search evidence, not permission to register a special application or rewrite a guest command.

Microsoft's WinExec contract, checked during this task, searches a bare image in the application's load directory, current directory, Windows system directory, Windows directory, then PATH. A supplied directory path does not invoke that bare-name search. An extensionless executable gets .exe; a final component containing a dot keeps its extension. Child current directory is inherited independently of executable lookup.

Preparation preserves the shared source window: Aranna released at 04:25:52Z, graphics claimed the next window at 04:26:13Z. Graphics released at 04:29:18Z; the repair was applied only after that release and explicit file claims. The patch carries the calling image and actual guest PATH through WinExec's private host handoff; production browser and CLI consume the same resolver. The Win16 task's retained image path takes priority over the root process identity. Search selects only file entries, and failed WinExec lookup cannot substitute a registered application's basename.

The Win16 adapter also needs to consume browser startup tickets before narrowing its return to 16 bits: truncating 0x57450001 produces error 1 even after successful launch. The regression covers pending startup without a duplicate launch, success after readiness, direct errors, and the far Pascal stack. This uses the existing startup handshake, not a guest workaround.

Repair 9b1da000e was committed with explicit paths and immediately pulled/pushed on main. The mandatory locked build passed. Production browser and CLI callback regressions cover application directory over cwd, successive system/Windows/PATH fallback, extensionless and quoted names, explicit relative paths, missing files, unchanged argument quoting and inherited cwd. The CLI test uses the real snapshot/export/merge transport with only the OS fork substituted. Native WinExec, Win16 WinExec, startup readiness, VFS launch, actual NE module queries, task argument/MMSYSTEM and PROGMAN controls passed. The module-query fixture now explicitly returns unsupported process_ctl=-1, instead of accidentally describing every call as a pending startup wait.

One fresh box, bx_8p8f4s4j, was created at 04:33:52.608Z with a persisted UUID Idempotency-Key; original expiry 06:03:52.464Z. No old box was resumed. All 639 staged runtime/media pins passed SHA-256 verification before launch, and all 181 media files matched their manifest. The staged source identity is Git ce0330c444ea43742259624db5e9cd0240b2f156 plus the retained relevant diff and untracked-file patch; the executable-search repair is 9b1da000e. The standard module is e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467; compatibility is 04acbb89f66ad9814836b5e2c34de659b518f4e1994f50c4ee8069c991f02f72. Both exact modules are retained. Both modules passed PROGMAN, redraw, WinFlags, temporary drive, LZCopy and observer passivity controls on the fresh box. Production browser/CLI routing, native NE module queries and the queued-input validator also passed there.

Source/build/local runtime ownership was explicitly released at 04:34:44Z after independent remote pin verification; the original uses its immutable separate runtime. The original remained in that one browser session.

At 04:38:19Z the normal install-complete button caused setup32.exe C:\SIERRA.INF A to run. console.log records [ShellExecute] launching c:\windows\temp\setup32.exe from the caller's filesystem, then child ExitProcess code 0 and the real installed shortcut. The installation's ordinary warning/confirmation dialogs were answered through visible controls; no guest file, register, configuration, or API result was changed. The setup-complete capture and matching state are after-complete-ok.png and after-complete-ok-state.json.gz.

The real desktop shortcut was opened by two trusted pointer clicks, not by calling the shell's launch API. Its target is C:\SIERRA\L2Demo\LORDS2.EXE, SHA-256 c83df505c4be850c10425f1770090ba3ca31fa6b87d8bce2cde3c66ae11dd0dc; installed SMACKW32.DLL is 16725648b30a2bdcd2959a52ca92f9a0e14a7fc9fb26c84e90363c2631887f83. installed-identities.json records these unchanged bytes and the matching final state. The title screen rendered (installed-launch.png); ordinary clicks reached the demo greeting (after-title-continue.png), player/opponent confirmation (after-proceed.png), then the campaign map with its Game Objectives overlay (after-begin-conquest.png, 04:42:05Z). I personally reviewed each capture against the state snapshot made immediately afterward. The final game was running, with no recorded Win16 far fault. All instantiated module identities match the pinned standard module above.

The next concrete blocker is the harness's command deadline, not a demonstrated emulator fault. tools/web-input-probe.js, probeSteps, fixes the command lifetime at 420000 ms and throws ordinary-input queue deadline at line 297. It closed Chrome normally at 04:42:13.359Z, code 0, before objectives could be dismissed or a map/economy control exercised. original-output.log ends with that exact error and PAGE CLOSED +444.0s; the extra time includes launch and completion of an already queued step. Thus installed title/menu/map rendering and ordinary progression are verified, but player-controlled map gameplay is unverified. No original restart, second box, old-box resume, or guest bypass followed. A future separately authorized run needs sufficient command lifetime for installation plus actual map interaction; it should begin from this published repair and preserve the same identities.

Evidence caveats are retained rather than erased. The first proposed input was rejected locally by the production queued-input validator because it used unsupported click; no command was published or guest action taken, and subsequent inputs used supported canvasclick. A read-only HTTP assertion initially expected every response to hash-match; normal 404s and the dev server's documented localhost index injection invalidate that broad assertion. The corrected identity check confirms actual host.js and Wasm response bytes and instantiated Wasm identities; all 639 staged file pins independently pass. A later read-only process-context read found Chrome already gone after the normal command deadline. These are retained harness/diagnostic errors, not safety flags or replayed guest operations.

After normal close, all 181 original media files and all 639 runtime pins passed SHA-256 verification again. Unique logs, matching captures/states, module identities and cleanup evidence (84 files, retrieval hashes in retrieved-final.json) were retrieved before the explicitly delegated normal stop. remote-cleanup.json confirms harness 6718, server 6717, deadline guard 6719 and Chrome 6737 gone; HTTP 36939 and debug 46537 closed; no Chrome using the owned profile; /tmp/wine-assembly-input-BTuarR absent. Fresh box bx_8p8f4s4j stopped normally at 04:43:17.446Z and is archived, with completed non-forced stop. Its original expiry was 06:03:52.464Z; no active expiry remains. There is no live test runtime to hand back. Git plus retained relevant binary diff/new-file patch is the source recovery record; no source or media archive was made.