Icewind Dale demo — local candidate

The official English demo is pinned in test/candidate-corpus/manifest.json from Archive item icewind_dale_eng_demo. Its README expressly prohibits copying/electronic distribution, so the extracted fixture stays ignored and must not be deployed or rehosted.

2026-10-31 local COM lifetime audit; no repair contract established

Run 20261031T2220Z-icewind-local-lifetime performed only a cheap static audit of the retained original executable and current implementation. No browser, build, engine edit, guest edit or new gameplay capture occurred. Party Formation, creation, Prologue and tavern movement remain blocked. The earlier polling trace and its first-failure/final captures remain evidence of the earlier run only; expired asynchronous frame values are excluded.

The original call/dataflow now distinguishes ownership from session bookkeeping:

The implementation's activation path allocates a primary object, queries the requested ANSI interface, increments its reference count, and releases the temporary factory reference. The query returns that primary identity; the DirectPlay4A upgrade preserves inherited slots. AddRef uses the shared COM counter; Release decrements it and closes/frees the owner at zero. The audited stdcall cleanup is 24 bytes for CoCreateInstance, 16 for QueryInterface, eight for AddRef/Release and 16 for GetSessionDesc, matching return address plus arguments. No mismatched interface, missing activation reference or ABI defect was established on this path. This is a source audit, not a fresh runtime refcount measurement or comprehensive COM conformance claim.

Microsoft documents an acquired interface as a reference the client must release, and requires returned interface pointers to carry an added reference. Neither rule supplies a reason to clear the guest's stored output or fabricate an open DirectPlay session. The concrete unresolved dependency is the original engine's intended handling of provider zero + retained interface + active local-session byte: the inspected caller itself makes that combination and then polls it. No proven generic API obligation resolves it. Work stopped at this dependency; changing NOCONNECTION, activation output, caller flags or release behavior would be speculative.

Evidence: audit.py, audit.json, audit.log, native-network.asm, focused native-activation.asm, native-message406.asm, native-release.asm, native-local-create.asm, native-poll.asm and RECOVERY.txt in this run. No heavy gate or browser lock was acquired; the current Return Fire allocation was left untouched. Boat ID/state/expiry N/A.

All 1,106 original pins and 27 CD2 joins matched again. All 867 compiler pins matched; the 528-file source comparison found only the separately owned Return Fire edit to src/09a8-handlers-directx.wat. source-audit.json preserves the audited excerpts and hashes. The concurrently rebuilt local module was hashed as local build output only; there is no actual HTTP-loaded module in this run. The first scratch assertion used a transposed IID text suffix (a1a7); the PE bytes decode a7a1. That failed assertion and the corrected audit are retained.

2026-10-31 native local-session ordering and cleanup branch identified

Run 20261031T2210Z-icewind-session-order reused the original retained 2058 installation and read-only CD2 mapping. One browser ran 22:11:54–22:14:42; three balanced native Escape presses and one trusted Create Game click returned to GAME MODE. Fresh failure/final captures are retained; this run did not capture the brief CLEARING GAME DATA screen. No Party Formation, character creation, Prologue or tavern movement; no engine repair or build.

Original disassembly and the fresh read-only trace narrow the cause beyond the previous late, quarantined observer:

Thus an earlier failing initialization/Open call was not found: native code deliberately bypasses both calls in this local path. The retained COM object is nevertheless queried, and the valid outside-session result triggers native cleanup. This is not a defined generic API contract defect and is not permission to change GetSessionDesc's result. The remaining question is why this original local path retains and polls that object. A generic repair still requires an independently defined contract and focused oldFAIL/newPASS test.

The exact exported scratch observer passed cheap tests before launch: copied shared views, RegionMap.REGIONS, unchanged guest memory, original log return/receiver/error forwarding, COM/vtable resolution, filtering before the bounded cap, and error containment. The shared-decoder oldFAIL/newPASS used an explicit strict-browser decoder mock; Node itself accepts shared views. Browser observer errors and dropped records were zero. Live memory is asynchronous: the guessed savedInitializeResult in record 13 belongs to an expired frame and is excluded. Only qualified active-frame reads support the result above. The previous late armDesc/armSafe results remain quarantined.

All 1,106 original files and 27 CD2 joins matched, as did the prelaunch 528 source and 867 compiler pins. The uncached actual HTTP module was 1,898,762 bytes, SHA-256 d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8. Exact geometry/input, 14 HTTP guard cases, five isolated-server statuses and original read-only CD2 prerequisites passed. No guest edits, copies, reinstall, manufactured markers, injected saves or API-result changes occurred.

Controller, supervisor, launcher, monitor, board watcher, attached inspection helper and six adopted children were durably reaped with zero status; Chrome closed normally. The stop-request sender exited one after an acknowledgement timeout because the controller exited after capturing its final image without writing the command acknowledgement; this is retained, not reported as a successful sender wait. Final audit found all 23 owned PID/start identities absent, whole Chrome/global crashpad empty, ports 44041/42313 closed, profile gone, nine trusted events balanced, and lock free. Minimum free space was 2,562,224,128 bytes. A source-file change by the next worker after runtime (src/09a8-handlers-directx.wat, OWN 22:14:49) is recorded separately from the unchanged original inputs and loaded module. NEWFULLRELEASE was announced at 22:17:05; Boat ID/state/expiry N/A.

Evidence: session-boundary.json, session-observer-latest.json, exact observer and tests, native-*.asm, 09-first-failure-before-teardown.png, 10-final-before-teardown.png, stop-request-wait.json, full-release-final.json and RECOVERY.txt in the run directory. The combined recovery driver, previously unlaunched, was browser-tested in this run.

2026-10-31 original Create Game dispatched; creation still blocked

Run 20261031T2150Z-icewind-native sent trusted native Create Game clicks in the unchanged original demo. It reached the labelled CLEARING GAME DATA progress screen, then returned to the labelled menu. Party Formation, character creation, Prologue DONE, unpaused tavern, and paired floor-click movement were not reached. There was one browser launch, 21:50:58–22:02:00, with six balanced mouse clicks within that session. No production repair or build is claimed.

The previous optional-DOM-method correction passed cheap tests but still omitted the renderer's exclusive presentation transform. Before any mouse input, browser inspection found its proposed rectangle (0,34,640,480) differed from the displayed native viewport (22,34,979,734). The corrected scratch helper requires an uncropped 640×480 native viewport and uses getPresentedRectClient(). Its native (480,175) maps to CSS (756.25,301.6041666666667). The exact exported helper passed scaled-coordinate, balanced-dispatch and ambiguous-crop refusal checks and was executed through Puppeteer in the same browser. Actual DOM evidence records trusted mouse move/down/up events; the native last-message data independently contained WM_LBUTTONUP at (479,174). The attached helper inherited FD9 pointing to the lock inode while the original runtime supervisor retained its exclusive flock. It did not acquire another lock or launch another browser.

Before the later diagnostic logging error, the native Create Game transition already returned to the menu. Native session initialization performed the compound-address sizing/allocation calls; a read-only return observation captured CoCreateInstance(CLSID_DirectPlay, IID_IDirectPlay4A) returning S_OK and output 0x08221620. The existing API trace then recorded GetSessionDesc(NULL, size) from native return address 0x008D4D12, followed by DestroyPlayer(1) and Close. No Party Formation transition was accepted.

Later instrumentation captured that query returning 0x887700AA (DPERR_NOCONNECTION), size zero, shared session state zero and session pointer zero. The next native LeaveCriticalSection, return address 0x008D4D2E, exposed the saved HRESULT at sizePtr-8, consistent with the original function's disassembly. The actual object's slots 6/22/24/38 resolve to CreatePlayer/GetSessionDesc/Open/InitializeConnection as expected. No Open call was observed. This does not establish a generic API violation: the query result is valid outside an open session, and the first native caller or ordering error that causes the menu return remains unknown.

Late scratch diagnostics had real failures: one helper syntax error, a TextDecoder attempt on a shared view in a host-log observer, and an incorrect RegionMap.regions lookup. The bounded replacement copies the view and reads RegionMap.REGIONS. The logging failure happened after the original menu return had been retained; later query results are explicitly quarantined from claims about an uncontaminated first cause. No guest bytes or API return values were patched. Optional missing music-file probes and directory-open errors were retained without treating them as a repair contract.

The exact HTTP guard's 14 cases and isolated server's 200/400/404/416/500 checks passed. All 1,106 original pins, 528 source pins, 867 compiler pins, and 27 original CD2 joins matched before and after. The retained installation and original C: paths plus read-only D: CD2 presentation were reused directly; no reinstall, download, fixture copy, manufactured marker/volume, KEY/INI/BIF rewrite, save injection, WININIT edit, or foreign cleanup occurred. Runtime HEAD was b7ea3fcc500ecd346dcf656a4c4f37c35b7fcadc. The actual HTTP-loaded module was HTTP200, uncached, 1,898,762 bytes, SHA-256 d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8. Chrome 151 used isolated Worker/Threads, WebGL, native 1024×768 viewport, default audio and normal clock. No FPS, audible-output, or gameplay claim.

After the preteardown failure capture, controller, supervisor, runtime launcher, board watcher and monitor all completed durable waits with code zero. All 26 attached Node helpers were durably waited; two diagnostic helpers truthfully exited one. Six adopted children were reaped, Chrome exited zero without a signal, all 49 recorded own PID/start identities were absent, ports 37827/44305 closed, profile gone, all 24 trusted input events balanced, and FD9 free. Minimum observed free space was 2,488,537,088 bytes; start reserve exceeded 2.5 GB and runtime remained above the 2.1 GB early-stop threshold. Boat ID/state/expiry N/A. Explicit NEWFULLRELEASE was announced at 22:02:56.

Evidence and recovery: session-boundary.json, 10l-cocreate-result-returns.json, 10q-final-contracts-safe.json, 10r-object-contract-vtable.json, native-geometry.json, displayed-native-tests.json, trusted-input and actual HTTP receipts, reviewed captures, full-release-final.json, and RECOVERY.txt. The original log is retained as browser-console-before-safe.jsonl.gz plus browser-console.jsonl. native-layer.cjs preserves the tested input helper; recovery-browser.cjs integrates it but that combined driver has not been browser launched. Next: establish the first native session caller/order between successful COM activation and the first session query/menu return; require a defined contract and oldFAIL/newPASS regression plus build before any generic production repair. Do not change the original installation to bypass character creation.

2026-10-31 original optical scan and labelled native menu reached

Run 20261031T2137Z-icewind-optical executed the unchanged original demo's optical scan and reached its labelled native menu. Character creation, Prologue DONE, tavern, and floor-click movement remain unverified: the first Create Game click failed in this run's scratch coordinate helper before any mouse event was dispatched. This is a harness failure, not an observed guest disc failure or an emulator API violation. There was one browser allocation, no retry, and no production change or build.

At 21:43:18, the actual guest called GetLogicalDrives with return address 0x0043AD68; the host returned mask 12 (C:/D:). Two GetDriveTypeA calls returned to 0x0043ADFC; the second host lookup reported type 5. The native GetVolumeInformationA call returned to 0x0043AE65 and proceeded to FindFirstFileA through the original MFC wrapper, return address 0x00929F05, for D:\cd2\data\iwdcd.2. Its VFS result was handle 1778384913, entry iwdcd.2, size 1. original-optical-runtime.json retains these actual calls and results. Direct Win32 BOOL/handle return logs were not captured; the successful volume branch is established by the subsequent native marker lookup. There is no observed failing optical call to fix. The intro then read the original C:\CD2\data\MVEfile1.bif; two ordinary Escape presses reached the detailed GAME MODE menu with CREATE GAME, LOAD GAME, MOVIES, and QUIT GAME labels. No character or gameplay is inferred from that menu.

All 1,106 retained original files (1,079 installed files totaling 470,164,222 bytes and the 27 CD2 files), 528 source pins, and 867 compiler pins matched before and after. The EXE/KEY/INI identities below are unchanged. The run reused the retained native installation directly and presented the same original CD2 URLs under read-only DRIVE_CDROM D:, preserving the original C: paths. No reinstall, download, full fixture copy, marker/volume/KEY/INI manufacturing, BIF expansion, save injection, WININIT edit, or foreign cleanup occurred.

The exact executed response callback passed 14 cheap cases before launch: optional HTTP404s, same-origin HTTP200 script, isolated/unisolated/missing document, data/blob URLs, required asset/script/module HTTP failures, and module hash mismatch. The scratch server additionally passed actual 200/400/404/416/500 requests with consistent COOP/COEP headers. It applies headers before the static handler's success/error branches. In the browser, crossOriginIsolated=true, normal Worker/Threads and WebGL were verified; the previous optional404 guard failure did not recur. Audio/clock remained default, with a native 1024×768 viewport. No FPS or audible-output claim.

Heavy work waited for the current SlaveZero producer 01a193e2-c9b5's 21:43:06 NEWFULLRELEASE, then independently checked actual PID/start identities, whole Chrome/global crashpad, profiles, ports, inputs, and the inherited FD9 lock. Runtime Git HEAD was independently captured as 5efa420458517d3eacbab60f27946e9cdef80a8d. The actual HTTP-loaded module was independently hashed: HTTP200, no cache, 1,898,762 bytes, SHA-256 d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8.

The sole browser ran approximately 21:43:11–21:44:51. At the first Create Game request, the scratch helper treated the DirectDraw OffscreenCanvas as a DOM canvas and threw TypeError: c.getBoundingClientRect is not a function. The mouse never moved or pressed; input evidence contains only the four trusted, balanced Escape key edges. The reviewed browser-failure.png and browser-final-before-teardown.png genuinely capture the labelled menu before teardown. executed-browser.cjs preserves the actual driver. The retained browser.cjs now guards the optional DOM method and falls back to the displayed renderer canvas and native client rectangle. Cheap tests extract the exact helper body: oldFAIL1/newPASS3 for OffscreenCanvas, DOM canvas, and absent native layer. This geometry correction has not been browser-executed.

The controller has a durable wait with exit 1, honestly preserving the harness failure. Supervisor, runtime launcher, board watcher, and monitor have durable successful waits; six adopted children were reaped. Chrome exited 0 without a signal. Final audit found all 23 own PID/start identities absent, whole Chrome/global crashpad empty, ports 43113/46145 closed, profile gone, inputs released, and FD9 free. Minimum observed runtime free space was 2,668,838,912 bytes. Boat ID/state/expiry N/A. Runtime release was announced separately from pending documentation publication.

Recovery: RECOVERY.txt, original-optical-runtime.json, first-causal-failure.json, header-guard-tests.json, layer-geometry-tests.json, runtime-git.json, the actual HTTP module receipt, and full-release-final.json. The next authorized task should cheaply verify the native canvas coordinate helper, reuse this exact original optical presentation, and perform native character creation, Prologue DONE, then paired unpaused floor-click movement. Do not rerun Setup or change the original KEY/INI to address a scratch geometry error. The run's 44 artifact references and four screenshots pass existence and containment checks; notes pass git diff --check. The global checker remains non-green with 9 historical containment errors, 200 missing references, and 114 invalid records; none names this run.

2026-10-31 original optical scan identified; runtime stopped by harness

Run 20261031T2126Z-icewind-disc did not establish original-demo startup or gameplay. Static analysis found a missing prerequisite in the preceding C-only presentation, and a cheap VFS test passed with the original CD2 tree on D:. The sole browser attempt then stopped during asset loading because this run's scratch HTTP response guard incorrectly rejected optional 404s. No runtime disc-scan result, menu, character creation, Prologue DONE, tavern, or floor-click movement is accepted. There was no retry or emulator change.

The unchanged original EXE's scan at 0x0043ACE7 calls GetLogicalDrives at 0x0043AD62, tests each assigned letter with GetDriveTypeA at 0x0043ADF6, and accepts only type 5 (DRIVE_CDROM). It calls GetVolumeInformationA(root, NULL, 0, NULL, NULL, NULL, NULL, 0) at 0x0043AE5F. After a successful volume probe it concatenates the root, cd, digit 2, \data\iwdcd., and digit 2. Its call at 0x0043AF05 reaches the MFC finder at 0x00929EC7, which calls FindFirstFileA at 0x00929EFF. Thus this scan requires D:\cd2\data\iwdcd.2, not D:\data\iwdcd.2. Disassembly and import/string references are retained in optical-scan.asm, find-wrapper.asm, and static-imports.json. This is a static necessary-condition diagnosis; the failing original guest call has not yet been observed dynamically.

The predecessor supplied the 27 original CD2 files only at C:\CD2\. C: reports fixed disk and is excluded by that scan; its original marker does not satisfy the optical path. The new private app entry additionally presents those same retained URLs below D:\CD2\, with D: recorded as type 5 and read-only through the existing generic VFS interfaces. The installed C: paths and native INI aliases remain present. No ISO, volume label, serial, marker, KEY, INI, resource, save, or decompressed BIF was manufactured. The prerequisite test verifies the old missing-path failure, drive enumeration/type, root enumeration/chdir, original one-byte marker 32 hex, successful marker lookup, and write-protect error 19. These checks do not prove guest startup. No generic implementation violation was shown, so no production API change or build was justified.

All 1,106 retained input files were freshly size/hash checked: the 1,079 installed files totaling 470,164,222 bytes plus all 27 original CD2 files. The original EXE/KEY/INI pins recorded below remain unchanged. All 528 accepted source and 867 compiler pins also matched before and after. No reinstall, download, fixture copy, recovery inventory, WININIT edit, or foreign cleanup occurred.

Heavy work waited for the fresh corrected Nox jump/action task's 21:30:16 NEWFULLRELEASE, followed by an independent actual gate. The sole attempt ran at approximately 21:30:42–21:30:44 with default Worker/Threads, WebGL, audio/clock behavior, isolation, and a native 1024×768 viewport. Crucially, actual-http-loaded-module.json hashes the actual browser response body: HTTP 200, no cache, 1,898,762 bytes, SHA-256 d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8. The runtime Git HEAD was not independently recorded; source/compiler and loaded-module identities are byte-pinned instead.

The scratch listener wrongly required COOP/COEP headers on every HTTP response. The optional build-info.js, favicon, and font-bundle 404s lack those headers and set its closing flag. The inspected pre-teardown browser-first.png shows the download window; its state has no guest windows, running=false, and an empty drive trace. The queued Escape was never processed; trusted-input evidence is empty. Exact guest entry-point execution is unproven. There is no additional failure screenshot. The executed driver is preserved separately as executed-browser.cjs. The corrected scratch guard checks document isolation, retaining the independent mandatory loaded module check; eight focused cases pass, including four failures of the old guard. The corrected driver has not been browser-executed.

Supervisor, controller, board watcher, and monitor have durable successful waits; six adopted children were explicitly reaped. Chrome exited 0 without a signal. The final audit found all 20 recorded own PID/start identities absent, whole Chrome/global crashpad empty, ports 39977/37637 closed, profile gone, inputs empty, and FD9 free. Minimum observed free space was 2,802,823,168 bytes. Boat ID/state/expiry N/A. No FPS/audio/gameplay claim.

Recovery is in this run's RECOVERY.txt, corrected browser.cjs, mount manifest, prerequisite/guard tests, and full-release-final.json. Reuse the retained native installation; do not rerun Setup or legacy preparation. Next requires a new authorized runtime allocation using the corrected guard and original D: presentation, then an actual optical-call trace before native character creation, Prologue DONE, and paired unpaused movement acceptance. The 42 run artifact references pass schema, existence, and containment checks; the notes pass git diff --check. The global evidence checker still reports 9 historical containment errors, 200 missing references, and 114 invalid records, with no issue naming this run.

2026-10-31 original native installation complete; browser reports No Disc

Run 20261031T2058Z-icewind-install completed a fresh ordinary installation from the retained authenticated standalone ZIP. The inspected Setup Complete screen says installation finished; its native Finish button produced guest [Exit] code=0, followed by an explicit successful Popen.wait. The installed destination contains 1,079 files totaling 470,164,222 bytes. No legacy preparation, host cabinet/BIF expansion, filtered KEY, rewritten configuration, save injection, or emulator change was used. The prepared fixture, authored state, and foreign WININIT prototype were not changed or reinventoried.

Before extraction, the ZIP's 645,596,503 bytes, SHA-1 cc6345ae7decb43341bde25bb1302f422a42666c, and SHA-256 992b3d9054ff089f40dba40debd0fe902117c1a21f80d7e34678ae3313c9ff80 were freshly verified. All 56 extracted files passed ZIP CRC checks and were hash-pinned without transforming their contents. Heavy work waited for the current Nox direct-route successor's 21:03:51 NEWFULLRELEASE, then independent actual PID-start, whole Chrome/global crashpad, profile, ports, and FD9 checks. That predecessor explicitly lacked durable wait receipts; this run checked actual absence instead of treating the old 20:43:34 release as sufficient.

The documented bootstrap/replay commands below were reused with paths under this run: original-source/icewind_dale_eng_demo/Setup.exe, bootstrap/, and installed-vfs/. The original Setup emitted a 111-file capture and its 557,056-byte _ins5576._mp, SHA-256 a8657371f03e2e66db951c3dcd3aeb42c576894908ca2eb1b3806aa0404cb083. The engine and both emitted seed DLLs are pinned in final-check.json. Only SHA-identical source/capture files were hardlinked to keep export space available; the guest still decompressed its original cabinet itself.

The native Welcome, License, Recommended, CPU Speed, Destination, and Program Folder choices retained their defaults. At batches 6,100, 12,100, 20,100, and 24,100 the inspected copy screens showed 24%, 41%, 64%, and 86%. The post-copy DirectX 7.0a prompt was declined with its native No command, ID 7. At batch 30,201, 10-after-optional-directx.png shows Setup Complete. The attempt to clear Readme with dlg-input-click:501 did not toggle its check; the retained control dump and 11-finish-options.png show it still checked, with Run Game unchecked. Finish ID 1 at batch 30,213 requested the Readme shell open and exited normally. No successful Readme launch is claimed. The controller reported “guest exited while command pending 0” because natural guest exit interrupted its requested 50-batch step; the guest Exit log and durable exit-0 wait distinguish this from a deadline or synthetic success.

Original installed file Bytes SHA-256
iddemo.exe 6,283,264 b94816d10029cb99c0315f175330c917be2fff298394e853e2764973d8d13af4
chitin.key 216,476 c04920810d6663f6e23779c16dadbed88408b4a0528fd79c686d81fd535f6292
icewind.ini 449 eec07e40a941ea22c9ebf19631250bac021ecc57cdb8b251b3506b2c004be9fb

The installer itself finalized the earlier 316-byte relative INI to HD0:=C:\Program Files\Black Isle\Icewind Dale Demo\, CD1:=C:\Program Files\Black Isle\Icewind Dale Demo\CD1\, and CD2:=C:\CD2\. These are guest-written bytes, not a host correction. All 27 exported CD2 files match the authenticated source sizes/SHA-256; 246 installed override files and the entire installed closure are pinned. EXE/KEY match their retained original pins. The new output manifest is a local pin with native-installer provenance, not an independently compared upstream installed-tree manifest.

The optional fresh Chrome 151.0.7922.108 probe served those original installed and CD2 files at their exported guest paths, with an isolated profile, default Worker/Threads, WebGL, audio/clock behavior, and trusted balanced Escape inputs. A private app-registry response changed only asset URLs, guest paths, and working directory; no tracked registry or guest bytes were edited. The native log reported “No Disc Detected. Game Aborted.” at 21:15:59. There was no accepted menu, character, Prologue DONE, tavern, or movement, and no retry or save injection. Later gray frames are not gameplay. The next investigation is the original CD2 optical-drive presentation and native disc scan, preserving this KEY/INI/export. Do not assume that another INI rewrite or a prepared KEY is required.

The accepted canonical module file stayed d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8; all 1,395 accepted source/compiler pins matched before and after. No build was needed because no emulator implementation changed. The browser-loaded module hash was not independently captured. All installed closure hashes still matched after the probe. No FPS or audible-output result is claimed.

Space recovery and an ownership mistake

Safe net reclaim was 2,013,306,880 allocated bytes after correction: lossless gzip of stopped terminal transcripts and released completed-run data, plus 277 SHA-verified hardlink replacements of identical retained WASM files. Per-file original/decompressed SHA-256 and size pairs, original timestamps/modes, and hardlink sources are retained in the run's receipts. RECOVERY.txt and recover-transcript.py restore a selected compressed original with hash verification, retain the gzip, and require more than 2.1 GB free after restoration. Hardlinked evidence must be detached by verified copy/replace before future editing. No original package/fixture was compressed or removed. Minimum recorded free space during compression was 2,729,975,808 bytes; final runtime release had 2,873,475,072 bytes free.

The initial release filter incorrectly matched NEWFULLRELEASE as well as explicit RELEASE. It compressed the active Nox profile's ActorSafetyLists/9.8200.9609/listdata.json, a scope violation. Its original and decompressed size/hash matched before replacement, but Nox's subsequent own profile cleanup removed the directory and gzip, so that temporary file is no longer recoverable here. No foreign process was signaled; the cause of Nox's separately reported supervisor exit 143 remains unknown. The full mutation audit and correction were announced on the board. Forty-eight older files with uncertain release status were restored and hash-verified, with their gzip archives retained. The filter was corrected to explicit release records; restored storage and the profile incident are excluded from the safe net reclaim above. Eighteen tracked evidence originals were also restored byte-identically, retaining their gzip archives; their storage is deducted from this final total. Guarded operations also stopped on an absent target and a changed-size file without modifying those files.

The single combined runtime lasted approximately 10.5 minutes, within the 17-minute target. Extraction, bootstrap, engine, installed audit, and browser controller have durable successful child waits; supervisor tool sessions also returned 0. At 21:19, all 23 recorded own PID-start identities were absent, whole Chrome/global crashpad was empty, ports 34721/40747 were closed, the profile was gone, inputs were released, and FD9 was free. No runtime signals were needed. Boat ID/state/expiry N/A. This NEWFULLRELEASE handed the runtime to the waiting Nox successor; it is not a claim about later foreign processes.

Recovery starts with this run's retained installed-vfs/, not the legacy prepared fixture or the old partial exports. runtime.py, extract-original.py, and the bootstrap launch.json preserve the exact native install route. installed-all-files.json, installed-closure.json, original-cd2-joins.json, completion-reviewed.json, engine.log, browser console/input records, and full-release-final.json retain the evidence. The run's 31 artifact references pass existence/containment checks, and the notes pass git diff --check. The repository-wide evidence checker remains non-green (9 containment errors, 200 missing references, 114 invalid records); none names this run. Compressed historical originals require the documented restore recipe before consumers expecting their former raw paths can read them. Next: qualify the original CD2 browser mount, then require native character creation, Prologue DONE, and paired unpaused floor-click movement images.

2026-10-31 original package recovered; installation awaits disk space

Run 20261031T2045Z-icewind-recovery recovered the exact standalone original package into its own ignored directory, without running the legacy fetch or preparation recipe. The prior 8f72fa264 original-install absence findings were reused; the prepared fixture was not inventoried again or modified. No installed export was recovered and no guest ran.

The retained private file is scratch/runs/20261031T2045Z-icewind-recovery/icewind_dale_eng_demo.zip:

Check Result
Size 645,596,503 bytes, matching Archive metadata
SHA-1 cc6345ae7decb43341bde25bb1302f422a42666c, matching the candidate manifest
SHA-256 992b3d9054ff089f40dba40debd0fe902117c1a21f80d7e34678ae3313c9ff80
ZIP directory 60 entries, 679,877,981 uncompressed bytes; no absolute or traversal paths
Original Setup.exe 73,728 bytes; SHA-256 5656e87da0641c9dcfcd0ee8949ce72b3fa6a7d0e8b1fd985a16f6bd6c34ce52

Archive metadata and the source HEAD were checked before transfer. The public download URL returned HTTP 500 before any bytes were received. The storage host named by that item's metadata returned HTTP 200 and the expected length: https://ia600609.us.archive.org/33/items/icewind_dale_eng_demo/icewind_dale_eng_demo.zip. The successful transfer computed both hashes while writing a unique partial file, fsynced it, and renamed it only after the expected size/SHA-1 matched. Reading Setup.exe, README.TXT, and DirectX/dxsetup.exe from the ZIP also passed their ZIP CRC checks; this is not a full cabinet extraction or install test.

The transfer started with about 3.40 GB available and retained at least 2,758,266,880 free bytes during transfer. At the extraction decision, 2,758,172,672 bytes remained. Full source extraction alone would leave 2,078,294,691 bytes, below the requested 2.1 GB early-stop threshold and 2.5 GB reserve. Source plus the historical 470,164,222-byte installed game would leave only 1,608,130,469 bytes, before bootstrap capture/export overhead. Consequently no source extraction, unshield operation, installer, build, CLI guest, or browser was launched. The authenticated ZIP is the completed checkpoint; original installed KEY/INI/CD2 closure and browser gameplay remain unverified. Existing prepared assets/authored state and foreign WININIT WIP were preserved.

Next: make at least another 2 GB available without deleting foreign work, then extract this authenticated ZIP into a new ignored source directory. Use the documented native Setup.exe bootstrap and guest-emitted _ins5576._mp replay below, replacing all historical /private/tmp outputs with unique run-local paths. Budget for the source, copied bootstrap capture, and installed export together. Require actual Setup Complete, guest exit 0, and hashes of the original installed KEY, installer-written INI, complete CD2, override set, and served closure. Do not present cabinet extraction as an installed export. Then apply the current browser-release/actual-process/FD9 gate before the native character → Prologue DONE → unpaused tavern route and trusted balanced floor click with paired screenshots. No loading, menu, paused HUD, or historical screenshot is new gameplay evidence.

Evidence: source-metadata.json, source-routing.json, direct-source-head.json, package-authentication.json, zip-directory.json, checkpoint.json, and the two recovery scripts in this run. No runtime lock, profile, server, held input, or test box was created; Boat ID/state/expiry N/A. The current Nox NEWFULLRELEASE was observed, but no runtime gate was consumed.

2026-10-31 original-browser prerequisite audit: blocked before launch

Run 20261031T2040Z-icewind-browser inspected main input revision 6ae0e926a61e7e6e0edb3d099c9a0ebaf16e91fc. No new browser gameplay or movement acceptance was obtained. The existing local fixture is the legacy prepared asset set, not the intact original installation described below. No browser, CLI guest, build, or acquisition was launched.

Cheap verification recorded SHA-256 and sizes for all 920 local files (903,400,058 bytes). The registered icewind_dale_demo EXE plus its 919 file entries all exist; every declared registry size matches. This is a local inventory pin, not a new upstream-package comparison. The actual EXE is 6,283,264 bytes, SHA-256 b94816d10029cb99c0315f175330c917be2fff298394e853e2764973d8d13af4, matching the existing acceptance pin. Asset identity prevents treating this as an original-install route:

Before considering another build, 1,151 preceding accepted build-input pins from 20261031T2026Z-nox-browser were compared: 1,150 match. Only index.html differs while the current Nox worker owns its selector repair. The existing WASM SHA-256 remains d39f362187ce89abda47fa53bca6bb6576badd8d21b78e44c6fbad973d1636d8. This is a file/source comparison, not proof of a browser-loaded module. No emulator failure was demonstrated and no implementation/test paths were edited. Foreign lib/wininit-boot.js and test/test-wininit-boot.js remain untouched.

Recovery: recover the intact private guest-installed export, including its original KEY, installed INI, full override set, and original CD2 layout, or recover and authenticate the standalone original demo package and complete its ordinary installer into a separate context. Preserve the existing fixture and authored state. Do not reconstruct a key by reversing the legacy filtering or inject saves/configuration to satisfy the route. Pin every file actually served and use the existing original-tree import path before diagnosing a loader bug. The manifest package SHA-1 remains cc6345ae7decb43341bde25bb1302f422a42666c; it was not checked against newly acquired bytes in this run.

Only after that prerequisite is met, wait for CURRENT nox-candidates-browser-gameplay-main-oct31 NEWFULLRELEASE (or its explicit terminal no-runtime release), then independently check FD9, whole Chrome, global crashpad, PID starts, profiles, and ports before heavy work. This audit did not consume that dependency or acquire FD9; no runtime/profile/port or held input was created. Boat ID/state/expiry are N/A.

The next acceptance remains the ordinary native menu → authored character → Prologue DONE → unpaused tavern, with reviewed before/after screenshots of the same character moving to a distant floor click under trusted balanced DOM input. Party Formation, loading, a paused HUD, NO DISC, and red log glyphs do not satisfy it. Native save evidence counts only if the guest writes it. Evidence is in the run's registered-files.json, fixture-inventory.json, prep-summary.json, source-parity.json, and lifecycle.json; no screenshot exists because no guest ran. Historical CLI movement below remains historical.

Legacy fetch and test (the fetch recipe requires unshield; this is not evidence that the original installer works):

node tools/fetch-candidate-corpus.js --id=icewind-dale-demo
node test/test-icewind-dale-demo.js

Original installer investigation

Latest acceptance: original-installed gameplay and quicksave

The local FindFirstFile correction below passes a fresh launch of the clean /private/tmp/iwd-profile-fixed-installed-vfs, not the diagnostic cache export or legacy extracted fixture. Original EXE, INI, KEY, CD2 files, and override resources were retained. No guest memory patch or host-expanded archive was used. The game recreated its own cache during this run.

The single low-priority headless process used --no-build --no-threads, --control-stdin --frozen, --batch-size=200000, --time-scale=10, --repaint-every=10, and the internal --max-seconds=900 guard. The game was advanced in short stdio steps and explicitly quit at batch 1996, exit 0. This is a functional result, not a performance measurement.

Inspected checkpoints:

The native mpsave/000000001-quick-save contains icewind.gam (3520 bytes, GAMEV1.1, contains codex), icewind.sav (1670, SAV V1.0), worldmap.wmp (8072, WMAPV1.0), icewind.bmp (23462), and portrt0.bmp (1678). All five were exported and checked under /private/tmp/iwd-original-native-saves/program files/black isle/icewind dale demo/mpsave. The default session also contains codex in its 2416-byte GAM. This supersedes the older claim that the numbered quicksave path was unavailable.

Fresh-process reload is now verified as well. The same CLI command mounted the clean original installation first, then the full native save export with a second --vfs-tree. No cache files were carried over. At batch 330, /private/tmp/iwd-reload-list.png shows the numbered quicksave, its thumbnail, and the character portrait. Load at (541,102) causes the guest to expand its original CD2 archives again. A nearly stationary loading bar is not evidence of a hang: the AR100A output handle advanced from 5,004,288 to 5,410,816 bytes between batches 1230 and 1530, and AR100D was being written by batch 2330.

Batch 2830 reaches Party Formation with codex. Done at (552,432), batch 3030, restores the tavern at batch 3130, with the character at the saved position near (500,300): /private/tmp/iwd-reload-tavern.png. Clicking (452,330) and stepping 40 moves the character there, visibly confirmed in /private/tmp/iwd-reload-move.png. The historical red Paused for saving game log line remains visible but does not prevent this movement. The process explicitly quit at batch 3170, exit 0. Shutdown still reports a held main-thread critical section at 0x006c8f94; this test does not establish that all synchronization cleanup is correct.

Remaining limits: browser acceptance of this original VFS is not yet verified. The Prologue narration body is still blank. Earlier shutdown diagnostics report one abandoned parked EnterCriticalSection and one held main-thread critical section; successful gameplay does not prove that scheduler issue fixed. The legacy acceptance script below still mounts its modified fixture and must not be cited as the original-install regression.

Original-install browser probe: not accepted

The temporary /private/tmp/iwd-original-browser.js serves the same original installation and native save export through range-backed providers. It uses headless Chrome, cross-origin isolation, frozen stdio stepping, and a 900s internal deadline. It does not mount the legacy fixture. The first harness attempt omitted profile preparation and stopped on VfsPendingError for icewind.ini. Calling the existing mediaImport.materializeIniFiles helper before launch corrects that harness omission without changing file contents.

The subsequent probe still does not establish browser gameplay. After a relaunch without premature Escape input, it creates JigSawedME and opens the GUI archives and data/mvefilel.bif, but the captures at frozen steps 450, 750, and 910 remain gray. The window has no _dxFrameLayer at that checkpoint. Main EIP is 0x008b68c5; four cooperative secondary threads exist, with final EIPs 0x005b3344, 0x00437dd3, 0x00437d49, and 0x0043d3eb. Escape at steps 750 and 830 did not expose a menu. The last capture is /private/tmp/iwd-original-browser-menu.png (despite the filename, it is not a menu). The browser and its temporary server were explicitly closed, exit 0. Rendering, intro progress, and input routing still need to be distinguished before assigning a runtime cause; no browser fix is claimed.

Original first-area load: false local file matches

On the Lobby3A build, the clean original-installed VFS completes character creation through Sound and Name. The inspected /private/tmp/iwd-original-party-ready.png at batch 695 shows the created codex fighter. Accepting the party enters Starting Game, then the guest itself expands the four original area CBFs into cache/data:

Archive Expanded Bytes SHA-256 of Guest Output
AR100A 6613876 4134c41a68425cb9a6c98fb5cbb146d6be9aa883aa76e626acb1c7a22ca330a6
AR100B 8898048 1e06cd252940e44040b55c48feae9ed72c0ca52491f71a2c723fbbf8ae269328
AR100C 6199168 87b4f35baba9403c1a5a0efb8f80fe2ed2b43efd604766522533609b5cbd05a1
AR100D 5072154 5bc21fcd750516e2e3f1d97bca9bb7fb9382c3de8fd24736ac6af4020926a648

A read-only, in-memory zlib comparison found every byte equal to the original CBF payload's expansion. No reference output was mounted or substituted. This rules out decompression corruption for these four files.

The run subsequently asserted at Infinity.cpp:1763, "Demand for WED file failed"; the inspected screenshot is /private/tmp/iwd-original-first-load-1245.png. The guest stack returns to 0x006559cb; resource object 0x4ff6c1d0 has locator 0x03600004. The original KEY resolves that locator to AR1006.WED in BIF 54, data/AR100B.bif. Its expanded file contains the valid WED V1.3 resource at offset 23108, length 6586. The failing BIF object's stored path instead names the nonexistent override/data/ar100b.bif.

VirtualFS.findFirstFile was allowing exact C-drive probes in absent directories to match the same basename anywhere in the VFS, including the real cache file. createFile already disallowed that fallback on C:, so the two APIs disagreed about the existence of the override path. The new regression fails before the correction; both APIs now reject missing local paths while retaining exact cache access and the existing non-C media fallback. VFS (30), lazy-entry (34), overlay (19), and ANSI/OEM filename tests pass. This is a general path-lookup correction, not a game-specific resource substitution.

Earlier Lobby3A checkpoint: character generation reached

The targeted trace after the DP4 fix identified a second rejected interface: IDirectPlayLobby2_QueryInterface at return address 0x008d3796 requested {2db72491-652c-11d1-a7a8-0000f803abfc} (Lobby3A), not another DirectPlay4 interface. Its E_NOINTERFACE path released the lobby and abandoned setup.

Lobby3A now upgrades the same ANSI object to the SDK's 19-slot layout, preserving all 15 inherited slots and appending ConnectEx, RegisterApplication, UnregisterApplication, and WaitForConnectionSettings. The signatures follow the Wine DirectPlay lobby header. These four new operations remain explicitly unsupported, with argument validation and no fabricated connection or registration success. Unicode Lobby3 remains rejected. This is ABI support, not full lobby functionality.

The real-thunk test-directplay4.js regression passes for both DP4 and Lobby3A, including identity, inherited slots, worker registry restoration, HRESULTs, output clearing, and stdcall cleanup. The inherited lobby-address suite also passes. Full build gates pass: native 1,076,563 bytes, compatibility 1,077,027 bytes, layout fc765043612db8cc.

A fresh low-priority, frozen CLI run used the original install's complete /private/tmp/iwd-profile-fixed-installed-vfs, without substituting the EXE, INI, KEY, or compressed archives. The trace now reports Lobby3A QI S_OK, followed by IDirectPlay3_EnumConnections. Inspected captures:

The process was explicitly quit at batch 390, exit 0. This fixes the observed Create Game blocker but is not gameplay acceptance. Next complete a character and verify first-area loading, unpaused movement, and native session saving using this original-installed VFS. Older acceptance below used the modified legacy fixture and cannot establish those results for the original install.

Earlier DirectPlay4 checkpoint

The pre-DP4 rejection recorded below is now fixed in compiled factory tests. QueryInterface upgrades the same ANSI object to a generated 53-slot vtable, preserving its 47 inherited entries and appending the six SDK tail methods. test-directplay4.js loads a Win32 PE and calls the generated thunks through the emulator dispatch loop, checking slot IDs, HRESULTs, output buffers, and stdcall stack cleanup. It also checks restoration of the appended thread vtable registry entry. The factory/identity suite passes 31 cases.

Synchronous local SendEx, send/receive queue queries, and cancellation of pending queue entries are implemented. Async production/completion and public group-ownership policy/notifications remain explicitly unsupported; both ownership methods preserve state/output rather than report false success. Cancellation tests inject pending entries and do not claim an async transport. The earlier internal ownership storage is not a completed public contract.

The full build passed (native 1,076,220 bytes, compatibility 1,076,684 bytes; layout d36314e0fcee18b2). A fresh frozen CLI run used only the original installer's /private/tmp/iwd-profile-fixed-installed-vfs, with the same EXE/cwd and no INI/KEY substitutions. The inspected menu at batch 300 is /private/tmp/iwd-dp4-menu.png. Clicking Create Game at (480,175), then stepping 100 batches, still produces Cannot connect to the game session; the inspected capture is /private/tmp/iwd-dp4-create-game.png. The process was explicitly quit at batch 400. This is not gameplay acceptance, and the new result disproves treating the missing IID as the sole session blocker. Next is a targeted COM/DirectPlay trace beyond the now-supported activation request. Older checkpoints below describe their respective builds.

Decoder overlap correction

The long install's exit counters reported roughly 101 million retired blocks but only 127,641 invalidation calls. The retirement counter also counts page_publish replacing overlapping entries; it does not prove that the guest modified its code. The hot cabinet loop at relocated 0x007b4d21 has ordinary interior branch targets. Alternating an outer entry with an interior entry made the one-owner-per-byte cache continually retire the other decode.

decode_block now emits an ordinary block end when it reaches an already cached instruction entry, preserving that suffix instead of overlapping it again. No guest address or cabinet format is special-cased. The focused regression in test-sparse-generated-code-cache.js failed before this change because the interior entry vanished after re-entering the outer prefix. It now requires both entries to remain cached while alternating them, and retains the subsequent shared-immediate rewrite check. Sparse and cross-instance invalidation, page-chunk allocation, and 138 x86 cases pass with this correction.

The actual clean installer run on this build passed inspected copy screens at batch 8750 (30%) and 18750 (56%). Its next large stdio step did not return a checkpoint for a long interval: a live process and CPU activity were not sufficient evidence of installation progress. Final output did establish a new stopping point at batch 27726: the installer left its copy-progress UI and trapped on the still-unimplemented WritePrivateProfileSectionA. The caller is 0x0041d366, returning to 0x0041d36c in the emitted InstallShield engine. Process exit was 1, not Setup Complete. Implement and test the real section-writing semantics before another full installer acceptance; do not silently return success. Use shorter step requests or a separately accessible control channel so progress and orderly-stop commands are not queued behind a long step.

WritePrivateProfileSectionA/W are now implemented in db04f9a8: section replacement/deletion writes the VFS file, retains an existing UTF-16LE file's encoding, and reports invalid parameters, missing parents, and read-only write failures. Single-key profile writes preserve the section-written keys. The storage regression and full build pass. test/test-profile-section.js also exercises both compiled WAT handlers with guest pointers, Unicode text, NULL deletion/flush arguments, BOOL/LastError, and three-argument stdcall cleanup. The original-installer acceptance below now passes; gameplay using the installed game's original INI/CD layout remains unverified.

Completed original installation

The clean guest-produced engine replay on the integrated profile-fix build (1,073,129-byte WASM, layout f73bfdc3f7f38137) completed the original installation. It passed the former profile-section trap, created shortcuts, and displayed the compatible-DirectX reinstall question at batch 27750. Answering No with dlg-cmd:7 reached the actual Setup Complete screen at batch 27800. Finish (ID 1) led to guest [Exit] code=0 and a successful VFS export at batch 27840. This is not a deadline-only or copy-progress acceptance.

The separate output is /private/tmp/iwd-profile-fixed-installed-vfs. Its program files/black isle/icewind dale demo subtree contains 1,079 files totaling 470,164,222 bytes. Read-only comparison with the existing installed-extracted/Recommended_compressed reference found 1,076 identical files, one changed icewind.ini, and two installer-created extras: readme.txt and uninst.isu. All 640 sound-set WAVs are present. The original CHITIN.KEY is byte-identical to the reference's original key, not the legacy modified CHITIN-full.KEY. The installer itself wrote these aliases:

[Alias]
HD0:=C:\Program Files\Black Isle\Icewind Dale Demo\
CD1:=C:\Program Files\Black Isle\Icewind Dale Demo\CD1\
CD2:=C:\CD2\

No host decompressor or INI/KEY patch supplied the installed payload. The export also retains the original cd2 source tree. Next launch must mount this entire VFS and use the installed executable's guest path and working directory, retaining the original CD2 layout and compressed resources. Do not substitute legacy expanded BIFs or the modified key as gameplay proof.

Inspected screenshots: /private/tmp/iwd-profile-directx-question.png and /private/tmp/iwd-profile-setup-final.png. The latter shows Setup Complete. Replay used --max-seconds=3600 --control=8137 --control-stdin --frozen. This terminal closed stdin, so all wizard actions and checkpoints used HTTP. Long step requests can exceed HTTP's 30-second response deadline while the guest keeps running: inspect /snapshot credits until zero before issuing another step. Independent file-size checks established continued progress. The process priority was lowered during high host load; no benchmark was run.

Two follow-up observations are not covered by this install acceptance: the Readme checkbox input issued a ShellExecute request before Finish, and VFS export reported an empty c:\windows file/directory collision, preserving the file as windows.__vfs_file__. Neither caused a setup error, but both deserve separate investigation rather than being silently treated as correct.

Original-installed game startup

A low-priority headless run mounted the complete exported VFS, loaded program files/black isle/icewind dale demo/iddemo.exe, and set both its guest executable path and working directory to that installed directory. No INI, KEY, CD2, override, or compressed resource was changed or removed. With --batch-size=200000 --time-scale=10 --repaint-every=10, Escape at batches 100, 160, 220, and 280 skipped the intro sequences. The inspected /private/tmp/iwd-original-game-later.png at batch 285 shows the full menu with readable labels, not just background artwork.

Clicking Create Game at (480,175) at batch 300 instead produced Cannot connect to the game session. The inspected error screenshot is /private/tmp/iwd-original-party.png at batch 400; despite its filename it does not show Party Formation. The run was explicitly quit at batch 450. This differs from the legacy modified-fixture acceptance and remains the next gameplay blocker. The error text alone does not establish whether DirectPlay address creation, COM activation, or another step failed.

A subsequent selective DirectPlay/COM API trace stopped at the internal 180-second deadline at batch 200, before the Create Game click. It provides no failing session API evidence. Note that --trace-api=Names also enables unfiltered file-read/find diagnostics through _debugReadFile and _debugFindFile, so even a nominally selective startup trace emits extensive resource I/O. Next investigation should enable the relevant tracing at the menu or suppress those independent file diagnostics while retaining the DirectPlay/COM trace, then compare with the legacy route on the same build.

The follow-up frozen trace identified a concrete interface gap during menu initialization, before Create Game: at 0x008d3680, the executable calls CoCreateInstance(0x00997c50, NULL, 1, 0x00997c40, out) with return address 0x008d3686. Decoding the original executable's GUID bytes gives CLSID_DirectPlay and IID_IDirectPlay4A {0AB1C531-4745-11D1-A7A1-0000F803ABFC}. The caller stores the HRESULT at [ebp-0x1c] and tests it after CoUninitialize.

The current dplay_query_interface_wa intentionally rejects this IID because only the 47-slot ANSI DirectPlay2/3 vtable exists. The compiled test-directplay-query-interface.js now reproduces the exact factory request: E_NOINTERFACE, cleared output, and no leaked temporary object. Its passing negative test describes a missing feature, not working IWD gameplay. The legacy fixture's earlier success does not establish that it still works on this newer COM implementation; the original-installed error should not be attributed to INI/KEY/CD layout without further evidence.

At that checkpoint, the next step was a 53-slot ANSI DirectPlay4 implementation. Its additional methods are group-owner get/set, extended send, queue inspection, and cancellation by message or priority, as defined in the Wine DirectPlay header. Simply accepting the IID on the shorter vtable would permit calls past its end. Ownership, queue state, cancellation, reference lifetimes, and all tail stdcall contracts need coverage before changing the current rejection test. The trace run was explicitly quit at batch 350; its inspected session-error capture is /private/tmp/iwd-original-dp4-error.png.

Implementation has started with an internal ownership-storage prerequisite: the local entity entry now retains an explicit owner ID. Assignment requires a live group and a live player; transfer replaces that ID, player destruction invalidates references to it, and group destruction/Close/slot reuse discard old ownership. Compiled entity tests cover those transitions and failed assignments preserving prior state. This is not yet GetGroupOwner or SetGroupOwner API support: default selection, membership policy, ownership notifications/migration, per-object session isolation, and public HRESULT contracts still need to be established. No DP4 IID or additional vtable slot has been exposed by this prerequisite.

Internal message storage now retains copied payloads, owner keys, sender and recipient IDs, unsigned priorities, and non-recycled message IDs. Queries count messages/bytes or find the oldest match independent of slot reuse; cancellation is owner-scoped and separates send from receive entries. Storage is bounded to 64 entries, 1 MiB per payload, and 4 MiB total. Close clears it even when no entity table has been allocated. The compiled test-directplay-message-queue.js covers those invariants, ID exhaustion, and copies spanning non-contiguous guest-page backing. Receive now consumes received entries through the public handler, with buffer-size negotiation, peek, FIFO, explicit sender/recipient filters (including system sender zero), and sparse destination copies. GetMessageCount reports received entries for the caller's object. Final COM Release removes that object's messages without removing another object's queue; non-final Release retains them. The compiled test exercises those handlers and their stdcall stack cleanup. Receive's copy/peek/filter behavior follows the Wine receive implementation. The storage tests inject received entries directly. A separate compiled test-directplay-send.js now exercises public Send-to-Receive delivery: local unicast, broadcast and direct group membership, excluding the sender. Each recipient gets copied bytes and its borrowed receive event is signaled only after all copies exist. Capacity failure rolls back this send without discarding prior messages or signaling events. Unknown flags fail, while streams, security and asynchronous modes explicitly return unsupported. Entity entries retain creator-object and event fields; sends require a local sender belonging to the caller and do not silently join unrelated objects. Close and final Release retire only the caller's entities/messages. Destroying a recipient discards its unread messages but destroying a sender does not retract already-delivered messages. Tests cover those transitions, two-object cleanup, capacity rollback, copied bytes, event ordering, and recipient slot 31. Provider/session initialization, network/session joining, anonymous sends, SendEx and asynchronous completion still need implementation. Existing entity enumeration/mutation methods also need a full session-isolation audit. The DP4 IID remained rejected at that checkpoint; see the latest result above. No new gameplay pass is claimed.

The original Setup.exe successfully emits its InstallShield 5.5 engine through guest execution. Replay that emitted engine, not a host-extracted cabinet. The following frozen CLI route uses an isolated build and leaves the legacy gameplay fixture untouched:

node test/run.js \
  --exe=test/binaries/candidates/icewind-dale-demo/icewind_dale_eng_demo/Setup.exe \
  '--vfs-include=**/*' --no-build --no-threads --quiet-api --quiet-blocks \
  --no-close --screen=800x600 --batch-size=100000 --tick-ms-per-batch=100 \
  --max-seconds=180 --control-stdin --frozen \
  --capture-launch=/private/tmp/iwd-original-bootstrap

Step 100 batches, then quit cleanly and wait for the capture to finish. launch.json identifies the guest-produced windows/temp/_istmp1.dir/_ins5576._mp (557,056 bytes), with its original zdatai51.dll and _wutl951.dll. The 96-file capture includes the original cabinet and CD2 source tree; those source files are copied, not decompressed by the host.

node test/run.js \
  --exe=/private/tmp/iwd-original-bootstrap/windows/temp/_istmp1.dir/_ins5576._mp \
  '--exe-guest-path=C:\WINDOWS\TEMP\_ISTMP1.DIR\_INS5576._MP' \
  --vfs-tree=/private/tmp/iwd-original-bootstrap '--cwd=C:\' \
  --dll-seed=/private/tmp/iwd-original-bootstrap/windows/temp/_istmp1.dir/zdatai51.dll,/private/tmp/iwd-original-bootstrap/windows/temp/_istmp1.dir/_wutl951.dll \
  --no-build --no-threads --quiet-api --quiet-blocks --no-close \
  --screen=800x600 --batch-size=100000 --tick-ms-per-batch=100 \
  --max-seconds=300 --control-stdin --frozen \
  --save-vfs=/private/tmp/iwd-original-installed-vfs

At batch 300 the Welcome screen is visible. Send {"cmd":"dlg-input-click:1"} and step 50 for the License screen; its Yes button is ID 6. Accept that button and step 50 for Recommended setup. Thereafter click ID 1 and step 50 at each of Setup Type, CPU Speed, Destination, and Program Folder. The unchanged defaults install to C:\Program Files\Black Isle\Icewind Dale Demo. At batch 750 the native install progress window is visible; at batch 1750 it shows 11%, with iddemo.exe (6,283,264 bytes), dialog.tlk (2,942,485 bytes), override resources, characters, and the first area archive actually written by the guest. At batch 3750 the progress is 18%; at batch 5750 it is 24%, with 284 files totaling 117,868,153 bytes under the destination (including the in-progress archive). These screens and file sizes were inspected on the isolated 6cada250 runtime. This is partial installation evidence, not yet a completed-install or fresh-installed gameplay pass.

The 300-second internal execution guard stopped normally at batch 7177 inside the guest cabinet DLL (EIP=0x007b4d3f), and the VFS export completed with process exit 0. This exit is a test deadline, not Setup Complete. The destination contains 284 files totaling 142,198,393 bytes. Read-only comparison against the legacy reference found 280 byte-identical files; Data/chranim.bif is still partial (52,695,040 of 69,527,564 bytes), and the original 316-byte icewind.ini differs from the legacy patched INI. The other two files are installer-created uninst.isu and readme.txt. The installed EXE SHA-256 is b94816d10029cb99c0315f175330c917be2fff298394e853e2764973d8d13af4.

Longer execution and interrupted-install replay

A fresh replay with --max-seconds=1800 reached inspected progress screens at batches 8750 (34%), 9750 (37%), 10750 (40%), 12750 (47%), 14750 (53%), 16750 (59%), 18750 (65%), and 20750 (79%). The internal deadline stopped at batch 21190, EIP=0x007b4edd, still inside cabinet decompression. It exited cleanly and exported /private/tmp/iwd-original-complete-vfs; despite that directory name, this is not a complete installation. Its destination has 302 files totaling 407,579,613 bytes, with 298 byte-identical to the legacy reference. The incomplete file is now Data/sndspell.bif (21,606,400 of 22,416,309 bytes). The other differences remain the unpatched INI and installer-created README/uninstall log. No payload bytes were supplied by a host decompressor.

Directly replaying the engine from that export exits with guest code 4 before its wizard: _ins0432.ini and _isenv31.ini have been consumed. Running the original Setup.exe again with the exported VFS succeeds and emits a fresh engine under _ISTMP2.DIR, preserving the installed files. This capture is /private/tmp/iwd-rebootstrap (426 files). Its engine reaches the same wizard, but selecting Program Folder produces unInstaller setup failed to initialize. You may not be able to uninstall this product. Dismiss the native message box with dlg-cmd:1; unlike the guest wizard buttons, dlg-input-click:1 alone did not dismiss it. This interrupted-install route is not a clean-install acceptance.

It also does not resume decompression: after 1000 additional copy steps the wizard shows 11%, and the previously complete Data/ar1000.bif has been replaced with an 8,632,320-byte partial file. The original partial sndspell.bif is untouched. This probe was quit explicitly; its separate /private/tmp/iwd-original-reinstall-vfs export must not replace the earlier output. A subsequent completion attempt needs one uninterrupted run with sufficient internal execution allowance, not repeated partial-VFS replay.

Next: require the actual completion screen and verify the complete output before changing the fetch recipe or registry. Then test gameplay using the original installed INI/KEY and CD2 layout. Do not reuse this partial directory as an installed fixture or assume the legacy KEY/CBF modifications remain necessary.

Legacy fixture preparation

The package is an outer ZIP containing an InstallShield cabinet and CD-resident data. The fetch recipe extracts the Recommended compressed group and merges the official CD2/Data tree into the local installed Data directory. Without that merge the real executable renders its “insert CD in D:\” screen forever.

The fetch recipe rewrites the extracted portable HD0:=.\ alias to the installed layout used by the emulator (HD0:=C:\, CD1/CD2:=D:\). This Infinity build's dot-prefix branch strips two characters from the original hd0: resource path, producing malformed C:\0:\... paths; the absolute installed aliases let it open Dialog.tlk and its BIF resources normally.

Local-session startup

The original IDirectPlayLobby2_CreateCompoundAddress shim returned S_OK and an output size of zero for every call. Infinity Engine uses the documented two-call sizing pattern in its local-session initializer: first it supplies a null address buffer and requires DPERR_BUFFERTOOSMALL (0x8877001E) plus the required byte count, then it allocates that buffer and calls again. Returning success from the size probe made the initializer abort, leaving Create Game to show “Cannot connect to the game session.”

The handler now sizes and packs each DPCOMPOUNDADDRESSELEMENT as a GUID, data length, and payload. test/test-directplay-lobby-address.js pins the null, undersized, and successful-buffer cases independently of the game.

The localhost dropdown must also mount every archive marked as installed (location=1) in CHITIN.KEY. The old menu-only 19-file manifest omitted BCSgen.bif and later installed archives; once DirectPlay succeeded, the installed-resource pass stopped at ChDimm.cpp:817, reported the CD as removed, and terminated. The dropdown now mounts all 34 location-1 archives present in the Recommended install.

CD2 and first-area gameplay

The official package's CD2/Data directory supplies 24 compressed area/creature BIFs, two movie BIFs, and IWDCD.2. Merely mounting those files under D: was not enough: manifest mounting registered each file and immediate parent but did not register the d:\ drive root. Consequently SetCurrentDirectory("D:\") and the game's FindFirstFile("D:\\*.*") disc scan failed even though D:\data\IWDCD.2 itself resolved. The VFS mount helper now registers the drive root and every ancestor, and Win32/DOS *.* enumeration also matches extensionless directory names such as cd2.

The CBFs are host-inflated once by local fixture preparation into normal BIFF files. Its matching CHITIN-full.KEY changes the 24 expanded BIFs and two direct movie BIFs from CD2 (location=9) to HD (location=1). The retail-derived KEY also advertises 2,464 resources whose backing BIFs are not in the official demo at all, including SNDVO.bif; leaving those entries active made the loader run its six-pass, 200 ms CD poll forever for an archive no disc in this package contains. The prepared KEY retains the 12,682 resources backed by the 60 installed/demo archives and removes only those dangling entries.

NPC and narration audio that retail stores in SNDVO.bif is intentionally loose in the demo. The dropdown mounts all 188 supplied Override/*.wav replacements. Non-audio override resources are not bulk-mounted at startup: some intentionally replace UI/game resources and activating the whole directory before the demo's normal phase switch trips ChUIControls.cpp:6668.

Character-generation sound sets

The executable's JigSawedME window title is the official demo build's internal name; it is unrelated to the Win16 JigSawed app elsewhere in the corpus.

After the Appearance dialog, the demo clears panel 45 and dynamically creates one control per directory returned by FindFirstFile(".\\sounds\\*"). With only Sounds/sndlist.txt mounted, the directory scan produced no sound set, the linked-list loop created zero controls, and the immediate lookup of control 0 returned null. The caller at 0x0069CDD1 then entered the assertion at 0x00534341 (ChUIControls.cpp:5945).

The Recommended install contains 16 official sound-set directories with 40 WAV files each (15 MiB total). The localhost manifest now mounts all 640 files, which creates those directories in the VFS and keeps Play usable after a set is selected. The exact acceptance completes Gender, Portrait, Race, Class, Alignment, Abilities, Skills, and Appearance, then requires the populated 16-set Sound panel after the formerly failing Appearance Done click.

The local acceptance skips the intro, requires the detailed menu frame, clicks Create Game, requires the large transition to Party Formation, opens a party slot, and drives character creation through the Sound panel. An error modal or a generic menu-frame change can no longer satisfy the test. The CLI capture must retain the menu labels and the Prologue title/buttons as real light GUI-font glyphs; detailed stone artwork without dynamic text is a failure, even if the same run later reaches gameplay.

test/test-icewind-dale-menu-web.js applies the same label-band gate to a fresh, no-cache Chrome page and also requires the browser VFS to contain the 2,942,485-byte Dialog.tlk plus Data/GUIfont.bif. It keeps sampling the actual 640x480 DirectDraw layer for another 30 seconds after the first complete menu, so a transient first-good frame cannot satisfy the browser regression. The repeated blank-label reports were not explained by a stale tab: a later fresh-page report disproved that earlier diagnosis. Repeated exact Chrome runs kept 738 label glyph pixels throughout the sustained sampling window; no runtime change is claimed without a reproducible failing transition.

It then names the character CODEX, accepts the party, waits through the real first-area resource load, and requires the native PROLOGUE chapter screen. The chapter narration body is currently blank, but REPLAY and DONE render. Pressing Escape hides this panel without completing it and exposes the tavern under Infinity's Paused for chapter text lock—the exact state previously misclassified as playable because the acceptance treated red status glyphs as success. The corrected route activates DONE, captures the unpaused tavern, clicks a distant floor point, and requires the created character's pixels to move. Acceptance then requires the native multiplayer-session state at mpsave/default/icewind.gam and verifies that its party bytes contain codex; the old NO DISC IN DRIVE D: screen and a loading image can no longer count as gameplay, and neither can a static paused HUD. The demo's Q path did not produce a complete numbered save slot under emulation, so the test does not mislabel that unavailable path as the persistence contract.

The app opts only authored state into browser persistence (Characters, Save, and MPSave). A fresh-Chrome regression writes the native default multiplayer session through the browser VFS, reloads the page, attaches a new VFS under the same icewind_dale_demo app id, and verifies one restored app-scoped entry whose bytes still contain codex. This covers the actual multiplayer session path written during character acceptance rather than treating temporary files as saved character state.

An exact browser run with cross-origin isolation and Threads enabled reached Party Formation with four secondary guest workers. The manifest-driven CLI acceptance reaches the first-area HUD and native default session save with the same dropdown asset list.

Call-form census (2026-09-29, box 1, uop-tier-design §15.1)

Window: batches 6000-6600, walking in the tavern. Measured with --handler-hist --edge-hist --uop-census and analysed with tools/call-form-weighted.js. As % of all block entries:

Nothing call-shaped is worth a mechanism here.

Redundant installer ZIP reclaimed (2026-11-01)

The coordinator removed the 645,596,503-byte retained ZIP after freshly verifying its SHA-256 992b3d9054ff089f40dba40debd0fe902117c1a21f80d7e34678ae3313c9ff80 and comparing all 56 file members byte-for-byte by SHA-256 against the unchanged retained 20261031T2058Z-icewind-install/original-source/ extraction. Installed files and original extraction remain available. Recovery uses scratch/runs/20261031T2045Z-icewind-recovery/recover-package.py and the standalone source ZIP, with the pinned hash above. Evidence: scratch/runs/20261101-coordinator-reviewed-compression/icewind-zip-reclaimed.json. Free disk rose to 2,674,139,136 bytes; no game execution or new gameplay claim.