Silver original Windows demo

Status on 2026-10-13: original main menu and dock scene rendered; player-controlled gameplay not verified.

Original fixture and mounting

The Infogrames 1999 Windows demo is archived at https://archive.org/details/SilverDemo. Publisher-era install.ini, demo readme and licence identify the original demo. No installer or licence prompt was executed or accepted in these investigations.

SilverDemo.zip is 129,637,371 bytes, SHA-256 645db1f10f9fd77fcc22f3140648fc154e3d5680a9f88cd7fe8d834268e887ea (SHA-1 a6af4d96fdc74b9649b1d96b513c416f8c65650b). Native extraction yields 3,123 ZIP files plus 24 files from the original InstallShield cabinet. All 3,147 mounted C-drive files preserve the original file bytes. silver.exe is 1,694,720 bytes, SHA-256 2301beb5100c6eb73eadcb700a2ce65e7b1221d2390207cbe3f07bc64df73ff8. Original mss32.dll and smackw32.dll are native DLL seeds.

The executable checks logical drives for GetDriveTypeA == DRIVE_CDROM, then searches for SILVER\LEVELS\VERDANTE\DAVHOUSE\DAVHOUSE.SCT (68,940 bytes). The existing harness's D drive has CD-ROM semantics. Keeping the extracted tree at C:/SILVER and mounting the unchanged original ZIP with --zip=SilverDemo.zip --zip-root=D:/SILVER/ supplies the 3,123 read-only media files. This passes the original CD search without game configuration changes, invented registry keys, API overrides or a fabricated ISO.

Requested flip-chain failure and fix

The original 8-bit startup path sets DDSD_CAPS|DDSD_BACKBUFFERCOUNT (0x21), caps 0x4218, and two backbuffers before CreateSurface at 0x41b8a3. GetAttachedSurface at 0x41b936 requests 0x4004 and succeeds; the next call at 0x41b980 requests 0x4000 on that first backbuffer and fails with DDERR_NOTFOUND. The actual error-dialog stack contains return 0x41b9af.

The old implementation allocated only one backbuffer regardless of the requested count. The generic fix allocates the complete chain with distinct storage, marks only the first attached surface BACKBUFFER, rotates all backing buffers on Flip, validates target overrides, reports the requested count, and rolls back partial allocation failures. Linear ownership preserves release behavior; a weak tail-to-front lookup avoids a reference cycle. Existing explicit and mip attachments remain covered.

This follows the documented DDSCAPS roles and Flip associations and target override. The change was reviewed/integrated as 3cc6e1c64; the original tested three-file commit is e14f5d97c3cd2854bc1401ce3f51f1ef5ddb47bb.

Canonical build from base 68e5e094a138c89f8c47898f2d68a64ab7a8318f plus the tested source produced module SHA-256 46c1e23aabca26de5e459f21463d3e1167f3a71d95e573ffedecefe7b31765f7. The DirectX source SHA-256 is 0bd81ce191e29a690ae560fadcde314b0c777692fc5d023b32f1be12ada603be.

Nine focused suites passed: complex-surface release, explicit attachment, mip chain, surface caps, Surface3 descriptor, surface enumeration, attachment enumeration, D3DIM worker parity and GPU asynchronous Flip. The expanded complex test covers double/triple/four-buffer storage and color rotation, target/caps failures, retained intermediate references, COM-slot exhaustion and DIB allocation rollback. On unchanged base source, the same test's narrow negative control fails at the second requested backbuffer with DDERR_NOTFOUND.

Reviewed original runtime and remaining route

With the exact original C/D fixture, the new module passes startup and renders the Silver main menu (Start, Load, Options, Quit Game). The first absolute mouse move/click at (320,362) and Enter did not establish gameplay. No approval or licence prompt was answered.

The last step request began at batch 232 and requested 120 batches. It never returned a control reply before manual shutdown, while final statistics show 352 batches. Logs contain repeated zero-byte reads at silver2.smk EOF 0x2c4a8, notably at batch 261. These observations do not establish a guest deadlock: step completion is settled on the next harness loop, and stdin commands serialize behind a pending step. The final PNG is byte-identical to the reviewed batch-222 menu capture.

Source inspection also found that the first absolute handleMouseMove has no previous DirectInput point, so it initializes that point without generating relative motion. The reviewed custom pointer remained around the menu's center; the Start click may therefore have missed the guest's cursor. A prepared follow-up uses the ordinary renderer motion contract to deliver directDx:0,directDy:112, captures the pointer before clicking, and limits subsequent steps to at most 20 batches. This remains an unexecuted hypothesis.

The prepared driver uses independent loopback HTTP controls so a pending step cannot block snapshot/capture/quit. A timeout is recorded with before/after batch and credit values, then captured and stopped; it never silently counts as a successful step. No production harness or guest state is patched. --trace-api is omitted because even a filtered trace sets _debugReadFile and caused the prior 345 MB repetitive log. Output remains drained with an 8 MiB retained-log limit and explicit omitted-byte counts.

Evidence is in scratch/runs/20261013T2254Z-silver-flip-original/investigation.json: 74 transferred artifacts were SHA-256 verified, including a compressed copy of the full raw log verified after decompression. Prior CD and startup evidence is in scratch/runs/20261013T2237Z-silver-original-cd-startup and scratch/runs/20261013T2232Z-silver-original-startup. Supervisor 177329, driver 177330, guest 177353 and their process groups were absent at 2026-10-13T22:56:19.682Z; the runtime slot was released. Follow-up execution requires a new coordinated runtime grant.

Relative-pointer replay, 23:11 UTC

The next granted run used the same source closure, module and original media. The relative (0,+112) motion visibly moved the custom pointer onto Start. After inspecting that capture, a normal left-button down/up advanced out of the menu. This confirms the previous absolute-pointer route had not tested a correctly positioned Start click.

Loading initially showed black frames and an occasional small Silver logo. Repeated snapshots often landed in the original loop at 0x43f96c, but later EIPs varied and the scene progressed. A read-only capture at batch 482 showed an 18,000-slot table with 8,253 used entries. The original instructions scan eight-byte entries for a zero first word. This evidence does not support calling that loop a deadlock. The small steps all replied through the independent HTTP channel; no pending-step recovery was needed.

By batch 1082 the game rendered a dock with a sailing ship, soldiers and other characters, plus a Grunt portrait. The original readme documents Escape as a cutscene skip. Escape during this scene opened an in-game menu containing Load, Options, Quit Game and Back. A second Escape, intended to return from that menu, instead reached the original promotional end card. The exact game branch responsible for that transition remains unverified. No character movement or combat was demonstrated, and this result does not establish whether the route was interactive gameplay or a scripted demo sequence.

The next ordinary route should inspect the original Start/demo and Escape semantics, avoid that second Escape, and attempt a visible world interaction under a new runtime grant. Do not modify the game or invent a save/config to skip the sequence.

Evidence: scratch/runs/20261013T2311Z-silver-relative-start/investigation.json. All 224 artifacts (13,379,717 bytes) were transferred and SHA-256 checked. The full output log is 6,258 bytes with zero omissions. Representative captures are 002-pointer.png, capture-47.png (dock), 005-cutscene-skip.png (in-game menu), and 006-return-world.png (end card; the filename records the attempted action, not its outcome). Harness quit returned zero; supervisor 193010, driver 193011, guest 193019 and all three process groups were absent at 2026-10-13T23:14:54.862Z.

World-controls replay, 23:35 UTC

The fresh run used the same pinned module/source and original fixture, with no Escape during startup or the dock sequence. Relative Start positioning again worked. The driver accepted 25 commands before harness quit (command 26); every actual step was at most 20 batches. Five ordinary left clicks at the inspected dock pointer (365,410) were interleaved with waiting. Captures show the dock crowd moving and portraits changing through Grunt, Fuge and the Duke. The final capture is still the dock at batch 3672. Neither those scripted movements nor portrait changes prove player control; the evidence does not isolate clicks as the cause of dialogue progression.

The original HTML readme, instdata/gamedata/common/readme/index.htm, lines 193–216, describes running down stairs toward Granddad, crossing the bridge left, entering David's house, opening the chest, collecting sword/shield, equipping the sword and returning for training. This identifies a useful interactive checkpoint, but does not document how this demo's dock sequence transitions there. The original dockside.sct contains @verdante and @davhouse; strings alone do not establish script control flow.

The separate original instdata/gamedata/instsupp/readme.txt, Gameplay / Skipping Cutscenes, documents Escape. A bounded, distinct next input probe is one Escape at the dock, immediately inspect the result, and, if the previously observed in-game menu appears, use relative pointer motion and a normal left click on the visible Back item. Do not send a second Escape. This is an untested route, not a promise that Back will skip the sequence. If a house/Granddad or other identifiable player scene appears, first capture without input, then click two distinct clear-ground destinations and verify hero displacement against fixed scenery before following the documented house/training objective. If the dock remains, record the unresolved transition; do not treat more animated frames as gameplay or invent a save.

Evidence: scratch/runs/20261013T2335Z-silver-world-controls/investigation.json. All 114 remote artifacts (13,365,642 bytes) were transferred and SHA-256 verified, including commands, replies, launch manifests and the full 8,203-byte log (zero omissions). Reviewed captures include advance-5.png, click-15.png and advance-25.png. Postrun hashing verified all 3,147 original C-drive files, the original ZIP, 130 source-closure files and the unchanged 46c1e23a… module. Harness quit returned zero at 23:45:04; supervisor 212438, driver 212439, guest 212464 and all three process groups were absent at 2026-10-13T23:45:09.545Z. This follow-up only packaged evidence and inspected source; another runtime requires a new coordinated grant.

One Escape and reviewed Back, 2026-10-14 00:00 UTC

Run 20261014T0000Z-silver-escape-back retained the same source/module and original fixture. At the dock, one ordinary Escape initially left the same image; 100 further batches exposed Load/Options/Quit Game/Back. Relative pointer motion placed the cursor on Back at (320,450); its reviewed left click returned to the dock. Another 200 batches still showed the dock. No second Escape or guest-state change was used, and no player control was established. All 3,147 post-run original hashes matched; the driver quit normally and all three PIDs/process groups were absent at 00:04:35.516. All 67 transferred artifacts (6,803,478 bytes) were individually hash-checked; evidence is scratch/runs/20261014T0000Z-silver-escape-back/investigation.json.

A source-only follow-up checked original installer metadata: the Silver Demo shortcut invokes Silver.exe without extra arguments, so a missing documented command-line switch is not supported by that file. Original sound/fugeapt1.wav and fugeapt2.wav are format17 stereo44.1kHz with fact-chunk durations19.608s and58.788s; the B variants are15.734s and42.478s. These durations do not establish which clip currently gates the script. Before another replay, correlate actual loaded speech, mixer completion and script progress using existing bounded tracing; do not infer a broken audio API merely from a long animated sequence.

Automated original-media recovery, 2026-10-14

Run node tools/fetch-candidate-corpus.js --id=silver-demo. Fresh acquisition verified the previously recorded ZIP SHA1/SHA256, preserves source/SilverDemo.zip unchanged, extracts its3123 files under game/SILVER, then extracts the24 original data1.cab files into the same directory. Both sets have no overlapping filenames and no content changes. Canonical silver.exe matches the prior recorded SHA256. Complete3147-file hashes are in test/candidate-corpus/silver-demo-files.json; node test/test-silver-demo-recovery.js verifies pristine output. Source media and cabinet outputs were independently extracted and byte-compared.

This adds a recovery candidate only. Original launch still requires C:/SILVER and unchanged source/SilverDemo.zip mounted at D:/SILVER/ with CD semantics, as documented above. Generic candidate CLI does not yet encode that dual mount; do not use its default single-directory execution as a compatibility verdict. No browser/desktop registration, installer/licence prompt, new player-control validation, audio/FPS claim or public deployment.

Offline browser-evidence recovery, 2026-10-14

Run scratch/runs/20261014-silver-history-recovery recovers retained Boat command outputs without restarting the old sandbox or executing the game. boat --json info bx_yqnspudj reports stopped, with a verified snapshot at 02:54:21.026Z. boat --json exec bx_yqnspudj -- ls -ld /home/user/scratch/runs returns HTTP409 machine_not_running. This is a stopped-machine observation, not proof that the disk artifacts were deleted. No SSH, resume, replacement job, browser, build or guest execution was used.

boat history bx_yqnspudj COMMAND_ID remains readable. Its response is JSON with a data string; decode that string before interpreting base64. Fifteen retained archive-chunk outputs reconstruct the complete 02:22 archive: 5,622,000 bytes, SHA256 7fc9ad29807389c214c8af0efd60648012fa1f91b3c643aef210d6215e8cba23. The stored packaging output cmd_r2ytupol81dk independently records that hash. All 96 extracted files match the original artifact manifest. The archive and extracted files are inside the recovery run (about 14 MiB total). The committed docs/silver-history-recovery-20261014.json retains chunk command IDs/offsets and the checks needed to repeat this recovery. Read only those stored commands; do not execute the historical command strings.

The 02:22 archive contains both runs, drivers, input commands, logs, module responses, and pre/postflight manifests. Its actual module response is 1,796,313 bytes, SHA256 02f7ed397f92fb8e4ed81798b0e2ba3405b27115b9887ce5a52ee4c82c2936d6; recorded source HEAD is 56dc827cc64241cf763fff0d821c6a3fbd195a25. Do not replace this with today's HEAD: preflight includes the actual source closure hashes. Browser is Chrome/151.0.7922.108, headless, 1000x800, --enable-unsafe-swiftshader; the reviewed UI shows Software GPU and Threads enabled. The route creates a temporary silver_probe app, mounts all original files at C:/silver and the unchanged ZIP at D:/SILVER, and uses native Mss32.dll/Smackw32.dll. It is not a registered lib/apps.js Silver route.

The newly reviewed run-two/pointer-two.png shows the menu with the pointer on Start. The 02:22 driver lacks relative:true and holds the button only 80ms, so it is not the later corrected input replay. Historical command cmd_bpoch8nvo8sg records adding relative:true and replacing that hold with 500ms, a passive snapshot, then another 500ms. Packaging commands preserve the later ordinary-input sequence and summaries, not the complete later runs.

Stored output cmd_m2cwqpgklgjx recovers a JPEG derived from the 02:31 run-three/final.png. Newly reviewed: the original outdoor house/stairs/ bridge scene is visible. This is a lossy derivative, not the original PNG, and a single image does not prove player control. Its local filename is cmd_m2cwqpgklgjx.jpg. The 02:31 complete archive remains unavailable locally (recorded SHA256 405e36702214669c7d580a91aecbd28da362b96cdac8b78257e108e949b4d050). The final 02:36 archive remains unavailable locally (8,911,631 bytes, SHA256 93fc85c1f5b6fc5de897fc96499c337b239bc8c090a2b7bbf387a8233ac240ec). Nine requested image histories return empty data, done:true, with offset == size despite nonzero original output sizes; exact responses are retained in history-receipts.json. Do not describe these missing images as visually reviewed. Historical summaries report native completion status2 and ring segments0/1; those later raw timing files were not recovered here.

The retained original ZIP was freshly rehashed against the pin above. /tmp/silver-source matches 3,123 inventory files, with zero mismatches; all 24 cabinet payload files, including silver.exe and native DLLs, are absent from that particular tree. Do not launch it as a complete fixture. The existing corpus extraction recipe can restore the cabinet files from the retained media when disk/runtime coordination permits; no download is needed.

Next bounded ordinary-input comparison (not executed)

First restore/verify all 3,147 fixture hashes and both original mounts. Use the recovered browser driver as a route reference, explicitly correcting relative:true and the one-second Start hold as recorded above. Pin the actually served module/source, browser, GPU backend and worker settings; do not assert that a new build matches the 02:22 module. Obtain the sole runtime slot before launching. Bound each fresh browser session to 300s with retained logs and normal owned-process cleanup.

Run a no-world-input control after the reviewed Start click. Capture the house scene against fixed scenery and observe whether movement/dialogue continues automatically. In a separate matched session, reach the same observable scene/readiness checkpoint, capture before input, then use normal relative pointer motion and a one-second left hold on one clear destination; release and capture at fixed intervals. Repeat at a distinct visible destination only if the first comparison is interpretable. Record host down/up and actual guest input consumption if existing diagnostics expose it. Stop at the deadline or unresolved dialogue/readiness; a changed portrait, continued script, host event, or animated scene is not movement acceptance. The original readmes describe house/chest/sword/training objectives but do not document the ground-click/selection contract. Do not assume that two clicks failing proves an input defect, or alter guest state to force control.

Acceptance remains unfulfilled: establish repeatable input-caused hero displacement relative to fixed scenery, distinguished from the no-input control, and review the actual before/after images with source/input/backend identity. No new gameplay, audio, FPS, or engine-fix claim results from this offline recovery.

Complete local fixture restored, 2026-10-14

Follow-up scratch/runs/20261014-silver-cabinet-restore restored exactly the 24 missing cabinet files (39,027,302 bytes) directly into /tmp/silver-source using the existing tools/is5-extract.js and original data1.cab. It did not re-extract the ZIP, download media, execute the installer, or launch a guest. Before extraction, all 3,123 existing files matched the pinned inventory; the cabinet catalog matched the 24 absent paths and sizes exactly. Extraction was refused for any existing destination, mismatch, or projected free space below 2 GiB plus 8 MiB metadata margin. After extraction all 3,147 original SHA256 values match test/candidate-corpus/silver-demo-files.json, including the previously pinned silver.exe. The original 3,123 files remain unchanged.

The source-preparation script, preflight, extraction log and validation are retained in that run. docs/silver-cabinet-restore-20261014.json preserves the command, extractor/inventory/media hashes, resulting disk space, and exact mounts. The earlier missing-cabinet finding above describes the pre-restoration state and is now resolved for this tree only.

For the next separately coordinated browser run, serve /tmp/silver-source as the complete C:/silver tree and launch C:/silver/silver.exe with original Mss32.dll and Smackw32.dll. Serve the existing unchanged scratch/runs/20261014-silver-recovery/SilverDemo.zip as /SilverDemo.zip, then use HttpRangeProvider.open('/SilverDemo.zip') and ZipMount.mountZip(wine._helpCtx.vfs, provider, {zipPath:'SilverDemo.zip', root:'D:/SILVER/'}) in the recovered route's beforeRun hook. Preserve D-drive CD-ROM semantics. This is fixture readiness only: no new browser, build, game, input comparison, or player-control result has been produced.

Bounded idle/input attempt, 2026-10-15

Evidence: scratch/runs/20261015-silver-idle-input. Both authorized300s sessions used source9449e6c0d, identical fingerprint dc29bc7dc19724ea78671352de90649b6e6dbc1adaf9339c5850de733d16d50a, Chrome151.0.7922.108 headless1000x800, Software backend and isolated guest main Worker. Both actual browser module responses matched70c6d329d51eee5db553715f53ffc47f3e6e2491b54ae5327f1b8b012427d655. Pre/postflight941source and3147original fixture hashes passed. No guest file, clock, save or runtime state was patched.

Idle131122 ran00:15:17.793–00:20:28.992Z; input134433 ran00:21:11.543–00:26:14.916Z. Both controllers closed at their300s deadlines and supervisors exited0 with no survivors, postflightPassed:true and errors:[]. The deadline includes launch, original media mounting, image review and cinematic progression; supervisor completion includes cleanup. Each run executed only the same reviewed relative(0,+112) Start positioning and one-second normal left hold. Held snapshots record hostMouseButtons:1 and later release0. This proves host delivery, not actual guest input consumption by itself. Actual menu departure and natural dock progression were visually reviewed.

Idle naturally reached the house by00:19:55.351Z. The world marker was processed00:20:07.162Z; idle0/2/5/10-second captures were retained, but no20s capture fits the deadline. They show David beside Granddad while the portrait disappears, without world input. Control readiness remains unproven. A brief read-only attachment to the same browser during startup used defaultViewport:null, preserved1000x800, saved passive-observation.json and disconnected; it showed the apparent DLL-loading pause had progressed.

The second run reached the house by00:25:28.062Z; house-ready-check at 00:25:54.468Z shows the portrait gone. Its camera/character placement has a small offset from the idle image, so do not claim exact scene alignment. Crucially, controller shutdown was00:26:12.877Z. The attempted mark-world request8 and relative ground request9 were written only at00:26:21.118Z and 00:26:21.590Z, after shutdown. They are retained with timestamps in input/late-unexecuted-requests.json; executed commands.json ends at7. Neither world request executed, no ground hold was sent, and no movement response was tested. A missing after-command-9 capture is therefore expected. This is an incomplete comparison due to the observation/input window, not evidence of rejected game input or a game bug.

The runtime slot was released after terminal cleanup confirmation; no restart or extension was attempted. The two complete artifact inventories comprise 61idle files/8771557bytes and48input files/5942831bytes, independently byte/SHA256-verified on transfer. Representative reviewed images are idle/pointer-settled.png, idle/idle-10s.png, input/house-ready-check.png and input/final.png. No gameplay qualification results from this attempt.

Next execution needs a separately coordinated budget with enough time after natural house arrival for image review, a settled pointer capture, a normal hold and matched observations. Base that budget on the measured startup/ cinematic times above; preserve an explicit hard wall-clock limit. Do not silently convert the two late request files into completed input evidence, force readiness or call the stationary pictures proof that control works.

2026-10-15: paired ordinary movement verified

The revised 600-second pair establishes narrowly scoped player-controlled movement in the original demo. Evidence is self-contained under scratch/runs/20261015-silver-600-comparison/{idle-600,input-600}. Remote controller source was 77e0380e3fdca6a44f2b31088760285b91192ddb, with the unchanged original C:/silver fixture and original ZIP D:/SILVER mount. Both loaded WASM SHA256 70c6d329d51eee5db553715f53ffc47f3e6e2491b54ae5327f1b8b012427d655. The identical source/module/fixture fingerprint was c4a60b4f0954f01a9eb9f56ec206d34a736b6fc78001c4cbcfcccbe67133cbf7. Chrome151.0.7922.108 used Software rendering, isolated guest main Worker, and the unchanged1000x800 viewport. All941 source and3147 original fixture files passed pre/post integrity; no guest state or clocks were modified.

Idle148910 started00:39:27.392Z and exited cleanly00:45:03.351Z. After normal Start and natural cinematic progression, reviewed portrait-free house captures idle-0s through idle-20s show David stationary beside Granddad; Granddad's animation alone does not establish control. Input152857 started 00:45:14.550Z and exited early00:54:12.009Z, within its600s deadline. Its portrait-free forecourt matched the idle scene with a small camera offset, so comparison uses the doorway, bridge and character relationships rather than claiming pixel-identical scene alignment.

Input commands.json records every command0..14 executed. Command10 held the ordinary left button1000ms at guest(410,330), after relative(+90,-32) and inspection of ground-pointer-settled.png. David moved right toward Granddad, visible in ground-1-20s.png. Command13 held1000ms at(260,355), after relative(-150,+25) and inspection of bridge-pointer-settled.png. ground-2-20s.png shows David on the left stone bridge, reversing the first displacement while the fixed house scenery remains unchanged. Both holds have held-button receipts and complete0/2/5/10/20-second observations. This directional response to two ordinary destinations, against the stationary20-second idle baseline, establishes player movement for this route. It does not establish combat, saving, completion, audio quality, performance or broad playability. The controller's conservative automatic playerControlVerified:false field remains unchanged; visual review supplies the scoped movement assessment in the run result.

Both supervisors returned code0, postflightPassed:true and survivors:[]; both errors.json files are empty. Runtime was released immediately after cleanup. The earlier300s failed-to-execute comparison remains preserved unchanged. Work pauses after this evidence checkpoint at the user's request.