Darkstone demo 1.0

Software manual combat, October 30

Run scratch/runs/20261030T2206Z-darkstone-software-combat/ establishes original skeleton combat on the software backend. It reuses the retained committed-main 10952264c runtime compiled at 21:54:53.532 UTC, module 4649ee5fcf5a10a0212e690d826940bf0ba6445bea678bcac3978c0464bd67c6 (1,891,324 bytes), independently hashed from the actual HTTP response at 22:03:35.209 UTC. All 15 original installed files match before/after hashes. Chrome 151.0.7922.108 uses isolated guest Workers and software D3DIM. No production code, guest memory, exports, saves or input queues were changed.

The original keyboard configuration/readme map F to automatic/manual combat. Native F in Town visibly displays Combat Mode: Manual; native I shows LIFE 35/35, MANA 25/25. The established Town exit and ground clicks (.65,.35), (.78,.60) reach the skeleton clearing. Read-only state before attack at 22:06:52.082 UTC is LIFE 32/35, EXP 0: the enemy has already damaged CODEX while manual mode remains selected. Native left holds at (.64,.48) for 4.5 seconds and (.66,.48) for 20 seconds leave the enemy alive. After correcting the target to its torso, native left at (.625,.46) from 22:07:42.698 UTC for 18 seconds defeats it. Personally inspected third-settled.png shows its absence; subsequent victory-inventory.png and survival-final.png confirm survival with LIFE 17/35, MANA 25/25, level 1 and an advanced experience bar. The successful hold is not represented as one instantaneous hit.

numeric-state.json records EXP 0 → 79 and health through the encounter. The samples only read shared linear memory and locate the live CODEX record; the original executable's inventory disassembly identifies name at entity +0x30 (0x4b0756), level at +0xca and experience at +0xcc (0x4b0aba, feeding the experience bar). In this run the name is at linear 0x6ec03c, so experience is at 0x6ec0d8. These are run-specific observed addresses, not a production ABI. The visible inventory independently checks health and the experience-bar increase; it does not print numeric experience. Unused copies of the character record remain unchanged and are retained in the raw samples. Enemy numeric HP was not decoded.

Owning main legacy endpoint 1 has advancing native software completed draws, zero observed fallback software completions and no guest trap in captured slices. Exact counters, native inputs, times, hashes and personally inspected screenshots are retained in the run. A read-only inspection connection briefly resized the page to 800×600; it was restored to 1280×900 before the attack. The game canvas remains 640×480. Child renderer ownership, audio, FPS, save/reload and broad rendering parity remain untested. No production blocker reproduced and no repair was invented. Quit was acknowledged and PAGE CLOSED logged, but the outer browser harness returned 143, cause unknown; local process/port cleanup is recorded separately. No Boat or assigned box was used.

Original skeleton combat, October 30

Run scratch/runs/20261030T2158Z-darkstone-combat/ personally inspected combat on committed main 10952264c12cb1d4e2f2a096f59e8b367a1fb82b, after alpha repair 328b6631a. Canonical compilation of committed sources produced module 4649ee5fcf5a10a0212e690d826940bf0ba6445bea678bcac3978c0464bd67c6 (1,891,324 bytes); the Chrome HTTP response matches. The frozen archive excludes other workers' uncommitted changes. All 15 installed original assets match the served manifest and remain unchanged after the run. No guest memory, exports, save files or synthetic input queues were modified.

Chrome 151, isolated guest Workers, WebGL, 1280×900 page and 640×480 native canvas followed the established CODEX warrior creation and Town exit. In Ardyl, ordinary ground clicks (.65,.35) then (.78,.60) reach the clearing with a skeleton beside CODEX. clearing.png shows its raised sword; combat-before-live.png already shows automatic melee engagement. A native left-button hold at (.64,.48) for 4,500 ms starts at 21:58:12.086 UTC, logged as trusted DOM input at page (611,510). attack-held-after.png shows white death particles and the skeleton gone; attack-settled.png and combat-completed.png confirm its persistent absence. Native I opens the inventory: inventory-health.png explicitly reads LIFE 32/35, MANA 25/25, with a nonzero experience bar. Initial Town health looks full, but no initial numeric health or enemy HP was collected. Automatic combat preceded the explicit attack, so this does not isolate the click as the sole cause of damage or death. It establishes original combat during ordinary native input, beyond locomotion. No combat blocker reproduced and no runtime repair was necessary.

Owning legacy endpoint 1 has advancing WebGL submissions, finishing at 103,283 draws / 1,287,543 consumed calls, zero observed errors/fallbacks and zero native/fallback software completions; declinedClears=1 remains explicit. Software combat, child render ownership, audio, FPS and save/reload were not tested. Refused perf-telemetry connections remain in the console log. Quit was acknowledged and PAGE CLOSED logged; the outer browser and server returned 143, cause unknown. Postflight verifies debugger port 35309 and private server port 18175 closed, with no own runtime remaining. No Boat was created or touched. Exact inputs, times, frame hashes, source/module/assets, recovery recipe and limitations are retained in the run. A useful next step is manual-mode combat with an initial numeric health baseline, followed by software-backend combat; this run makes no manual-only attribution claim.

Package and installation

The local package comes from the Windows 98 A-D archive recorded in sources.md. Its original Setup.exe is an InstallShield bootstrap, not the game executable. tools/install-darkstone-demo.js runs that bootstrap inside Wine-Assembly, captures the _INS5576._MP child it emits, and then drives the guest installer through Welcome, destination selection, program-folder selection, file copy, and Setup Complete.

The installer must see the package recursively: dvoices1.mtf is under the source data/ directory. Mounting only top-level files correctly produces InstallShield move-data error -113; the installer tool uses **/* and checks the complete payload instead. The host copies files only after the guest installer has written them.

Pinned SHA-256 values:

The resulting installed/ tree contains 14 files and 33,471,003 bytes before manifest generation. The app and generated manifest now use only that tree. Run the authentic installer with:

node tools/install-darkstone-demo.js

Distribution status

The installed readme calls the package DARKSTONE Demo 1.0, describes its reduced music and voice assets, and lists Windows 95 or 98 as supported. It does not contain an explicit grant to redistribute the package. The original installer and its generated playable tree therefore remain local and gitignored. sources.md links the preserved original distribution so it can be obtained independently; Wine-Assembly does not publish these game files.

Gameplay route

test/test-darkstone-gameplay.js launches the registered app through one headless CLI process with cooperative scheduling, frozen stdio control, and the CLI's own --max-seconds guard. It skips the logo, selects New Game and One Player, creates a warrior named CODEX, and moves the new champion into the first party slot. The slot assertion is scoped to the portrait region so a card merely following the cursor cannot pass.

Demo 1.0 enters Town directly after party confirmation; unlike the retail manual's route, it does not display a difficulty menu. The test waits through the forest loading animation. Its live-town check requires the blue mana HUD that is absent from the loader, then holds the right-arrow key and requires a substantial pixel change from camera rotation. This is a functional acceptance test, not a performance benchmark. Screenshots default to the system temporary directory.

Run explicitly with:

node test/test-darkstone-gameplay.js

Retained gameplay recovery, October 7

Before assigning a new-game run, a read-only audit recovered all ten existing captures from /tmp/wine-assembly-darkstone-gameplay/, with file mtimes on October 6 from 06:31:23 to 06:31:46 UTC. They are copied and SHA-256 indexed in scratch/runs/20261007-darkstone-retained-gameplay-audit/. The reviewer codex:darkstone-refill-worker personally inspected 08-town-before-input.png and 09-town-after-input.png: both show Town, the champion and full HUD; the road, trees and gate rotate substantially between them, consistent with the ordinary Right camera-input route documented above. This is historical controlled gameplay, so Darkstone is excluded from a never-qualified new-game refill. No game was rerun during recovery.

The original invocation, execution start/end, source/module hashes and exact renderer were not recovered. The result records these as unknown, retains file mtimes separately and timestamps the recovery itself. It does not qualify the current module or establish gameplay FPS. The later October 6 browser results only reached the main menu and do not replace this evidence.

--stuck-after kills the asset load, and it looks like an emulator bug

Fresh browser menu checkpoint, October 15

Run 20261015-darkstone-software-menu captured and reviewed the ordinary main menu on an explicitly older admitted source snapshot: controller commit eb5a9559fbb4209823cd0473e5dc5bb6c4d7b3bb, production-equivalent to the verified IW 6c22121e build, canonical module 04ede400070b3acbdcc2ad5f3c033bcf051e3d2da9a4d28a449324d23ab0ba17. This does not test the later GetParent/main changes.

The owning page's actual wine._wasmModule object hashes to that module (1,803,941 bytes), independently matching the served HTTP and disk bytes. All 15 installed fixture paths matched locally and over HTTP before launch and remained unchanged afterward. Chrome 151.0.7922.108 ran headless with SwiftShader enabled, isolated guest Workers and the software backend request.

At 16:35:22.622 and 16:35:25.031 UTC the completed main-Worker slice sequence advanced 4691 → 4804 and replay commands advanced 774343 → 821142, with zero reported fallbacks and zero GL triangles. These are replay/presentation counters, not software raster-draw consumption evidence. Backend remains unknown. A child Worker also exists, and the collector does not attest its render ownership. The console retains missing ole32.dll/comctl32.dll built-in fallback messages; no crash was reported.

The reviewed screenshot shows New Game, Load a Game, Configuration, Credits and Exit Game. Console resize is 640×480: do not blindly reuse the older 800×600 route coordinates. Before another click, capture the live canvas dimensions and map the visible menu target. No menu click, party creation, Town arrival, control, combat, audio or FPS was tested in this checkpoint.

The outer controller 247615 finished normally at 16:35:25.152 UTC. Independent postflight at 16:35:49.700 verified no active runtime processes, unchanged source/dependency/browser/fixture inventories and preserved 937/a948/04ede/0dbf/e26a modules. Earlier staging-only attempts stopped before the browser on an absent sparse-checkout app document and an inaccessible unrelated Chrome extensions directory. No permissions were changed; the successful preflight records sparse omissions and pins regular Chrome files.

Recovery bundle in the shared checkout: scratch/runs/20261015-darkstone-software-menu/evidence.tar.gz, 1,532,791 bytes, SHA-256 984d1aa01ce7720bf8ee3f54e39618e0c52c97f039df9f11c86e584ea697d3f4. The screenshot is darkstone-menu-evidence-3/capture/menu.png inside it, 478,739 bytes, SHA-256 7bc3eb8915e4f09dbf609f717e41fa075b51eb262c50700e7fc7355d12fbf8b5. The archive contains the exact executed outer controller, invocation, pre/post inventories, fixture hashes, console and slice logs. Extract into the same run directory to recover the relative paths. Darkstone has no candidate entry in the current candidate-corpus manifest, so this remains investigation evidence rather than assigning another application's dashboard candidate ID. See docs/darkstone-backend-checkpoint-20261015.md for the staged recipe.

Driving this app by hand with a plain test/run.js command line ends at batch 1183 with no error line, no crash dump and no unimplemented API — the process simply stops. That is not the guest. It is run.js's own --stuck-after, which defaults to 10 (test/run.js:773): after ten batches that retire no new EIP the harness decides the run is wedged and ends it.

The loop it trips on is legitimate. 0x004e7b2b is the demo's asset-load loop reading ddata.mtf, and a load that reads 33 MB out of one archive spends a long time in one tight block by design. Nothing is wrong, so there is no diagnostic to find; the run just disappears.

The fix is one flag:

--stuck-after=1000000

With it the demo runs straight through the logo, New Game, One Player, warrior creation and the forest loading animation into Town, with the full HUD — portrait, health and mana orbs, potion belt and minimap — on a champion standing on the forest road. Verified on the remote bench box under --d3d9-renderer=software --d3d9-programmable.

Reach for this reading whenever a bounded run of any asset-heavy app ends early and silently: the absence of an error is the tell, because every real failure path in the emulator prints something.

The host's per-object GetTickCount loop and the clock-spin detector (2026-10-10)

After "Create a session", the host's main thread stamps every object with GetTickCount and makes no other API call in between, with only 5-8 blocks between reads (w4). The clock-spin detector took that for a busy-wait and parked once per object, leaving the host on the white Session menu for minutes. a7e5f4f75 makes a read count only when the callee-saved registers match the previous read in that context, so a per-object loop no longer qualifies. Host-only A/B, same build (run 20261010T1720Z-clock-spin-regs-check-boat): --spin-regs-check=15 gives 0 parks and Town as HOST; =0 gives 1213 parks and the frozen Session menu. The two-seat test no longer needs --no-spin-park for this.

Multiplayer over the virtual LAN (2026-10-10, w4)

test/test-darkstone-vlan-gameplay.js (heavy, boat) plays two seats to the Town together. Menu route (800x600 coordinates): New Game (400,260) -> Multiplayer (400,374) -> create a champion as in single player -> OK (425,570) -> Available Services: TCP/IP (400,374) -> Session: Create a session (400,324) / Join a session (400,424). Join asks "Enter IP address, or Name of Host": click the field (400,372), type with keydown + di-keydown + keypress per character ('.' is VK 190; the champion-name field works without keypress, this one does not), OK (400,420), then double-click the session in the list (400,330); a single click only highlights it.

DirectPlay: CoCreateInstance of an IDirectPlayLobby2 (only for GetConnectionSettings, i.e. "not lobbied") and of IDirectPlay3, used as IDirectPlay4 when available (flag [0x52aa64]). The game's wrapper object lives at 0x8a9f00: +0x6704 is-host, +0x6708 "session lost, stop sending", +0x670c own DPID, +0x6710/+0x6714 the player event and a quit event, +0x6718 the receive thread, +0x6720 joined/in-world, +0x6724 CreatePlayer succeeded, +0x673c/+0x6744 received message count/bytes. Thread 0x40bb70 waits on the two events and runs the receive loop 0x40c010 (Receive(DPRECEIVE_ALL); sender 0 -> system handler 0x40cee0, else app handler 0x40d390, a jump table on the first byte, types 0x65-0x7f at 0x40d878). Every send goes through 0x40cb60 (size per type from the table at 0x40ce68), which calls GetMessageQueue(SEND) and then SendEx with GUARANTEED|ASYNC|NOSENDCOMPLETEMSG (0x601); it treats only DPERR_CONNECTIONLOST (0x88770168, logs "SESSIONLOST", sets +0x6708) and DPERR_BUSY (retry x5) specially. Before 47aafe16e our SendEx refused ASYNC and nothing ever left either seat.

Join protocol: the guest, inside its EnumPlayers callback, resets the event at [0x647ee0], sends type 0x6c (8 bytes) and waits WaitForSingleObject(10 s) up to three times at 0x416e4b (timing out tears the session down via 0x40bf10). The host's 0x6c handler 0x40d980 either pings 0x7f every second (host not yet in the world, +0x6720 clear) or, normally, announces a block transfer with 0x6d (16 B: block count, size; guest handler 0x411850) and sends the world in 0x6f blocks (536 B, guest handler 0x411940, "receiving %d/%d..."), each acked by the guest with 0x6e (12 B). After the last block the guest sends 0x7e and 0x40e1a0 sets +0x6720 and signals [0x647ee0]. The demo's town took 5 blocks. Headless this is slow but correct: run.js pumps a seat that is not blocked on the wire only every 64 batches, one block round trip per pump, so the test waits 800 batches after joining.

2026-10-15 fresh paired backend evidence

October 17 retained-image review: the later software ground-click run exited 0 and shows the first WebGL destination at a closely matching gate/tree/feet pose. WebGL has a substantially larger lower-left black triangle; software still has a smaller wedge and foreground obstruction. Backend disagreement is established, but neither correctness nor the responsible draw/state is. See the retained same-pose comparison in darkstone-backend-checkpoint-20261015.md for PNG hashes, consumption counters, cleanup receipts and the proposed single-frame draw-input diagnostic. No renderer change or new run accompanied this review.

The original installed demo reached Town through ordinary character creation on both software and WebGL using a fresh normal build at source fadaff08a (module SHA256 9b57583c1ca8887f7195a268c94312ca6352eb646f4071c740c5a75aff1eb69e). Actual before/idle/held-ArrowRight PNGs establish camera control in both arms. Owning main endpoint 1 has advancing completed native software draw calls in software and advancing actual WebGL submissions in WebGL. Observed WebGL error/fallback counters are zero, but declinedClears=1 remains explicit. Child renderer ownership, locomotion/combat/audio/FPS and original-27 membership are not qualified. See docs/darkstone-backend-checkpoint-20261015.md for exact landmark counters, independent cleanup and the full recovered archive SHA256 25252b50372c11589f30439d0a13ab39b09576b5844c1855721c26a429c2aa8f. Earlier c7e/04ede screenshot-only evidence remains backend-unknown.

2026-10-15 locomotion follow-up: the no-rebuild WebGL capture on paired module 9b57583c showed two ordinary left-click road targets changing character position relative to Town gate/tree landmarks, with a following camera and stable intervening idle. No camera key was used. Main endpoint draws advanced with zero observed errors/fallback sites/direct or fallback software completions; declinedClears remained 1. The session later hit its input-queue deadline (exit 1), so the additional final-idle screenshot is absent; both independent cleanup/preservation checks passed. Full qualified result and archive SHA56acbc258d810f764e91edcc0644e45816d4ebd668f53b8f7174e1c6f2476087 are in docs/darkstone-backend-checkpoint-20261015.md. This is movement evidence, not an overall successful terminal run, FPS/audio/combat qualification, or original-27 membership.

Fresh registered Chrome Town/camera run, October 30

Run scratch/runs/20261030T2058Z-darkstone-browser/ uses the retained installed EXE b43db5e1… and the complete registered darkstone_demo manifest, without an installer or any change to the missing-Setup test guard. Pinned browser source is main ba96fc3ad84edee6c899762c596e8b1bf2bb2195; all tracked pinned sources match that commit (25 unrelated untracked tools are separately listed). An in-memory private localhost server holds source/module/assets fixed during both arms. HTTP response bodies identify module 6085eff52c9be987f67d83563602a0ea89e0c4bac9c36f9cf11171bd3adada4e (1,883,551 bytes); WebGL's built-in compile identity independently matches it. No WAT fallback occurred. The exact runtime archive, source inventory and all 15 fixture HTTP/local hashes are retained. Chrome 151.0.7922.108 runs headless, 1280×900, isolated guest Workers, native game canvas 640×480.

Personally inspected menu, creation, party, Town-before, idle and held-Right captures on software and WebGL show the original CODEX warrior, Town label, blue mana HUD and rotating gate/tree/road landmarks. The five-second idle control changes only 160 software / 110 WebGL world pixels above channel delta 20; camera input changes 446,306 / 419,279 of 515,372 world pixels. This is camera response, not locomotion or combat. The initial foreground palisade obscures much of the view; no general rendering-correctness claim follows.

At Town before/idle/after, owning legacy endpoint 1 reports software completed native draws 726155→831005→859360 (slice seq 31868→37906→40499), zero fallbacks. WebGL actual submissions (draws) advance 30638→41656→45232 and consumed calls (drawCalls) 395419→488046→517185 (seq 25448→33252→35657), zero errors, zero fallbacks, zero direct/fallback software completions. declinedClears=1 remains explicit. Actual WebGL renderer reports ANGLE Intel/Mesa UHD620, despite the probe requesting SwiftShader flags; this does not qualify physical hardware. Child render ownership is not attested by the collector.

Use inspected geometry, not old 800×600 coordinates: normalized visible-canvas New Game (.5,.455), One Player (.5,.54), Create (.18,.68), name (.55,.82), Create/OK (.53,.95), party slot (.66,.375). One Player needed a second ordinary click in each arm, after the first screenshot still showed the mode menu. Exact input commands and landmark timestamps are retained. Both explicit quit commands were acknowledged and logged PAGE CLOSED; both outer processes returned 143 during teardown, cause unknown. Do not describe that as a clean terminal exit. Postflight verifies no owned Chrome or private server remains. No Boat was created or touched, no public deployment, no Telegram. The bounded Town/camera/backend route passes; audio, FPS, save/reload, child ownership, combat, locomotion and full-game correctness remain untested. Recovery recipe and portable investigation result.json are in the run directory.

Native locomotion and Town exit, October 30

Run scratch/runs/20261030T2110Z-darkstone-locomotion/ replays the archived ba96fc3ad… runtime directly from pinned-runtime.tar.gz in a private localhost server's memory. All 1,292 served files match the preceding run; both Chrome HTTP module hashes are 6085eff5…, with the original installed EXE b43db5e1… and all 15 fixture paths unchanged. This is a retained-build run, not a test of later main. No installer, guard, runtime source, guest memory or game export was modified. Chrome 151, isolated Workers, 640×480 game canvas and 1280×900 page reproduce the CODEX warrior/party/Town route.

The installed readme's section IV assigns movement to the left mouse button; left/right arrows rotate the camera. No camera arrows were used here. After personally reviewing Town before/idle, an ordinary unmodified left click on the visible road at normalized (.45,.36) in WebGL produced a turning/running pose, the Land of Ardyl loader, then the Land of Ardyl world/HUD. The software arm's nearby road click (.43,.44) independently crossed the same exit. Thus these road points are near the Town boundary, not suitable for a small within-Town displacement control. Do not mistake the loader for a settled world or reuse these points as camera-only evidence.

Ardyl initially has substantial terrain occlusion in both arms. In software, ground click (.65,.35) moved out from behind that terrain into a clearing: the subsequent captures show CODEX running, then standing beside a birch. A second reviewed bare-ground click (.78,.60) shows him running across the clearing, with the following camera translating fixed tree landmarks. In software/ardyl-move-idle.png, the birch base is about 90 page pixels right of the character's feet; in software/clearing-settled.png it is about 160 pixels left. The birch remains at the latter location in clearing-idle.png while a skeleton approaches CODEX. These are approximate image-space observations, not decoded guest world coordinates. Actual movement and scene transition are established; combat correctness, audio, FPS and save/reload are not.

Input evidence retains trusted DOM left-button presses, measured canvas geometry, renderer routing and the host's actual consumed WM_LBUTTONDOWN (0x201, wParam 1) / WM_LBUTTONUP (0x202, wParam 0) for guest hwnd 0x10001. The observational onInputTrace callback changes no routing. Owning main legacy endpoint 1 advances throughout: software native completed draws at Town before/idle are 460775/507742, after the exit 594423/639937, and across the final clearing early/settled/idle 842430/879323/912682. Software fallbacks remain zero. WebGL actual submissions at Town before/idle are 22930/23778 and at Ardyl arrival/idle 41919/43014; consumed draw calls advance 287249/294337 → 457109/467261, with zero observed errors/fallbacks and zero direct/fallback software completions. declinedClears=1 remains explicit. Child rendering is unsampled; no general rendering-parity claim.

Software acknowledged quit, logged PAGE CLOSED and exited 0. WebGL's outer probe exited 143 before acknowledging the queued quit (cause unknown); its exact remaining private Chrome PID was identified by the unique profile and terminated. Postflight verifies both owned profiles have no running Chrome, the private server is stopped and port 18164 is closed. No Boat was created or touched, so no box lifecycle handoff applies. Evidence, commands, exact hashes, input logs, renderer samples and an archive-backed recovery recipe are retained in the run. No runtime defect blocked control and no code repair was needed. Next useful work is terrain/foreground occlusion diagnosis using captured draw inputs, rather than changing input or repeating camera qualification.

2026-10-30 owning foreground alpha diagnosis and repair

Evidence: scratch/runs/20261030T2124Z-darkstone-occlusion/. Retained baseline is still module 6085eff5…, source ba96fc3a…, original installed 15-file fixture. baseline-equivalence.json verifies all 1292 frozen served paths against the retained browser archive. This is deliberately distinct from the repaired main runtime below.

The apparent terrain wedge is foreground geometry, but its opacity has a renderer defect. In ardyl-default-draws.json, the submitted tree atlases d3dim17913/d3dim17920 cover native pixel (320,300). Road depth is 0.027448246; foreground draws 3311/3348/3350 have depths about 0.009261/0.008689. All have positive RHW, enabled LEQUAL depth testing, no depth write, SRCALPHA/INVSRCALPHA blending, and diffuse alpha 64. The tree atlas and the game’s F10 overhead view identify legitimate canopy; removing it analytically reveals road. Neither depth ordering nor the atlas alone justifies its former full opacity. The GPU descriptor selects opaque texture alpha (255), discarding the vertex fade. A bounded owning software trace independently records the original surface as RGB565 (format1, 16bpp, type2) with legacy render state21 modes2/4. Full raw input bytes, texture copies, screen-point analysis and inspected frames are retained. The earlier ardyl-draws.json filename actually refers to a Town building; use ardyl-default-draws.json for the Ardyl causal sample.

Legacy D3DTBLEND_MODULATE selects diffuse alpha for a texture without alpha; MODULATEALPHA multiplies texture and diffuse alpha. This agrees with Wine’s legacy device implementation (fixup_texture_alpha_op and TEXTUREMAPBLEND translation). Merely decoding RGB565 to RGBA does not give the original texture an alpha channel. Main repair 328b6631a42bbb50267478f391a4e6b7f3ebcd11 tracks the active legacy blend mode separately from raw TSS queries, resolves original format at draw time, and shares lowering between native software and GPU descriptor paths. An explicit stage-zero combiner setting overrides legacy mode; other-stage/filter changes preserve it. Other legacy blend modes and unknown format behavior retain their existing fallback.

Focused native combiner/descriptor controls pass for RGB565/XRGB1555/XRGB8888, alpha-bearing formats, MODULATEALPHA, raw Device3/7 queries and explicit TSS overrides. Disabling the lowering rejects the observed no-alpha case (2 !== 3). Existing software/GPU wrapping checks and the required locked full build pass. Evidence logs retain exact commands and outcomes.

The repaired original-game retest uses committed-main source 328b6631a…, compiled at 21:43:26Z to module 4649ee5fcf5a10a0212e690d826940bf0ba6445bea678bcac3978c0464bd67c6 (1,891,324 bytes). The dedicated compile reads committed source, excluding another worker’s concurrent CRT WIP. repaired/pinned-runtime.tar.gz and its source/dependency manifest freeze the runtime; actual HTTP module hashes match in both Chrome151 arms. No input or asset repair was applied.

Personally inspected repaired/town-live.png shows the original palisade fading and exposes CODEX/road, where baseline town-idle.png was opaque. repaired/ardyl-live.png and ardyl-fixed-draw-capture.png similarly reveal CODEX through the original canopy. The owning GPU capture includes 58 alpha64 submissions using alphaOp4, replacing opaque alpha selection. A separate bounded observational replay tap in the repaired software owner records 100 alpha64 commands with original RGB565/16bpp and RS21=4/active legacy=4 (repaired-software/alpha-original-format.json). This tap skips non-TL records and records its exclusions; it changes no render result and is not a complete frame trace. The standard repaired format trace stopped on an unsupported legacy call with zero rows; that limitation is retained, rather than represented as successful raw capture.

Both arms repeat native original movement from Town into Ardyl, then a reviewed bare-ground click (.65,.35) moves CODEX across the clearing. Inspected before/running/final frames show translating tree landmarks and standing beside the birch (WebGL 21:48:32Z, software 21:50:47Z). move-idle.png in WebGL still depicts running; use move-final.png for standing. The loader screenshot is not counted as the settled world. Software town-idle.png, ardyl-before-move.png and move-final.png independently show the repaired fade and displacement. This is the requested post-repair gameplay check; prior movement remains established.

Main endpoint1 renderer counters advance with zero observed fallbacks; WebGL samples have zero errors and zero direct/fallback software completions, while software native completed draws advance. Child rendering is unsampled. No native Windows reference, all-scene rendering parity, audio, performance, combat or save/reload qualification is implied. WebGL acknowledged quit and PAGE CLOSED+281.7s, but the outer harness returned143 (cause unknown). Exact counters, times, limitations, cleanup and software exit are recorded in the run’s result/README. No Boat or parent lifecycle was touched.