Dark Reign 2: paired movement prerequisite, 2026-11-04

WebGL gameplay remains blocked at original startup: D3D7 mip sampling is missing. Ordinary registered Launch rejects both HAL and RGB Emulation for “doesn't support mip mapping,” then reports “no hardware d3d devices” and “Can't start hardware accelerated DirectX 7.0.” No mission, selection or move command is reached on this backend. This is a first-dependency result for GLD3D-CORPUS-27 work, not paired gameplay success or a historical membership assignment. The corpus status notes leave the exact 27-ID set unresolved.

Evidence: scratch/runs/20261104-dark-reign2-paired-movement/.

Retained software result

Reuse the accepted November 1 ordinary selected-unit movement result, published in dd82ceb03, at execution baseline 84358006a197266af7bb1f17ebf80de605d5d736. I inspected its move-before.png.gz and move-end.png.gz again: the selected vehicle moves from roughly (352,288) to (180,225), and soldiers translate across the ground; the pipe, pillar, crates, ridge and unselected foreground soldier stay fixed. The accepted neutral comparison and consumed Shift-M/ordinary left click establish command attribution. There was no new software execution. This retains evidence at that revision; it does not requalify software on today's module or assert arrival of every selected unit.

The accepted native module was dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4; its recorded companion was c405379ece71b0f89e4230ad4a9f0325a0b4f23abf05ef4a748671030038633e. accepted-software-references.json records existing evidence paths and full hashes without duplicating its archives, media or source trees.

Original WebGL attempt and first dependency

Runtime source is exactly committed main 1188e9bf57ef35bd3363fb88b7ca0da442ce696d, the assigned baseline. Committed files were read with git show and transferred individually to a fresh runtime staging directory; foreign WIP was excluded. No clone, worktree or source archive was created. runtime-pins.json records 1,461 staged file identities; relevant-source.patch is empty for tracked runtime paths at preparation. The original 29 game/manifest pins are in original-media-pins.json. The executable is the existing private fixture test/binaries/candidates/dark-reign2-demo/extracted/Files/dr2.exe, 3,969,076 bytes, SHA-256 6aa93ded56537ff9ceb6f47d0bbe2de74b5b2b83c6932d61eb98e69d797756ed. The guest identifies build 873, August 16, 2000. Media was provisioned at its runtime path, never copied into evidence.

The normal page at index.html?debug&candidates&frozen&gl-renderer=webgl used the actual dark_reign2_demo selector, WebGL renderer selector and Launch button. Frozen stepping changes scheduling only. The first request asked for 64 batches; the guest exited after 12, producing an incomplete-step response, not a frozen-step timeout. browser/startup-64.png was inspected: the page shows the original launch and exit, with no mission. There were no game inputs. The first observer lost the VFS during normal runtime teardown, so the exit code alone was not used as proof of a mip rejection.

One diagnostic replay retained a read-only reference to the live VFS before stepping, without changing guest bytes, config or renderer. Its sole request asked for 16 batches and again completed 12 before exit. The retained browser-diagnostic/guest-logs.json records both mip rejections at vid_enumdx:1164, followed by the fatal startup error. Its startup-exited.png was inspected as well. No retries, mission inputs, synthetic fixture pixels or patched game state were used after this diagnosis. Both requests stayed below 64 batches; Launch used balanced trusted mouse press/release events. No game keys or buttons were held.

The page selected WebGL, but no owning WebGL context was created in either run. The page and attached guest-worker observations establish module compile/instantiate/send/receive identity; they do not establish a GPU draw, GPU vendor, unit displacement or backend parity. Both original runs loaded the canonical module; its retained companion is listed separately:

The canonical bytes were recovered directly from the diagnostic observer and match the staged module. Both module files are preserved under modules/; the compatibility companion was not executed by these Chrome launches. Every successfully served file matched its pin (176 responses per run). HTTP 404 records are retained, including the absent optional stdole2 typelib and DLL search alternatives. Symbol initialization also reports missing data; the explicit device rejection, not that warning or the final exception code, is the diagnosed graphics dependency.

Why there is no caps-only repair

At this exact source revision, src/09ab-handlers-d3dim-core.wat:$d3dim_fill_device_desc7 advertises triangle filter caps 0x030003c3 with GPU enabled. The native-only branch adds 0x0003003c, producing 0x030303ff for software triangles. The original's previously accepted enumeration analysis checks descriptor offset 0x60 against mask 0x3c; the GPU mask has no accepted bit.

The production GPU descriptor lowers filtering to a single linear flag; lib/d3dim-gpu.js binds one image and sets mip: 0. Thus the missing caps reflect missing implementation. A causal generic repair requires explicit mip-chain metadata/upload, separate MIN/MAG/MIP lowering, bounded partial and malformed chain handling, and actual minified/perspective rendered-pixel regressions through the owning WebGL path before advertising these bits. Existing software mip implementation and GL/D3D8/9 sampling do not supply the D3D7 transport contract. That cross-layer implementation is outside this bounded first-dependency task. No production code, caps, tests or assertions were changed. The exact committed closure was recompiled on the leased box with flock /tmp/wine-build.lock node tools/build-compile-wat.js; validation passed and both resulting module hashes exactly reproduce the staged binaries (compile.log, recompiled-module-sha256.json). This is compiler/module reproduction, not the full tools/build.sh gate or a new regression-suite pass. All 1,461 runtime pins still matched after execution/recompilation.

Next work is that D3D7 WebGL mip dependency, followed by the equivalent ordinary Demo Mission 1 / Normal / select group / Shift-M / left-click movement route and inspected before/after frames with an actual owning renderer. Software movement evidence remains retained; paired gameplay remains open.

Runtime handoff

Fresh box bx_44kza28q, TTL 3,600 seconds, expiry 2026-11-04T02:47:04.645Z, shutdownOwner=coordinator, was recorded before use. SSH preflight returned api_key_action_forbidden; it was not retried and no credentials were changed. Separately authorized API boat exec succeeded. The two browser segments ran 01:49:08–01:50:46 and 01:50:46–01:51:08 UTC, well below 25 minutes, and both Chrome processes exited zero. Helpers and browser endpoints were closed; process checks are retained. The box is left RUNNING/idle for coordinator recovery/snapshot/shutdown. The worker performed no stop, delete, resume, public deployment or approval action, and did not inspect or continue the excluded graphics diagnostic.