DirectDraw system-memory blits and KKnD capability admission — 2026-11-06

The public COM matrix established software system-memory Blt support on both shipped modules, exposed packed-24-bit scalar corruption, and found incorrect success for GetCaps(NULL, NULL). Both defects are repaired. HEL now reports the demonstrated support; HAL makes no new hardware claim. The one original registered KKnD session passed DirectDraw setup, reached the main menu and visible Evolved gameplay, and showed unit selection brackets. Exact mission identity and commanded unit displacement remain unverified. Evidence: scratch/runs/20261106-kknd-system-memory-caps/.

The original KKnD setup rejection is inferred from retained modal stack buffers and original disassembly, not a captured GetCaps HRESULT or branch trace. HAL and HEL both contained dwCaps=0x24041, omitting DDCAPS_CANBLTSYSMEM=0x80000000. Omitting this optional bit is not itself an implementation defect. See the prior original session and its retained capability-boundary.json.

Microsoft documents GetCaps as separate hardware-driver and HEL outputs, and requires failure when both outputs are NULL. DDCAPS distinguishes system-to-video, video-to-system and system-to-system transfers. A measured CPU implementation can establish software support; it cannot establish hardware acceleration.

The public COM fixture generates guest x86 indirect-vtable calls with independently specified slots/argument counts and exact stdcall stack checks. It creates surfaces through DirectDraw7 and includes a legacy DirectDraw1/Surface1 indexed caller-to-video case. Both shipped modules were instantiated directly, with exact module bytes/hash recorded. No private handler is invoked and no synthetic surface metadata is installed. Factory dispatch uses the existing catalog-thunk harness; normal PE loading and COM-thunk initialization precede the calls.

The matrix covers allocated system-memory storage, caller-owned padded storage, video-labelled offscreen storage, full copies, subrectangles, destination-bound trimming, nearest-neighbor stretch, source keys, disjoint explicit clip lists and color fill. Formats are indexed8, RGB565, RGB555, ARGB1555, ARGB4444, BGR24, XRGB32 and ARGB32, with descriptor masks checked. Expected pixels come from an immutable source snapshot and coordinate arithmetic, independently of WAT traversal. Full destination rows, padding, caller allocation guards and source immutability are checked. References, caller storage after Release, descriptor guards, malformed GetCaps output and both-NULL GetCaps are covered.

Blt documents all four memory directions, null rectangles, stretch and color keys. Destination-bound trimming without an attached clipper is recorded as existing emulator behavior, not a native acceptance oracle. BltFast is documented only for display-memory surfaces and must reject an attached clipper. System-memory BltFast success is recorded separately as an existing emulator extension, not capability evidence.

The rejecting controls cover exact DDERR_UNSUPPORTED for BltFast with an attached clipper, unchanged destination bytes on that failure, atomic rejection of a malformed second GetCaps buffer, and exact DDERR_INVALIDPARAMS for both NULL outputs. Formats are matched between source and destination: this matrix does not qualify cross-format conversion, negative pitch, 4bpp, arbitrary ROPs, alpha blending, or all invalid rectangle/flag combinations. Source inspection suggests broader format-admission and BltFast out-of-bounds behavior deserve separate qualification; neither is silently treated as proven support here.

Before repair, each shipped module produced 422 observations and 1,075 COM calls, with 21 identical failures: five BGR24 operations in each of four storage configurations (stretch, keyed stretch, explicit clipping, keyed clipping, fill), plus both-NULL GetCaps. Ordinary packed-24 copies already passed. Scalar paths incorrectly used DWORD loads/stores and four-byte pixel indexing for three-byte pixels, damaging neighboring pixels or padding. The repair uses exact three-byte loads/stores and the actual byte stride. Both-NULL GetCaps now fails before writing either output. The repaired pixel matrix passed on both modules before changing capabilities; see baseline-*.log and repaired-*.log.

Only HEL adds DDCAPS_CANBLTSYSMEM: HAL remains 0x00024041, HEL is 0x80024041. HEL's system-to-video, video-to-system and system-to-system directional caps at offsets 172/216/260 advertise ordinary DDCAPS_BLT=0x40; their additional color-key, effect and ROP fields remain clear. This deliberately conservative report follows the measured CPU path and does not assert GPU acceleration. It does not reinterpret the earlier optional omission as a defect.

Final validation passed 422 observations and 1,076 COM calls per module, including DDCAPS sizes 316/364/380, directional claims, legacy indexed Blt, exact ABI cleanup and lifetimes. The final exact-clipper-HRESULT assertion also passed both modules. flock /tmp/wine-build.lock bash tools/build.sh passed after the production changes. Existing focused tests passed: GetCaps size validation, Blt clipping, explicit clip lists, 3,640 color-key cases and 276 BltFast snapshot/overlap cases. Logs and focused-tests.json, final-qualification.json, exact-reject-qualification.json retain the results.

Shipped module Bytes SHA-256
standard 2,178,624 3b7511d12f3e8d364eb6bf29cdeed8b53af07abc53644d4f186e46f8620da149
compat 2,181,270 33c0e72903f542222ec7bb1330a3a408743765184b4a7b8da451e24faf612f41

The exact predecessor releases were observed in the required order and cleanup was independently checked before the 08:20:03–09:05:03 UTC active window. Builds, tests and the sole original ran serially under shared locks. No boxes, additional workers, guest patches or media copies were used. Git SHA plus binarydiff/new-file patches preserve source identity; existing pinned media is referenced by path/hash. File-only recovery controls and required-size/hash rejection controls passed before original startup. One identity-control harness invocation failed on an absent gate record before opening a browser; the honest gate record was supplied and the generated identity controls passed. This was not an original-game retry.

The original started at 08:28:23 using the established real-origin recovery. Actual parent VFS entries contained authentic DDRAW (352,528 bytes, SHA-256 0a37a508dc6703f0998fb41d8fd5b03d7bbb5d466165dce5b45559d85174ca26, not a static module) and KKND.EXE (465,408 bytes, SHA-256 b2e1a476bfe766edf802c9fa4d0fed2ed841a8d1b50e1a8ee27ceae33afe5d8a). Cross-origin isolation, 512 MB memory, and the actual instantiated owning CPU Worker module were verified through guestWorker.link, owner slot/endpoint identity, and CDP bytecode hash matching the standard module above. The generated two-module identity control demonstrated discriminating actual instances. Startup and terminal executor records retain that identity; no debugger breakpoints or guest writes were used.

Personally reviewed main menu (menu-02.png), Evolved mission menu (evolved-menu.png), gameplay (mission-briefing.png, despite that capture name), and selected infantry (group-selection.png). The route toggled to Evolved and clicked the visible “The Walkover – Easy” text before OK, but the menu supplied no independently established selection marker and the loaded mission name was not recovered. Do not label the scene proven Walkover. The first single-unit click had no clear marker; reviewed drag selection then produced brackets around four infantry. The following right-button action at native (480,330) changed the camera. movement-command.png and movement-after.png therefore do not establish unit displacement under a stable camera. That is the first remaining input boundary; no further gameplay inputs or original restart followed. A read-only detail-crop helper captured the wrong UI region (selection-detail.png); it is not selection evidence. Trusted input records include mouse-down, while completed mouse-up commands, WM button-up messages and zero button state establish release.

The original closed at 08:38:11. Local browser/server/profile and holder/watcher resources are independently checked in final cleanup; no test boxes were assigned (ID/state/expiry N/A). FPS and audio are unqualified. The next task is to establish the actual mission selection and KKnD command semantics, then prove unit displacement with a stable camera under a separately authorized original session. No further DirectDraw capability claim is needed to explain this session's startup progress.