Native Win16 cursor loading and public publication
The generic cursor path now resolves supported predefined system IDs and module-qualified NE GROUP_CURSOR/CURSOR resources, including ordinal IDs and names passed as far strings. USER.69 calls the shared SetCursor implementation and returns the previous narrowed handle. GetCursor and ShowCursor use the same per-thread public state as the host renderer. NULL removes the cursor; invalid handles/modules and missing or malformed resources fail without changing the selection.
NE group entries select an actual child image, whose LOCALHEADER supplies the hotspot. The existing bounded classic DIB decoder copies the AND/XOR planes into owned bitmap records. Loaded cursors retain module ownership, survive DestroyCursor/DestroyIcon, and retire when their module unloads. Unload clears threads selecting those retired records. A shared NE cursor resends its pixels to each selecting host instance because the host CSS cache is private to that instance. Private cursor caching and destruction remain covered by the adjacent cursor/icon suite. Unsupported image encodings fail instead of returning a fabricated cursor.
The existing browser rasterizer represents the monochrome AND=1/XOR=1 desktop-invert combination as opaque black. This fix preserves that policy; the evidence establishes mask transparency and decoded visible pixels, not desktop-XOR parity. System cursor pixels are supplied by the browser from the observed CSS shape; no system-cursor PNG was fabricated.
The real native baseline completed its far-Pascal calls but returned handle 257 for both system IDs and published no cursor event. Its assertion failed on actual distinct-handle semantics. The repaired native fixture executes imported USER/KERNEL calls through the production WASM, not test-only WAT API wrappers. It proves:
- System 0x7f84 and 0x7f85 produce distinct handles and actual host
ew-resizeandns-resizeoutput; previous-handle and NULL results agree with those selections. - Two different NE cursor images, raw CURSOR lookup, repeated shared loads, named far-string lookup, and identical IDs in two loaded modules preserve identity and actual distinct host pixels.
- AND-mask transparency, visible pixels and the two resource hotspots reach the shipping host CUR encoder. The selected PNGs are decoded from that actual encoder output and were personally viewed. They are synthetic fixture cursors, not game captures.
- Invalid module/ID/handle, malformed resource and NULL/invalid destruction fail; shared destruction preserves the resource; unloading one module leaves the other cursor usable, and unloading the selected module clears the public selection.
- A real registered native window gains and loses focus without changing its installed cursor. Two WASM instances over one memory execute native far calls on different input threads, preserve the first thread's selection, and publish pixels into the second host's initially empty cache. Both Pascal stacks restore.
The complete locked build and focused suites cover native cursor publication, native modal destruction, resource paths and handle reuse, shared Win16 allocators, window positioning/DefWindowProc, cursor/icon image creation and lifetime, GetCursor/thread isolation, and renderer cursor input. Exact exits and logs are recorded under scratch/runs/20261114-win16-native-cursor-publication/, including the native/watchdog/supervisor waits, both EOFs, PID absence, build/module hashes and Git patches.
Publication uses a complete HEAD-plus-owned Git-object source view for all build gates, compiler inputs, census files and Node host modules. source-view.cjs supplies the candidate blobs without modifying foreign shared-tree files; its per-process read log identifies each supplied blob. This is an actual compiled and executed source view, not an index-only claim. The three existing DX SDK media-size corrections for camera.x, mslogo.x and sphere2.x are an explicitly coordinator-authorized prerequisite: the committed table failed the official media gate against the actual files. No media bytes were copied or changed. The final shared-tree restoration is checked byte-for-byte, followed by another full locked build and focused checks before SOURCE_READY. Foreign COMCTL/NE/census hunks are excluded from this commit.
Static inspection of the original test/binaries/wep16/WEP4/JEZZBALL.EXE (SHA-256 0450d0559fa5957d813bdd7be1f71122bf23047869b610802386320f109cb705) finds no CURSOR/GROUP_CURSOR resources. WEP4UTIL.DLL also has none. Original segment 1 at 0x290b imports USER.173 with NULL instance and integer IDs 0x7f84 and 0x7f85. The native fixture proves those exact public ID/output distinctions. It does not establish which orientation the original had selected, or prove that cursor loading caused the prior closed run's zero-wall outcome.
No fresh JezzBall original was launched, and the closed original was not replayed. The existing guarded route requires personally reviewed orientation evidence and complete board evidence; its native-cursor route requires a captured cursor image. CSS system-cursor metadata alone does not satisfy that image review. There is no manufactured selected/orientation flag, private game-state read, tracking, solver, gameplay write or restart. Original causal/game-progress proof remains UNPROVED and needs a separately bounded, qualified route. Native Quit was not attempted or conflated with host close. No box was created: ID/expiry null, state NOT_CREATED, stop receipt NOT_APPLICABLE.
Quake3 then TetraVex16 actually released before heavy execution began. Their native/watchdog/supervisor waits, EOFs, heavyReleased and PID absence were personally checked; ancillary outer UNKNOWN statuses remain separate. Final commit, publication and restoration identifiers are in result.json. Git plus the recorded binary patches is the source archive; no source tree, media or WASM copies were placed in the run directory.