feat(d3d11): multi-screen segments under the weave-on-scanout split (M6 on hybrid boxes)

[dfattal]

What

Multi-screen M6 on Windows now works on the default in-process D3D11 path on hybrid boxes, i.e. under the #918 weave-on-scanout split (ADR-039). A window straddling two panels is woven per screen without DXR_D3D_FORCE_GPU=scanout / DXR_SPLIT_SAME_ADAPTER=0.

Design

• One segment path, no third path. d3d11_dp_weave is the weave site on both sides of the split (the "split bypasses the windowed weave site" trap is a Vulkan-compositor thing, not D3D11). The !c->split_active gate is removed.
• Segment DPs on the device that presents. The segment manager was already created on d3d11_out_device() and hands its DPs d3d11_out_context(); under the split that is the output (scanout) device, where the primary DP already lives. create_dp_d3d11_for_screen(inst, device, context, NULL, binding, ...) is simply handed the output device. No plug-in change: the Leia factory (leiasr_d3d11_create_for_screen) uses whatever device/context it is given, and the primary DP has always been made on the output device the same way (d3d11_make_dp(..., out_dev, out_ctx, true), which tells the plug-in nothing extra).
• Atlas source = the egress slot. Under the split atlas_srv at the segment site is the egress slot being woven: the output-side copy of the composed atlas, already cropped to the box that slot was painted at. The per-segment crop, viewport + scissor, canvas = segment and set_present_origin are exactly the non-split ones.
• Flat-2D fill moved to the output composite unit. It went through comp_d3d11_renderer_blit_rect, which records on the renderer's (app) context: wrong device under the split. It is now comp_d3d11_outcomp_blit_rect, the same HLSL / linear sampler / opaque blend / rasterizer / depth state re-made on the outcomp's device (compiled on first use). Off the split the outcomp's device and context are the app's, so the pixels are unchanged; the renderer's blit_rect had no other caller and is removed.
• feat(d3d11): multi-screen M3 on Windows — per-segment views in the in-process D3D11 compositor #1863 per-segment views unchanged. The mosaic is painted by the renderer on the app device and crosses the bridge like any atlas.
• R2 resize lag: a slot one frame behind the window is cropped at its own view size against the live canvas: the same one-frame lag the single-DP split weave already has. Unit-tested (the crop still partitions the slot's tile exactly).

What is byte-identical

• A window on one screen under the split (the common case): comp_d3d11_segments_update returns false, and process_atlas runs with the same arguments as before. The only additions on that path are the segment manager's per-weave table update (no GPU work) and, once per manager on a split session that has ≥2 screens, one WARN line.
• The non-split segment path: same crops, same DP calls; the flat-2D fill uses the same shaders and states on the same context.
• The single-DP non-split path is untouched.

Not changed / out of scope

Zones / Local2D frames (non-whole-window canvas), zero-copy frames, shared-texture path, pinned sessions, service/hosted path, HWND-follows-majority.

Tests

• tests_comp_segments: new case, a split egress slot one frame behind a resizing window (tile 800/824/840 against a 1649 px canvas, seam at an odd column) still partitions its tile with the seam where the live window puts it.
• ctest -C Release: all segment / compositor tests pass. 3 tests_oxr_view_space_mode_* cases fail identically on unmodified main on the win box (environment: they need a >2-view DP and this box registers Leia SR).

Rig check (needs the two-panel win rig; not run here)

1. Dev runtime from this branch, no DXR_D3D_FORCE_GPU / DXR_SPLIT_SAME_ADAPTER in the env. cube_handle_d3d11_win launched non-elevated (via explorer.exe <bat> so XR_RUNTIME_JSON is honoured).
2. Place the client area across the seam at desktop x = 3840, e.g. client x 3016..4664 (1648 wide), with a per-monitor-v2 DPI aware SetWindowPos.
3. In %LOCALAPPDATA%\DisplayXR\DisplayXR_cube_handle_d3d11_win.*.log expect, in order:
   • D3D11 output-device split ACTIVE: weave/repaint/present move to '<Intel ...>' (or ADR-039 same-adapter ENGAGE)
   • D3D11 display processor created via factory on the SCANOUT device
   • segments: under the #918 output-device split — segment DPs are created on the OUTPUT device (LUID=..., device 0x...) (new)
   • segments: created a segment DP for screen 0x... ('\.\DISPLAY...', plug-in 'leia-sr') via create_dp_d3d11_for_screen
   • the plug-in's per-DP lines canvas=(0,0 824x...) and canvas=(824,0 824x...)
   • per-segment views: segments=2 ... -> total=4 (oxr_session)
   • absent: any segments: could not create a DP, D3D11 debug-layer device-mismatch errors, #918 R1 refusals beyond the first frame.
4. By eye: each half weaves and looks around from its own viewer, as with the knobs on 2026-10-08. Drag back to the laptop panel only: segments: retiring the segment DP ... after ~0.5 s and the frame is the plain single-DP split weave.

🤖 Generated with Claude Code