feat(d3d12): multi-screen segments + per-segment views for the in-process D3D12 compositor (M6 + M3)

[dfattal]

What

Multi-screen segments (ADR-047 D2, M6) and per-segment views (M3) for the in-process D3D12 compositor — the path the Unity display provider runs on. The twin of #1857 / #1863 for D3D11, mirroring the pre-split D3D11 behaviour (the #918 split path is untouched; nothing in compositor/d3d11/ changes).

ABI (append-only, no XRT_PLUGIN_API_VERSION_CURRENT bump — ABI stays 5)

• xrt_plugin_iface::create_dp_d3d12_for_screen after create_dp_d3d11_for_screen; announced by XRT_PLUGIN_IFACE_HAS_CREATE_DP_D3D12_FOR_SCREEN (same CREATE_ naming as the D3D11 / Vulkan macros), probed by xrt_plugin_iface_has_create_dp_d3d12_for_screen().
• xrt_display_processor_d3d12::set_present_origin after get_background_preview (slot 24); announced by XRT_DP_D3D12_HAS_PRESENT_ORIGIN, struct_size-gated helper. Plus xrt_display_processor_d3d12_tolerates_resample().
• tests_stereo_camera pins the new iface tail; docs/reference/xrt_plugin_iface.md documents the Windows per-screen slots.
• sim_display implements both on D3D12 (screen-bound instance answers its own screen's metrics and confines viewport + scissor to the canvas), so the CI self-test keeps a complete sim-display.

Runtime

• compositor/d3d12/comp_d3d12_segments.{h,cpp} — the D3D11 manager's twin (screen join, hysteresis lifecycle, decide loop, primary-first weave, flat-2D fill). D3D12 specifics: every segment records onto the one weave command list; each screen's crop has its own slot in a module-owned shader-visible SRV heap (descriptors are read at execute time); explicit COMMON → COPY_* → COMMON barriers on the crops and the source (D3D12 compositor: per-frame ResourceBarrier before-state mismatches (debug-layer id 527) — suspected in editor resize DEVICE_HUNG #747); viewport + scissor = segment before every DP (and the DP sets both itself — a command list has nothing to inherit). No deferred-release list: every weave ends in a full queue wait.
• comp_d3d12_compositor_set_screens + d3d12_segments_tick at the weave site (repaints included). Not segmented: zero-copy, the Epic: weave/repaint next to scanout — output-device split for hybrid GPUs #918 split, the epic: in-process tier parity — panel-rate weave everywhere (VK / D3D12 to match d3d11-hybrid smoothness) #1264 reroute, a zones / Local2D canvas, a pinned session, DXR_SEGMENTS=0, and a back buffer that does not match the client area. Window rect = the HWND's client area in device px.
• M3: comp_d3d12_compositor_get_segment_metrics / set_view_routing; renderer mosaic routing (projection layers per segment rect, the tile's CTS composition: projection layers are blitted opaque over the whole tile, wiping every layer submitted before them (QuadProjectionQuad / ProjectionQuadProjection FAIL) #1598 blend mode decided once; quads + equirect2 once per segment with that segment's camera); comp_d3d12_renderer_blit_rect for the flat fill; a routed frame is never zero-copy.
• oxr: session create / xrLocateViews / xrEndFrame hand-offs for D3D12; the Windows view-set capacity counts a D3D11 or D3D12 factory (oxr_segment_views_win_entry_has_dp, host-tested — a D3D12-only plug-in used to count zero screens).
• cube_handle_d3d12_win: the cube_handle_d3d11_win M3 port (renders every view set at its segment rect, logs view sets: ... once per change).
• Docs: status lines in comp-segments.md and roadmap/multi-screen.md.

Tests

• Local build (scripts\build_windows.bat build, then test-apps): both exit 0. check_displayxr_app.py on the D3D12 cube: clean.
• ctest -C Release: 100/103 pass; new/changed tests_oxr_segment_views, tests_stereo_camera, tests_comp_segments pass. The 3 failures (tests_oxr_view_space_mode_floor, _pinned, _floor_off) are environmental on this box (they need SIM_DISPLAY_OUTPUT=quad to take on the registered DP) and fail identically on a peer worktree build without this change.
• Not yet run on hardware (rig recipe below).

Rig check (two monitors, sim_display or Leia)

1. Plug-in pairing: the Leia side needs DisplayXR/displayxr-leia-plugin feat/win-d3d12-dp-segments built against this branch; sim_display needs nothing.
2. _package\run_cube_handle_d3d12_win.bat, drag the window so it straddles the seam between the primary (system-default) screen and the second one (at least 2 px on each side).
3. Runtime log (%LOCALAPPDATA%\DisplayXR\DisplayXR_cube_handle_d3d12_win.*.log):
   • segments: N screens, plug-in '…' — a window spanning monitors is woven per monitor (multi-screen M6, D3D12) (session start)
   • segments: created a segment DP for screen 0x… ('\\.\DISPLAYn', plug-in '…') via create_dp_d3d12_for_screen (two DPs: the primary + this one)
   • segments: D3D12 window x,y WxH -> … (INFO, the table)
   • segments: D3D12 per-segment views routed — 2 segment(s) x 2 view(s) … (INFO, M3) and the diag(oxr): log the per-segment view-set decision once per change (multi-screen M3) #1870 WARN per-segment views: segments=2 capacity=2 active=… -> total=4; … (a 0.000x0.000 m segment there means the screen has no physical size and the layout refused)
   • App log: view sets: routed [display 0x… views 0..1 at …] [display 0x… views 2..3 at …] (activeViewCount=4, rendering 4)
4. Dragging back onto one screen: segments: D3D12 window … is on the primary screen only — single-DP path, then after ~0.5 s retiring the D3D12 segment DP (INFO).
5. Kill switch: DXR_SEGMENTS=0 → single DP, view sets: single.

🤖 Generated with Claude Code