feat(win): multi-screen M6 — one D3D11 DP per screen, bound EXTERNAL weaver; nominal viewer, bound lens (D3) and own tracker (D4) on the second panel

[dfattal]

Multi-screen M6 (Windows): one D3D11 DP per screen, bound SR weaver, nominal viewer + bound lens on the untracked panel

Companion of displayxr-runtime #1857 and #1863, both merged and released in runtime v2.30.0. The runtime appended two slots (xrt_plugin_iface::create_dp_d3d11_for_screen, xrt_display_processor_d3d11::set_present_origin, no ABI bump); this PR implements them in the Windows D3D11 DP.

What changes

• leia_dp_factory_d3d11_for_screen: resolves the screen to an SR display id — the binding's own display_id, else this plug-in's per-monitor claim from probe_displays (M0, leia_win_claims_lookup) — and creates the weaver through leiasr_d3d11_create_for_screen.

• Per-screen weaver plan (leia_sr_multi_win.h, plain C, host-tested): from srGetRuntimeCapabilities (routing + binding + lens-binding caps) decide SrWeaverRoutingInfo{EXTERNAL, KEEP_DRAG_SNAP iff a real HWND} + SrDisplayBindingInfo{displayId} on SrWeaverCreateInfoDX11.pNext, with the Linux fallback order (drop the binding, then routing) on a refusal. An unbound DP (every pre-M6 caller) gets an empty plan: SDK routing, byte-identical to today.

• Display query by id: leia_sr_v2_query_display chains the binding on SrDisplayCreateInfo so a bound DP's geometry describes ITS panel; it also reports srDisplayGetIdentifier, which for the unbound query is the runtime's active display.

• Untracked screen = nominal viewer (David, 2026-10-07): a weaver bound to a display that is not the runtime's active one pins srWeaverSetSimulatedViewer (the SDK's own no-face viewer (0,100,600) mm split by 63 mm), re-applied before every weave; its instance is never srInitialized and late latching stays off (nothing to latch from until SR D4 gives that screen a tracker).

• Bound lens (SR D3, LeiaSR 2079+): when the runtime reports SrLensBindingCapabilities::lensPerDevice, that same DP creates a lens bound to ITS display and drives it from the runtime's 2D/3D requests — it must, since an EXTERNAL weaver never votes. The bound vote is context-owned and survives srDestroyLens, so it is disabled before destroy (a window leaving that screen hands the panel back to 2D). On a pre-D3 runtime that screen's lens is left alone, as an unbound handle would switch the ACTIVE panel.

• set_present_origin → srWeaverSetPresentOrigin before every weave on a windowless EXTERNAL weaver (device px relative to its screen; the effective phase is origin + viewport offset, as measured by the SR session). A weaver with the real HWND is left to the SDK.

• Scissor = canvas in process_atlas, alongside the viewport (the runtime contract for segment DPs; the SR weaver reads both).

• The per-weave diagnostic in process_atlas moves from process-wide statics to per-DP state: with two DPs alternating viewports every frame it logged a WARN per frame.

• Tracker per device (SR D4, LeiaSR 2141+): when the runtime reports SrEyeTrackerBindingCapabilities::eyeTrackerPerDevice, a weaver bound to a calibrated non-active display tracks that display's own camera automatically (EXTERNAL included). The screen-bound DP keys on it: tracked ⇒ no simulated viewer (a pin would override tracking), srInitialize on its instance, late latching on — the same as the active display; not tracked (gate tracker_per_device off, older service) ⇒ the untracked-screen path above stands. The bound lens (D3) is independent of tracking. This is what makes a straddling window tracked on both halves once the service gate is on.

• SDK pin → sr-sdk-v2-1.38.0.2141 (tag 27 lens-binding + tag 28 tracker-binding headers; dispatch table unchanged; the two Linux-only compose slots are placeholders, unused on Windows). CMake gates: DXR_LEIA_HAS_SR_MULTI_WIN, DXR_LEIA_HAS_SR_LENS_BINDING, DXR_LEIA_HAS_SR_TRACKER_BINDING.

• Per-monitor display info (get_display_info_for_monitor, multi-screen M1): every monitor this plug-in claimed reports its SR-calibrated physical size (the enumeration descriptor's cm, cached, non-blocking), native pixel dims, a nominal viewer scaled from the active panel's, and MANAGED eye tracking. Without it the runtime's per-segment views (#1863) refused the layout — the second screen had no metres on Windows (the EDID reader there never fills mm) — so a straddling window got ONE view set from the active panel's tracker: the DS1 half weaved from its own camera (D4) but looked around with the laptop's.

• New host test tests/test_sr_multi_win.c in the Linux lane (plan, fallback order, bound-lens gate, tracker gate, simulated-viewer gate).

Verified on the two-panel win rig

LeiaSR 1.38.0+2141, tracker_per_device=on, runtime v2.30.0 (+ the change-gated locate diagnostic of runtime #1870): cube_handle_d3d11_win straddling AL | DS1 — the DS1 DP is created for monitor 0xb72e… -> SR display 0x2 with plan{external=1 keep_drag_snap=0 bind=1}, tracked by its own camera, no simulated viewer, lens BOUND; the locate reports segments=2 … -> total=4; [0] 0.344x0.194 m; [1] 0.340x0.190 m; the pre-weave atlas carries a distinct L/R pair per segment; both halves weave (adjacent-column ratio 1.04 / 1.07, 2D fallback reads ≈0.54); and David confirmed by eye (2026-10-08): "each half weaves and lookaround locally" — hiding the DS1 camera degrades only the DS1 half. The SR session independently verified EXTERNAL + binding + window=0 weaving with per-device calibration on DX11/DX12/GL/VK, a two-weaver pair in one process, and a bound lens driving the DS1's own FPC.

Not in this PR: D3D12 / GL / Vulkan per-screen factories (D3D11 first per David), the service path, per-display trackers (SR D4, tag 28, a further SDK tag).

🤖 Generated with Claude Code