v0.26.0

Linux DRM/KMS backend — feature = "linux-drm" runs mirui directly against /dev/dri/cardN with a connector + CRTC + N dumb buffers, page-flipping via legacy drmModePageFlip on every frame_end. Companion N-buffer dirty mirror in core: Surface exposes multiple framebuffer slots and mirui copies each frame's dirty rects + scroll shifts into the inactive slots so the next rotation finds them ready. A handful of fixes ride along — hit_test no longer routes taps through scrolled-off rows, the cursor leaves no smear on a scrolled overlay, and SimAction::drag emits at least one PointerMove on low-tick-rate hosts.

Added

• feature = "linux-drm" — opt-in DRM/KMS backend on drm 0.14, drm-ffi 0.10, evdev 0.13, libc 0.2, and signal-hook 0.3. Mutually exclusive with linux-fb at compile time. Drives the full legacy modeset (acquire_master_lock → connector + mode + CRTC selection → dumb buffer × N → add_framebuffer → set_crtc), mmaps each dumb buffer for the surface's lifetime, shares the evdev pump and DPI scale heuristic with linux-fb, and rejects simpledrm / simple-framebuffer / efifb upfront so callers fall through to the fbdev path.
  • Page-flip + vblank wait — advance() queues a non-blocking drmModePageFlip with PageFlipFlags::EVENT; frame_end() polls the drm fd to drain the event before the next paint. An open-time probe gates the path so old simpledrm-style drivers fall back to set_crtc + wait_vblank.
  • Paravirtual scale + dirty notification — virtio_gpu / vmwgfx / qxl report bogus mm panel sizes; LinuxDrmConfig::scale = AutoDpi { .. } is downgraded to Fixed(1.0) for those drivers (caller Fixed(_) honoured regardless). flush() issues MODE_DIRTYFB so paravirtual cards sync the dumb buffer to the host (EOPNOTSUPP on real hardware is a non-error).
  • LinuxDrmConfig — card_path, connector_filter (matches xrandr / kmsprint-style names like HDMI-A-1), mode (Some((w, h)) matches by exact dimensions), input_path, overscan_inset_percent, scale, buffer_count (default 2).
• N-buffer dirty mirror core — Surface::buffer_count() (default 1, skips the mirror loop). FramebufferAccess::all_buffers() returns every slot with idx 0 = active; FramebufferAccess::advance() rotates to the next slot. RendererFactory::mirror_and_advance(backend, plan, transform) is default no-op; SwRendererFactory overrides it to copy each frame's dirty rects + scroll shifts into every inactive slot before rotating. The shared mirror primitives live in src/surface/mirror.rs.
• Surface::frame_end() — per-tick hook after render / render_dirty, called even on empty frames so vblank-bound backends can wait. Default empty.
• ColorFormat::BGRA8888 — byte 0 = B, byte 3 = A. LinuxFbSurface::format_from_var picks it when var.red.offset > var.blue.offset (the BGRX layout signal); LinuxDrmSurface pins it for DRM_FORMAT_XRGB8888. Cross-backend conversions wired through sdl_gpu (PixelFormatEnum::BGRA32) and the wgpu / web-canvas byte swap. Fast paths in src/draw/sw/{rect_fill, blit_fast, blend}/ aren't mirrored yet, so BGRA frames go through the slow path; RGB565 panels are unaffected.
• gallery::demos::widgets viewport scaling — DemoSize::for_viewport(view_w, view_h) reads backend.display_info() so the demo fills any aspect ratio. scale = min(w, h) / 128 keys row + tabbar heights to the shorter axis.
• gallery/examples/linux_drm_demo.rs — runs the shared hello scene on the DRM backend.
• Env var overrides — MIRUI_OVERSCAN_INSET=<n>, MIRUI_DRM_CARD=/dev/dri/cardN, MIRUI_DRM_CONNECTOR=HDMI-A-1, MIRUI_DRM_BUFFERS=<n>, MIRUI_DRM_MODE=WxH, MIRUI_SIM_OFF=1 (skip widgets demo's auto-cycle).

Fixed

• hit_test no longer routes taps to scrolled-off rows. A row scrolled past the top of its container had its post-shift rect drift outside the container's layout rect, and last-hit-wins traversal then let it steal taps from siblings outside the scroll — most visible as a TabBar above a LazyList where the list's first screen kept eating taps that should have hit the tab. hit_test now walks the parent chain and rejects any candidate whose probe falls outside an ancestor scroll container's visible viewport.
• Cursor leaves no smear on a scrolled overlay. When the cursor moved between frames and sat above a scrolling container, the in-place self-blit on the scroll area carried the previous-frame cursor pixels by (dx, dy), leaving a strip the existing overlay-rect repaint didn't cover. The dirty walker now records each out-of-scroll PrevRect and a post-walk pass folds prev ∩ shift.area translated by the shift back into the dirty bounds.
• SimAction::drag emits at least one PointerMove on low-tick-rate hosts. A 100ms drag on a 16-30ms tick went action_started=false → true → elapsed ≥ duration in two iterations and skipped the move branch entirely; the gesture recogniser then classified the result as a Tap. SimTimeline::drag_move_emitted now gates the final PointerUp: if no move ever fired, the catch-up tick clamps t to 1.0 and emits the move first.
• N=2 buffer-role rotation in LinuxDrmSurface::advance. advance() was page-flipping the slot mirui was about to paint instead of the one it just finished painting; the bug was masked on virtio_gpu by paravirtual scanout latency but surfaced as per-frame tearing on real hardware. The rotation now reads front_idx before incrementing.
• LinuxDrmSurface::open page_flip probe could hang. The probe called blocking receive_events() without gating on POLLIN, so a driver that returned Ok from page_flip but never delivered the event parked open() forever. Probe now gates on POLLIN and treats poll-timeout as page_flip_supported = false.

Changed

• linux-fb and linux-drm are mutually exclusive at compile time. Both backends drive the same panel — fbdev via mmap on /dev/fb0, DRM via master on /dev/dri/card0. compile_error! fires at lib compile time on Linux when both features are set.

Internal

• mirui::surface::linux::input and mirui::surface::linux::scale extracted from surface.rs (943 → 227 lines). Both gate on any(linux-fb, linux-drm), so the DRM backend reuses the evdev pump and DPI heuristic verbatim.
• mirror::blit_region format / stride / dimensions checks promoted from debug_assert_eq! to assert_eq! so any backend with mismatched per-slot dimensions panics loudly instead of silently corrupting buffers.
• SwRendererFactory::mirror_and_advance runs shift-on-inactives before blit-from-active, matching render_dirty's active-target order; the reverse smears the just-blitted dirty pixels.
• LinuxDrmSurface::flush is a no-op when buffer_count > 1; set_crtc in advance is what re-points scanout, so dirty_framebuffer on the new active wakes the host on a slot it isn't displaying.
• App::needs_full_first_frame — first persistent-path tick on a multi-buffer backend renders the full screen so every inactive slot starts from a known-complete image; cleared afterward.
• drm-ffi 0.10 added to the linux-drm feature for wait_vblank.