OrcaSlicer/TEXTURE_DISPLACEMENT.md

# Texture Displacement - Technical Notes

Branch: `feature/texture_displacement`. Reference for the feature as it stands: what it does, how the
algorithms work, and where the code lives.

## What it does

A paint-style gizmo (`GLGizmoTextureDisplacement`) that lets you:
- Paint one or more "layers" onto a model's surface, each a height-map texture with its own
  depth/tiling/rotation/offset/invert/tile-mode/projection-mode/blend-mode.
- Pick a texture from a shipped library (`resources/textures/displacement/`) or import your own
  (saved into `<data_dir>/textures/displacement/`, kept separate so app updates can't clobber it).
- Combine overlapping layers with image-editor-style blend modes (Add/Subtract/Multiply/Divide).
- Preview the true displaced result live, before baking (background job, not on the UI thread).
- Preview via a fast GPU shader instead (no real geometry movement) for a lighter-weight alternative.
- Bake into real mesh geometry on demand, restricted to the painted area only.
- Remesh and subdivide so a low-poly model has enough vertices to show fine detail.
- Unwrap a painted patch with a real CGAL LSCM parameterization and view it in a dedicated,
  dockable 2D "UV Editor" pane.

## Standard vs Pro mode

A two-position slider in the panel header, right of the Dock/Undock button.

**Pro** shows every mesh-preparation control; Remesh, Subdivide and Bake are run separately by the user,
in whatever order they like.

**Standard** hides all of it and folds one fixed recipe into the Bake button, because a height map only
ever *moves vertices that already exist* - painting onto an imported 12-triangle box and pressing Bake
would otherwise do nothing visible. Standard's Bake is:

1. `plan_remesh()` + `replace_mesh_keep_all_paint()` - isotropic remesh to 1 mm, sharp edges above 40
   degrees protected. Gives the subdivider an even starting density whatever the input looked like.
2. `plan_adaptive_subdivision()` + `apply_adaptive_subdivision()` - feature-adaptive refinement, max
   edge 20 mm, detail 0.02 mm, min edge 0.02 mm.
3. `bake()` - the ordinary background displacement job.

Both preparation stages are *planned* before the undo snapshot and *applied* after it, so a stage with
nothing to do is skipped without leaving an empty undo step. The standalone Pro buttons share the same
plan/apply split.

**All three stages sit under one undo step.** `Plater::take_snapshot()` records the state *before* the
change, so a single snapshot taken at the top of `bake_standard()` means one Undo returns the mesh to
exactly what was imported. `TextureDisplacementBakeInput::take_snapshot` lets the caller say who owns
the undo step - true for the Pro-mode button, false for the pipeline, whose background job commits long
after that snapshot's scope has closed.

The presets live in one place (`STD_*` constants) and `apply_standard_mode_presets()` pins the hidden
controls to them every frame while Standard is active, so the live preview cannot disagree with what
Bake will do. Switching to Standard also closes the subdivision preview, whose controls have just gone.

One control survives into Standard: **"Added triangles (k)"**, the subdivision budget. It is deliberately
*not* pinned - pinning would fight the user's own slider every frame - because unlike the rest of the
recipe its right value depends on the part rather than on the method (a big model, or a fine texture,
simply needs more triangles). Default 1500. The widget is one lambda shared by both layouts.

Standard remeshes *after* painting, so the remesh has to preserve paint: `ModelVolume::restore_painting()`
only remaps the four standard channels, so `replace_mesh_keep_all_paint()` additionally runs
`TriangleSelector::remap_painting()` over the eight texture-displacement masks. The Pro Remesh button
goes through the same helper. If the remap comes back empty the pipeline stops with a message rather
than baking a flat mesh.

## Architecture

### Data model (per `ModelVolume`)

Each of up to `TEXTURE_DISPLACEMENT_MAX_LAYERS` (8) layers gets its **own independent
`FacetsAnnotation`** paint mask - the same `TriangleSelector`/`FacetsAnnotation` machinery every other
paint gizmo (FdmSupports, Seam, MMU, FuzzySkin) already uses, just one full instance per layer slot
instead of one per volume. This is what makes layered/blended painting work for free: the same triangle
can be `ENFORCER` in layer 2's mask and layer 5's mask simultaneously, and at bake/preview time each
layer displaces the surface left by the previous one (image-editor-layer semantics).

Whole-stack settings (border handling, post-process smoothing) live beside the layers in
`texture_displacement_options` (`TextureDisplacementOptions`), since they belong to no single layer.

### Bake algorithm (`libslic3r/TextureDisplacement.cpp`)

`build_texture_displacement(base_mesh, layers, facets_data, options)` is **accumulate-then-displace,
and topology-preserving**: the returned mesh has exactly the input's vertices and triangles, in the
same order - only the positions of displaced vertices differ.

1. `its_compactify_vertices()` on a copy of the input. In practice a no-op (it only drops
   *unreferenced* vertices, and preserves the order and indices of the rest). It is there to
   guarantee the index alignment step 3 depends on.
2. Area-weighted vertex normals of the **undisplaced** mesh, computed once. Every layer both projects
   and displaces along these, so a vertex covered by several layers moves along one single well-defined
   direction. Where the paint does *not* cover every triangle around a vertex, the normal is recomputed
   from the painted triangles alone (the union over all layers, so it stays one direction per vertex):
   on the rim of a fully painted top face the whole-mesh normal is the 45-degree bisector it shares with
   the side wall, and displacing along that flares the rim outwards instead of raising it. Interior
   vertices are unaffected - all their triangles are painted, so the two normals coincide. Paint
   coverage per original triangle comes straight off `TriangleSplittingData::triangles_to_split`.
3. For each layer in slot order: deserialize its stored paint mask into a `TriangleSelector` against
   the **base mesh** (never against a previous layer's output), then
   `selector.get_facets_strict(ENFORCER)` → the painted patch. Two facts are exploited:
   - `get_facets_strict()` returns the mesh's **entire** referenced vertex array regardless of which
     state was asked for - only `.indices` is filtered by state. So `get_facets_strict(ENFORCER)`
     and `get_facets_strict(NONE)` share identical vertex indexing, which is what lets boundary
     detection be a plain index check instead of a position-hash lookup.
   - The selector's vertex array *starts with* the mesh's own vertices (extra ones created where a
     brush stroke split a triangle are appended after them), and `get_facets_strict()` emits the
     referenced ones in order. Combined with step 1, **selector vertex index `i` is our vertex `i`**.
     Split vertices live past the end of our array and are simply skipped - they sit on the paint
     boundary anyway (splitting only happens at partial coverage).
4. A vertex used by at least one **unpainted** triangle is a border vertex. Whether it moves is
   `TextureDisplacementOptions::displace_border`, and it does by default. Nothing can tear: the bake is
   topology-preserving, so a border vertex is *one* vertex shared by both regions and moving it simply
   tilts the unpainted triangles that use it. Pinning it instead clamps the outermost ring of relief to
   zero, which on a fully painted face collapses the pattern into a ring of steep ramps at the edge; it
   is kept as an option for when the relief must not spill past the paint at all. Either way the border
   drives the `edge_smoothing` falloff.
5. Per interior vertex: sample the height texture (`sample_layer_height()`, see Projection methods)
   and fold `height * depth_mm * (invert ? -1 : 1)` into that vertex's running total via the layer's
   `TextureBlendMode` (see Blend modes). A `visited` set makes each layer fold in exactly **once**
   per vertex, no matter how many of the patch's triangles share it - otherwise a Multiply/Subtract
   layer would apply two or three times over depending on local triangle fan-out.
6. Move each touched vertex along its (step 2) normal by its accumulated total.
7. Optionally (`TextureDisplacementOptions::smooth_*`) relax the result - see Post-process smoothing.

### Post-process smoothing

`smooth_mesh_vertices(mesh, movable, strength, iterations)` - Laplacian relaxation, run after all layers
have been folded in, restricted to the vertices flagged in `movable`. Each pass moves a movable vertex a
`strength` fraction of the way to the average of its one-ring, read from a **snapshot** of the previous
pass so the result does not depend on vertex order (a Gauss-Seidel sweep would smooth several times as
hard at the end of the array as at the start). Neighbours come from a CSR-style adjacency built once per
call. Topology-preserving, like the bake.

Its job is to round off the hard steps a bitmap height map leaves behind - a different knob from
`TextureDisplacementLayer::smoothing`, which blurs the *height map* before it is ever sampled.

Two ways in, sharing one set of settings on the volume:
- The **"Smooth result"** checkbox + "Smoothing (%)" / "Passes" ride along with Preview and Bake.
  `movable` is exactly the set of vertices the displacement moved, so the untouched part of the model
  keeps its exact geometry and the ring just outside the displaced set anchors the relaxation (the
  relief cannot creep outward).
- **"Smooth baked mesh now"** (`GLGizmoTextureDisplacement::smooth_model()`) applies the same settings to
  the volume's *committed* geometry, for relief that is already baked in. `movable` there is the painted
  triangles' vertices. Because smoothing never touches the triangle list, this is the one geometry
  operation in the gizmo that keeps **every** paint channel verbatim - it saves and restores the eight
  texture-displacement masks around `set_mesh()` rather than remapping or dropping them.

**"Ignore outer ring"** (`smooth_skip_border`, on by default) drops the patch's own outermost ring of
vertices from `movable`. That ring's neighbours *outside* the paint never move, so relaxing it drags the
rim of the relief down toward the flat surface and the pattern comes out half-melted where it meets the
edge. Held out, the border keeps the full depth the texture asked for and only the interior relaxes.
Turning it off softens the outer edge deliberately (a blunter version of the per-layer edge-smoothing
falloff). This is the *smoothing* rim, independent of whether that rim is displaced at all
(`displace_border`, step 4 above); both default to keeping the border sharp.

### Blend modes

`TextureBlendMode` {Add, Subtract, Multiply, Divide}, per layer, applied per vertex against the
total accumulated by the layers **below** it (lower slots). The quantity blended is a signed
displacement in **mm**, not a pixel value.

Add/Subtract are self-explanatory. Multiply/Divide are *scaling* operations and so need a unit
convention: they treat the layer's own value as a **factor relative to 1 mm**. That makes `depth_mm`
a gain, and - the property that makes a Multiply layer usable as a mask - a layer with depth 1 mm
sampling a white (1.0) texel multiplies by exactly 1, i.e. leaves the layers below unchanged.
Divide floors its divisor's magnitude at 0.05: a black texel samples to *exactly* zero, so the divisor
really does hit zero in ordinary use, and an unbounded `1/0` would fling vertices thousands of mm away
and poison the mesh's bounding box (and every plate/print-volume check downstream). The floor doubles as
a cap on how far Divide can amplify the relief beneath it: at most 20×.

The **lowest painted layer ignores its blend mode**: it has nothing beneath it, and Multiply/Divide
against an implicit zero base would annihilate (or blow up) it. Enforced in
`build_texture_displacement()` (the first layer to reach a given vertex always folds in additively) and
surfaced in the UI, which labels that layer "Base layer" instead of offering a control that does nothing.

### Projection methods

Five choices per layer (`TextureProjectionMethod`), all funneling through `apply_uv_transform()`
(scale by `1/tiling_scale`, rotate by `rotation_deg`, add `offset`). They are dispatched by
`sample_layer_height()`, which returns a **height**, not a UV - because Triplanar takes three
texture samples per vertex and so has no single UV that represents it.

- **Triplanar** (default) - samples the texture on all three world planes (`(y,z)`, `(x,z)`, `(x,y)`)
  and blends the three by the vertex's own normal raised to `TRIPLANAR_BLEND_SHARPNESS` (4). Hard-picking
  the single axis most aligned with the normal instead is discontinuous wherever that dominant axis
  flips: on a +X face the planar coordinate is `(y, z)`, on a −Y face it is `(x, z)`, so at the shared
  edge `u` jumps. A weighted blend is continuous across the transition by construction, since the weight
  of the axis being left behind falls smoothly to zero. This removes the hard *seam*; some cross-fade
  blurring in the band right at a 90° edge is inherent to triplanar mapping. A genuinely seam-free wrap
  around a box needs a real unwrap - that is what the LSCM mode is for.
- **Cylindrical** - wraps around an axis through the patch centroid, axis auto-picked as the world
  axis *least* aligned with the average normal (perpendicular to the outward radial normal, as a
  cylinder's own axis would be). `u = angle * local_radius` (arc length in mm), `v = distance along
  axis`. An approximation, not an exact fit for arbitrary geometry, and the axis/centre are not
  user-overridable.
- **Spherical** - longitude/latitude around the centroid, scaled by local radius. Same caveat.
- **LSCM** - real UV unwrap via `MeshBoolean::cgal::parameterize_lscm()` (CGAL's
  `Surface_mesh_parameterization` package, LSCM algorithm). Computed **once per patch** (not
  per-vertex like the others - it's a single global least-squares solve), then each vertex looks up
  its precomputed UV. Requires the patch to be a single topological disk (one connected component,
  one boundary loop) - `compute_lscm_uvs()` returns empty and the layer falls back to Triplanar if not
  (e.g. multiple disconnected painted islands, or a fully closed patch). CGAL's parameterizer needs a
  mesh with no isolated/unreferenced vertices, but `get_facets_strict()` returns the *whole* mesh's
  vertex array - so `compact_patch_with_map()` builds a clean sub-mesh plus an index map back to the
  original vertex numbering, purely local to this file.
- **ViewProjected** ("From view") - a flat projection along a fixed direction captured from the 3D
  camera, like a slide projector. `capture_view_projection()` takes the camera's right/up axes,
  transforms them into the volume's *local* frame (so the projection rides along if the part is later
  moved), and stores them as `TextureDisplacementLayer::view_project_right/up` (unit vectors, so the
  projected coordinate stays in mm and `tiling_scale` keeps meaning mm). `sample_layer_height()`
  projects `Vec2f(dot(pos, right), dot(pos, up))`. Single-valued per point, so - like LSCM but unlike
  blended Triplanar - the fast preview and UV-check overlay precompute it per vertex
  (`compute_layer_vertex_uvs()`) and drive the shader's `use_vertex_uv` path. Faces angled away from
  the projector smear; that is inherent to view projection.

  Two companions to this mode:
  - **Projection frame overlay** (`TextureProjectorFrame`, see below) - a semi-transparent window
    dragged over the 3D view whose border becomes the projection's edge. Applying it stores an exact
    **projective** map in `view_project_matrix`, which supersedes the affine `right`/`up` axes above
    for that layer (`view_project_projective`).
  - **"Project only on visible"** (`select_visible_faces()`) - repaints the layer with exactly the
    facets the camera can see, so the projected area matches the viewpoint the projector was captured
    from. Two tests: a facing test (normal vs. view direction, per triangle - under perspective the
    view direction varies across the model, so it is taken from the eye to each centroid), then
    `MeshRaycaster::get_unobscured_idxs()` on the survivors to drop facets hidden behind other
    geometry, so a concave part's far inner wall is correctly excluded. One ray query per front-facing
    facet, hence click-driven (on the checkbox and on each "Capture current view"), never per frame.
    It **replaces** the layer's paint rather than adding to it - "project onto what I can see" would
    otherwise accumulate every angle the user had ever looked from.

### Manual seams and island cutting

`TextureDisplacementLayer::lscm_seam_edges` - undirected mesh-vertex-index edge pairs the unwrap is
forced to cut along, on top of the dihedral-angle seams. `segment_into_charts()` takes a set of these
(translated from mesh → compacted-patch numbering inside `compute_patch_unwrap()`) and refuses to
union two triangles across a marked edge whatever their angle. Both the unwrap cache key and the
gizmo's `UVEditorState` include the seam list, so marking a seam (which leaves the paint mask
untouched) still forces a re-solve. Like the paint masks, seams are mesh-index-space and so dropped on
any topology change.

Two ways to write to it:
- **Mark seam (manual)** - a "Mark seams" click mode (`m_seam_edit_mode`) that suppresses painting. A
  click raycasts the volume (`m_c->raycaster()->raycasters()[idx]->unproject_on_mesh()`, `idx` = the
  volume's slot among model-part volumes), finds the facet's edge nearest the hit point, and toggles it.
  Marked edges render as a red overlay (`render_seam_overlay()`), pulled toward the camera so they read
  on top. This is the Blender mark-seam workflow.
- **Cut island (auto)** - `cut_island()` takes the selected chart's triangles (back-mapped from the
  unwrap via `source_vertex`), finds their 3D bounding box, and marks every edge that straddles the
  mid-plane perpendicular to the longest axis. The re-unwrap then splits the chart across its narrow
  waist. Exposed as the UV pane's **Cut** button.

### UV-check overlays (checker / distortion)

`resources/shaders/{110,140}/texture_displacement_uvcheck.{vs,fs}`, one shader with a `mode` uniform,
drawn over the painted patch (`rebuild_uvcheck_mesh()`/`render_uvcheck_mesh()`, P3N3T2: `normal.x` =
distortion, `tex_coord` = uv), pulled forward with a polygon offset. **Checker** samples a procedural
checkerboard at the layer's uv (per-vertex for LSCM/ViewProjected, in-shader triplanar otherwise) -
squares that stay square mean low distortion. **Distortion** colours each triangle blue→green→red by
`log2(uv_area / surface_area)` centred on the patch's *median* stretch (so a globally-scaled unwrap
reads as uniformly ideal and only relative stretch shows), averaged to vertices. A separate **Show mesh
wireframe** toggle draws the whole volume's triangle edges, rebuilt only when the vertex count changes
(not per stroke).

### Tiling

`DecodedHeightTexture::sample(uv, tile_enabled, tile_method)`. Two tile methods when enabled
(Repeat, MirroredRepeat). **When `tile_enabled` is false, sampling outside `[0,1)` returns `0`
directly** rather than clamping the *coordinate* into range, which would smear the border row/column of
pixels outward to infinity in every direction (streaky lines radiating out from the painted patch).

### Subdivision — two modes

**Uniform (`subdivide_mesh_uniform()`)** — whole-mesh, 1-to-4 split. Recursive edge-midpoint split with
a shared per-pass midpoint cache (keyed by sorted vertex-index pair) so triangles sharing an edge get
the *same* new vertex - capped at `max_iterations` (default 6). Whole-mesh so it never leaves a
T-junction, at the cost of densifying everywhere. Wired as a "Subdivide steps" slider (**0–5**, 0 =
no subdivision), Apply snaps back to 0. Drops texture-displacement paint (no remap) via the standard
`save_painting()`/`set_mesh()`/`restore_painting()` dance; the other four channels are remapped.

**Adaptive (`subdivide_mesh_adaptive()`)** — refine **only the painted area**, by **Rivara longest-edge
bisection**, which is *conformal by construction*. Only **terminal** edges are ever bisected - an edge
that is the longest edge of *every* triangle sharing it - which splits both those triangles along one
shared midpoint at once, so a hanging node is never created. The edge to split for a triangle that wants
refining is found by **longest-edge propagation (LEPP)**: walk to the longest edge of ever-longer-edged
neighbours until a terminal one is reached, and bisect that. Edge length strictly increases along the
path (ties broken by mesh-vertex key, which both sides of an edge compute identically), so the walk
cannot cycle, and Rivara's result is that repeating it refines the original triangle in a bounded number
of bisections. The transition triangles it pulls in just outside the painted patch are the graded band
that makes the size change conformal.

The win: a small decal on a big model no longer quadruples the *whole* model's triangle count.

**Run to completion, worst-first, against a triangle budget.** The refinement loop is not a fixed number
of sweeps: it holds every triangle that is over its criteria in a max-heap keyed by *how many times over*
it is, pops the worst, walks its LEPP, bisects, and re-scores. Edge adjacency (`nb[e]`, the triangle
across each edge) is built **once** and maintained incrementally through each bisection, so the cost
scales with the refined region rather than with the whole model. `max_triangles` is the only bound;
stopping on it leaves a perfectly valid, still-conformal mesh that spent its budget on the largest
errors. A fixed sweep count instead spends itself grading the *coarse surroundings* - whose edges are
the longest, so they win every terminal-edge contest - and never reaches the painted patch.

**It carries the paint forward**, which is what makes it usable (uniform subdivide drops paint). Because
the refinement is *driven by* the paint, the remap is trivial: `subdivide_mesh_adaptive()` fills an
`out_source[new_tri] = input_tri` map (children inherit their parent), and the gizmo rebuilds each
layer's mask on the new mesh - a new triangle is painted iff its source was fully painted in that
layer. `collect_paint_region()` derives both:
- the union refine-region: **exactly** the original triangles the brush touched, read straight off
  `TriangleSplittingData::triangles_to_split` (`serialize()` records an entry per original triangle that
  is either split - i.e. partially painted, the patch boundary - or carries a non-default state). No
  dilation: marking every triangle that shares a *vertex* with the patch drags in a whole fan of huge
  unpainted neighbours and refines *those* down to the resolution floor, since the height field the
  detail test samples is not restricted to the painted area. The conformal closure already grades the
  size change outward on its own.
- the per-layer fully-painted-triangle sets (a `get_facets_strict(ENFORCER)` sub-triangle with all three
  *original* vertex indices == a whole, fully-painted original triangle; a partial stroke's sub-triangles
  always carry a split vertex).

The other four channels ride the normal `restore_painting()` remap.

Both modes share the gizmo's Preview/Apply/Done flow; the **"Only painted area (adaptive)"** checkbox
picks the mode, and the adaptive preview follows the paint live (`rebuild_preview()` refreshes the
wireframe while the subdivide preview is open in adaptive mode). The panel shows the previewed triangle
count.

**Feature-adaptive (follow texture detail).** A sub-mode of adaptive (the **"Follow texture detail"**
checkbox) that puts triangles where the *displaced surface actually bends*, not evenly. A flat region or
a linear **ramp** needs no extra vertices (linear interpolation is exact for a ramp); what needs them is
**curvature** - the *second* derivative, not the gradient. So the extra predicate is a **chord-error**
test: sample the combined displacement at the triangle's three edge midpoints *and its centroid*
(sampling the interior is what catches a bump sitting inside a triangle, the blind spot of an edge-only
test) and take the largest departure from the flat triangle's barycentric interpolation. Refine while
that exceeds `chord_tolerance_mm` ("Detail (mm)"). Zero chord error on a ramp ⇒ untouched; high on a
bump/ridge/noise ⇒ refined until captured. Same conformal machinery, so still crack-free. The
per-triangle error is cached and recomputed only for the children of a split.

Four knobs bracket it, and all four matter:
- **"Max edge (mm)"** (`target_edge_length_mm`) is a **baseline that applies in feature mode too**.
  Without it the chord test aliases: a big triangle over a fine pattern can sample four points that all
  land at similar heights, report no error, and stall before refinement ever starts. The baseline
  guarantees a sampling density fine enough for the curvature test to see the texture at all.
- **"Detail (mm)"** is the chord tolerance above.
- **"Min edge (mm)"** is a hard floor under both, and is what guarantees termination across a sharp
  texture *step*, where the error never falls however fine the mesh gets.
- **"Added triangles (k)"** is the budget, passed as `max_triangles` (the model's own triangle count plus
  the slider, so the control still means something on an already-dense model).

The height field is `make_combined_displacement_sampler()` - it mirrors `build_texture_displacement()`'s
per-layer setup (decode, patch centroid, cylinder axis, blend order, "lowest layer folds additively")
but evaluated per point. Two deliberate simplifications, both erring toward *more* detail (safe -
over-refinement is never a crack): every sampleable layer is sampled at every point (no per-point paint
test), and edge-smoothing falloff is ignored. The first is *why* the refine region must not be dilated -
outside the paint the sampler still reports full relief. **LSCM layers are skipped** (no per-point UV); a
purely LSCM stack yields a null sampler and the code falls back to the length baseline alone. Per-vertex
heights are sampled lazily, so a small patch on a huge model never pays for the rest of it.

### Fast preview (GPU-only, no CPU meshing)

`resources/shaders/{110,140}/texture_displacement_bump.{vs,fs}`, registered as
`"texture_displacement_bump"`. Shades the *displaced* surface without moving geometry - active-layer
only, selected from the View row, and the default when the gizmo opens (`m_use_bump_preview = true`).
Vertex format is `GLModel::Geometry::EVertexLayout::P3N3T2`: `normal.x` carries the per-vertex paint
weight (0/1), `normal.y` flags the UV island currently being dragged, and `tex_coord` carries a
precomputed texture UV, so it can use `GLModel` normally instead of a hand-rolled VBO/VAO manager.

The mesh is **flat** (vertices not shared between triangles): every corner of a painted triangle gets
weight 1, every corner of an unpainted one weight 0. A coarse mesh needs that - one painted face of a raw
cube has no strictly-interior vertex, so per-vertex weighting would either bleed onto the neighbours or
vanish outright. Duplicating vertices costs no shading quality here because the shader takes its surface
normal from screen-space derivatives of position, not from a per-vertex normal.

**Both preview meshes work in the patch's vertex space, not the mesh's.** Those agree only until a
*brush* stroke splits a triangle: `get_facets_strict()` then appends the split vertices, so the patch
array is longer. `rebuild_bump_preview_mesh()` and `rebuild_uvcheck_mesh()` therefore index
`patch.vertices` throughout. The weight buffer is rebuilt at the same cadence as the true-displacement
preview (stroke-end/slider-release) but from the **live** `TriangleSelector` state, not the flushed model
facets, so it does not lag by a full model round-trip.

The perturbed normal is the analytic one for a height field `H = ±depth_mm · h(uv)` displaced along
`N` over any orthonormal surface tangent pair `T`/`B`:

    N' = normalize(N − (dH/da)·T − (dH/db)·B),   a = dot(p,T), b = dot(p,B)

The two slopes have to be genuine **mm-per-mm** derivatives for the preview's apparent depth to match
the bake's.

**Two projection paths (`use_vertex_uv` uniform):**
- **Triplanar (`use_vertex_uv = 0`)** - `uv` and the `T`/`B` axes are both derived in-shader from
  the dominant normal component, mirroring `project_planar()`/`apply_uv_transform()`, and the slope is
  formed analytically. `T`/`B` are the projection's axis-aligned pair, exact only when the face is
  axis-aligned; the shader drops the along-normal component to keep the gradient in the surface. Here
  one `uv` unit is exactly `tiling_scale` mm, so the `1/tiling_scale` gradient factor is right.
- **Precomputed UV (`use_vertex_uv = 1`, used for LSCM and ViewProjected)** - `uv` comes per-vertex from
  the CPU (`compute_layer_vertex_uvs()`, so island placement + tiling/rotation/offset are already folded
  in), and the perturbed normal is built with **Mikkelsen's method** ("Bump Mapping Unparametrized
  Surfaces on the GPU"): the surface gradient taken directly from the screen-space derivatives of the
  *sampled height* and position. **This makes no uv→mm scale assumption**, which is essential, because an
  LSCM map is **conformal, not isometric**: it is globally area-scaled but the *local* mm-per-uv varies
  across the chart, so a single global `1/tiling_scale` factor gets the apparent depth wrong. `dFdx(h)`
  captures the true on-screen rate of change however the chart is stretched. This path is also what makes
  the fast preview follow the UV editor: move an island and its uv - hence its shading - moves with it
  (the mesh rebuilds on drag-end, `on_island_edited(finished)` → `rebuild_preview()` →
  `rebuild_bump_preview_mesh()`). The branch is uniform and the paint weight gates by multiply, so the
  texture derivatives stay well defined. A triangle straddling a seam has a discontinuous uv → the
  `det≈0` guard skips it (a localised preview-only artifact, never in the bake).

**Parallax (triplanar path).** Perturbing the shading normal alone welds the pattern to the base surface:
it does not slide as the camera orbits, and does not get deeper as `depth_mm` grows. The triplanar path
therefore shades at the point the *displaced* surface would show at this pixel, found by **ray marching**
(parallax occlusion mapping). A point at ray parameter `s`, i.e. `P + V·s` (`P` the base point, `V` the
unit direction to the eye), sits at height `s·dot(V,n)` above the undisplaced surface. The displaced
surface lives in a shell between the extreme values of `amp·(h − midlevel)` - taken from both ends of
`h ∈ [0,1]`, so it holds for an inverted layer and a raised midlevel too, where the surface sits *below*
the undisplaced one. The march starts at the top of that shell, where the ray is outside the surface by
construction, and steps inward until the ray height drops below the sampled height. That crossing *is*
the visible point.

Solving `Q = P + V·(H(Q)/dot(V,n))` by fixed-point iteration instead is geometrically exact but the
divisor goes to zero edge-on; the sample then lands a large fraction of a tile away and the iteration
oscillates, which reads as a second, flat copy of the pattern ghosted over the real one. Clamping the
step to one tile does not help - a tile-sized shift lands on the neighbouring tile, the same pattern
again. Offset limiting (stepping along the tangential part of `V`) is stable but understates parallax
enough that the relief still flattens as soon as the camera tilts. Marching has neither problem.

The hit is interpolated between the last two samples, which keeps `PARALLAX_STEPS` (24) affordable, and
the whole march is skipped when sweeping the shell would move the sample point less than half a texel -
the head-on case, so the common view pays almost nothing. The 140 variant samples with
`textureLod(…, 0.0)` inside the loop, since implicit derivatives are undefined in non-uniform control
flow. Two uniforms exist for this: `midlevel` (parallax needs the real height, not just its derivative)
and `eye_model_pos` (the camera in the volume's local frame).

Parallax cannot change the model's silhouette or cast shadows; the View row's Normal mode is one click
away for that. The LSCM path stays plain Mikkelsen bump - it has no closed-form uv, so there is no cheap
way to re-project a marched position. One further approximation: the GPU sampler's wrap mode stands in
for `tile_enabled`/`tile_method`, so with tiling *off* the GPU repeats where the CPU returns 0 outside
`[0,1)`.

### On-canvas "Adjust Texture" gizmo

A per-active-layer toggle ("Adjust placement") that disables painting and shows a flat pan panel (free
2D drag on both axes) plus two arrows along the patch's own U/V axes (constrained single-axis drag).
Anchored to the painted patch's centroid/average-normal (`compute_layer_paint_anchor()`). Hit-testing is
screen-space distance/point-to-segment, not real 3D ray intersection against the handle geometry - simple
and good enough at this handle size.

### Projection frame overlay (ViewProjected)

`src/slic3r/GUI/TextureProjectorFrame.hpp/.cpp` - a semi-transparent, resizable `wxFrame` the user
drags **over the 3D view**, like a slide projector's gate. Whatever the model shows through it is what
the texture is projected onto, and the window's border becomes the hard edge of the displacement.
Press **Apply projection frame** and the gizmo reads the window's rectangle and commits it.

The window is deliberately **dumb**: it owns no placement state and reports nothing continuously. Its
position and size *are* the placement, read on demand at Apply - which is also when the expensive
visible-facet raycast runs. So dragging it is free and nothing recomputes until asked.

Plain 2D (`wxPaintDC`), not a `wxGLCanvas`: a second GL canvas would have to share the app's one real
`wxGLContext`. It only ever draws a bitmap and a border.

**The projective mapping (`apply_projection_frame()`)**. The frame defines a **screen-space** rectangle,
but the bake samples from a **local-space** position, so the two have to be reconciled.
`view_project_right/up` can only express an *affine* projection - exact under an orthographic camera, but
wrong under perspective, where the near end of a part projects larger than the far end and no pair of
axes reproduces that. So the layer instead stores a full projective map (`view_project_matrix`, row-major
3×4, `uv = (row0·p̃/row2·p̃, row1·p̃/row2·p̃)`), built like this:

- `K = projection · view · (instance · volume)`, i.e. local → clip, the same product the renderer uses.
  Note `Camera::get_projection_matrix()` is typed `Transform3d` (nominally affine) but its perspective
  form explicitly writes a `(0, 0, −1, 0)` bottom row into the underlying 4×4, so `clip.w = −z_eye` is
  genuinely carried. The build therefore multiplies **`.matrix()` products** (plain `Matrix4d`), never
  `Transform3d` products, which would not compose that row correctly.
- Window coordinates follow `igl::project`'s convention (as `CameraUtils::project` does), with y
  measured downward. Writing `uv = (win − rect_origin) / rect_size` makes u and v affine in
  `ndc = clip.xyz / clip.w`; multiplying through by `clip.w` leaves a plain linear combination of `K`'s
  rows, which is exactly the 3×4 matrix - the perspective divide survives intact.
- `w > 0` is checked rather than divided blindly. A point behind the projector has `w < 0` and divides
  to a plausible-looking but **mirrored** uv - the classic way a projected decal reappears on the back
  of a model. `project_uv_projective()` returns false there and the caller treats it as no height.

The map already includes placement, so `apply_uv_transform()` is **not** applied on top of it - the
window's own position and size are the placement, and the tiling/rotation/offset sliders would shove
the result off the frame the user just aligned. A "Clear" button drops back to the affine path where
those controls mean something again.

Apply also sets `tile_enabled = false`, so `DecodedHeightTexture::sample()` returns 0 outside `[0,1)`
and the border is a hard edge rather than the first seam of an endless repeat, and repaints the layer
via `select_visible_faces(&matrix)` - the frame's uv square clips the selection, which both matches the
paint to the border and keeps the ray queries proportional to the framed area instead of the model.

Owned by the gizmo and **destroyed** (not just hidden) in `on_shutdown()`. Closing it only hides it, so
reopening keeps it where it was left.

### UV Editor pane

`UVEditorCanvas` (`src/slic3r/GUI/UVEditorCanvas.hpp/.cpp`) - a standalone `wxGLCanvas` rendering the
flattened LSCM islands (per-island wireframe + outline + fill) over the height texture (background
quad tiled across the whole unwrap), with mouse pan/zoom. It is wrapped in a **`UVEditorPanel`**
(same file) that adds a button row (Frame / Snap / Avg scale / Cut / Join / Unjoin) and a status line
along the bottom naming the current gesture and the shortcuts in play. The *panel* is what is
registered as a `wxAuiPaneInfo` pane on `Plater`'s `m_aui_mgr`; `Plater::show_uv_editor(bool)`
shows/hides it (deferred via `CallAfter`, since the gizmo calls it mid-3D-frame), and
`get_uv_editor_canvas()` returns the inner canvas the gizmo talks to.

Deliberately **shares the app's one real `wxGLContext`** (`wxGetApp().init_glcontext(*this)`, the
same call `View3D`/`Preview`/`AssembleView` make) rather than creating an independent context like
`SkipPartCanvas` does elsewhere in this codebase - this is what lets it reuse the already-registered
`"flat"`/`"flat_texture"` shaders and `GLModel` as-is, instead of needing its own shader
compilation/VBO management.

**Geometry is uploaded once, in the unwrap's own (raw, mm) coordinates**, one `GLModel` set per island;
each island is then drawn through its own 2x3 affine (`island_transform_matrix()` composed with the
layer's tiling/rotation/offset) passed as the `flat` shader's `view_model_matrix`. A drag updates one
matrix per island and touches no vertex buffer - `on_island_edited(!finished)` calls only
`set_island_transforms()`, and the full `set_islands()` rebuild happens solely when the unwrap itself
changes (`unwrap_changed` in `update_uv_editor()`).

**Gestures** (canvas-owned, reported to the gizmo as incremental deltas via `IslandEditFn`): left-drag
= move, right-drag or **R** = rotate (hold **Shift** to snap to 15° steps - quantised on the
*cumulative* rotation, not each delta, so it doesn't judder, and accumulated incrementally so it
survives crossing ±180°), **S** = scale (R/S modal, click/Enter to confirm, Esc to cancel), wheel =
zoom about the cursor, middle-drag = pan, **Home**/**F** = frame all. Scale writes
`TextureIsland::scale`; "Avg scale" (`average_island_scales()`) sets every island to the mean, so
one island scaled by hand can be matched back to its neighbours' texel density. **Snap** (canvas-owned
`m_snap_enabled`, toggled from the toolbar) sticks a dragged island's nearest boundary vertex onto a
neighbouring island's at drag-*end* only - a magnet that re-applies mid-drag is very hard to pull out
of. Toolbar commands the canvas can't service itself (Avg scale, Cut, Join, Unjoin) are forwarded to the
gizmo via `CommandFn`; view-only ones (Frame, Snap) it handles directly.

## File map

**libslic3r (core, no GUI dependency):**
- `src/libslic3r/TextureDisplacement.hpp/.cpp` - data model, bake algorithm, projection methods,
  tiling, subdivision (uniform + adaptive longest-edge bisection), post-process smoothing
  (`smooth_mesh_vertices()`), and `TextureDisplacementOptions` (the whole-stack settings). See doc
  comments throughout, they're kept accurate and up to date.
- `src/libslic3r/MeshBoolean.hpp/.cpp` - `parameterize_lscm()` and `remesh_isotropic()` in the `cgal`
  sub-namespace, reusing the existing `CGALMesh`/`_EpicMesh`/conversion-helper infrastructure already
  there for mesh boolean ops. CGAL includes: `Polygon_mesh_processing/border.h`,
  `Polygon_mesh_processing/connected_components.h`, `Surface_mesh_parameterization/{Error_code,
  LSCM_parameterizer_3, parameterize}.h`. No new dependency - CGAL 5.6.3 is already vendored and the
  `Surface_mesh_parameterization` package headers were already present.
- `src/libslic3r/Model.hpp/.cpp` - the 8 named `FacetsAnnotation` fields + accessor,
  `texture_displacement_layers`, `texture_displacement_options`, and all the mirrored touch points
  (see Data model above).

**GUI:**
- `src/slic3r/GUI/Gizmos/GLGizmoTextureDisplacement.hpp/.cpp` - the gizmo and its whole panel.
- `src/slic3r/GUI/TextureLibrary.hpp/.cpp` - scans the shipped + user texture folders, imports an
  arbitrary image into the user folder (converting it to the 8-bit grayscale PNG libslic3r decodes),
  and loads a library file's bytes for a layer. The image→grayscale-PNG conversion lives here, on the
  GUI side, because libslic3r has no image toolkit; both the import path and the "pick a shipped
  texture" path go through the same one function.
- `resources/textures/displacement/*.png` - the 10 shipped height maps (Bricks, Grid, Hexagons,
  Knurl, Noise, Quilt, Studs, Waves, Weave, Wood Grain). All 512×512 8-bit grayscale and **seamless**
  (each is periodic over the full image in both axes, so tiling shows no seam). Generated
  procedurally; the whole `resources/` tree is installed recursively by CMake, so a new folder under
  it ships with no build-system change.
- `src/slic3r/GUI/Jobs/TextureDisplacementBakeJob.hpp/.cpp` - background bake commit.
- `src/slic3r/GUI/Jobs/TextureDisplacementPreviewJob.hpp/.cpp` - background preview compute
  (mirrors the bake job's shape but commits nothing to the Model).
- `src/slic3r/GUI/TextureProjectorFrame.hpp/.cpp` - the semi-transparent projection-frame overlay for
  ViewProjected layers (plain 2D `wxPaintDC`, no GL context - see its section above).
- `src/slic3r/GUI/UVEditorCanvas.hpp/.cpp` - the 2D UV unwrap viewer widget.
- `src/slic3r/GUI/Plater.hpp/.cpp` - `uv_editor_canvas` member, AUI pane registration,
  `get_uv_editor_canvas()`/`show_uv_editor()`.
- `src/slic3r/GUI/GLShadersManager.cpp` - registers `"texture_displacement_bump"`.
- `resources/shaders/{110,140}/texture_displacement_bump.{vs,fs}` - the fast-preview shader.
- `src/slic3r/GUI/Gizmos/GLGizmoPainterBase.hpp` - `PainterGizmoType::TEXTURE_DISPLACEMENT`.
- `src/slic3r/GUI/Gizmos/GLGizmosManager.hpp/.cpp` - `EType::TextureDisplacement` registration.

## Tests

`tests/libslic3r/test_texture_displacement.cpp`. Covers `decode_height_texture` round-trip, empty-layer
no-op, full-cube uniform displacement, a second layer over the same area contributing, all four blend
modes (table-driven), the lowest layer ignoring its blend mode, border displace/pin, post-process
smoothing and its mask guarantees, and adaptive subdivision: conformality (`every_edge_used_twice` on a
partially-refined cube - an exact crack detector for a closed mesh), the target edge length actually
being reached, the triangle budget capping the result without opening a crack, curvature-driven
refinement (a Gaussian bump refines at its centre, a linear ramp adds nothing), and the max-edge
baseline.

`BUILD_TESTS` is `OFF` in the checked-in build cache; flip it on to run them:

    cmake -S . -B build -DBUILD_TESTS=ON
    cmake --build build --config Release --target libslic3r_tests -- -m
    ./build/tests/libslic3r/Release/libslic3r_tests.exe "[TextureDisplacement]" --order rand