Commit Graph
20 Commits
Author SHA1 Message Date
Jackson OurslandandClaude Code 5962f56f87 Viewport: reuse still pictures and draw scenes into a 1x window pass (#4820) (#5187)
* Viewport: show the picture it drew last when nothing changed (#4820)

A frame redrew the whole scene whenever egui asked for a frame, even when
the camera, scene, highlight and overlays were identical: pointer moves over
empty space and UI animations over a still view paid for the full scene.

The renderer now hashes everything the picture depends on (scene keys,
camera, size, section, images, analysis, accessibility map, occlusion) and
blits the last offscreen picture when the hash matches. The first still
frame filters a supersampled picture down once and later frames copy it as
it is. A frame with the same content costs a blit instead of a scene draw.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* Viewport: draw the scene offscreen, into a 1x window pass (#4820)

The scene keeps the app's depth, stencil and MSAA, but egui's window pass now
has one sample and no depth. Whenever the two targets differ, the scene
(product previews too) is drawn offscreen and blitted, so the UI pass no
longer clears and resolves a multisampled window.

Over 8 s of pointer moves on the 450E rotor head, the app's GPU intervals fell
from 1325 ms to 1006 ms and egui_render from 892 ms to 509 ms; idle stays at
0%. A GPU test checks the split path's picture matches the direct one.

Co-Authored-By: Claude Code <noreply@anthropic.com>

* 3D mouse: a repeated report at rest asks for no frame (#4820)

A device may repeat its latest report while the cap rests. Publishing it
again asked egui for a frame although nothing changed. Only a report that
changes the reading now asks for one; while the cap moves, the view asks for
the next frame itself.

Co-Authored-By: Claude Code <noreply@anthropic.com>

---------

Co-authored-by: Claude Code <noreply@anthropic.com>
2026-10-23 13:42:47 +00:00
raykholo 8b1a722f06 Viewport: the ground shadow no longer sends every frame through the transparency path (#5230) (#5255)
* claim: #5230

* Viewport: draw the ground shadow on the direct path (#5230)

The ground shadow went into the model's translucent triangles, so every
frame with a body counted as see-through and took the order-independent
transparency path. Give it its own batch, drawn alpha-blended after the
opaque geometry as before, so only truly see-through bodies use that path.
2026-10-23 13:42:38 +00:00
raykholo 92d3a2fde5 GPU: a lost device stops drawing instead of crashing (#4172) (#4229)
* claim: #4172

* GPU: a lost device stops drawing instead of crashing (#4172)

After the driver resets the GPU (Windows TDR), the first queue write to meet the
lost device panicked inside wgpu-core (PendingWrites::activate unwrapping
DeviceError::Lost). Nothing watched the device for its loss.

- egui-wgpu: RenderState::device_lost (DeviceLost) is set by the device's lost
  callback. The painter draws nothing on a lost device, and before a frame's
  uploads it submits nothing (DeviceLost::probe), so a loss the driver reports
  is found by that submission instead of the first write.
- Pictures (thumbnails of saves, ui.render) watch the same flag: on a lost
  device the capturer refuses with GpuError::DeviceLost without touching it,
  and the picture comes from the CPU rasterizer. Headless capturers watch
  their own devices.
- The app, seeing the loss, closes keeping unsaved changes for recovery (as a
  forced quit does) and starts again; the new start says what happened. A
  second loss within ten minutes starts on the software renderer. The loss is
  noted in the crash journal.
2026-10-23 02:19:35 +00:00
Jackson OurslandandClaude Opus 5.5 75261cd797 Viewport: order-independent transparency for see-through surfaces (#4734) (#4903)
* Viewport: levels of detail for large meshes (#4724)

A large instanced mesh (20 k triangles or more) gets coarser copies, made
once on a worker (`mesh_lod`): a quarter, a sixteenth and a sixty-fourth of
its triangles, by quadric edge collapse (`geom::simplify`, after Garland and
Heckbert 1997; no fold-overs, no sheets joined, open boundaries kept), each
with its measured deviation from the full mesh. Each frame, as placements
are culled, each is drawn from the coarsest copy whose deviation stays
under 0.75 px where it stands (`placement_scale`, `level_for`). Until the
copies are made the mesh is drawn in full; the scene is built again when
they are. Display only: picking, measuring and exports keep the reference
mesh. Outside the kernel.

25 placements of a 2 M-triangle STL, 3600 x 2190 px, 4x MSAA, M2 Max
(`instanced_scene_gpu_benchmark`): 20.2 -> 1.2 ms a frame with all in
view, 21.9 -> 2.9 ms close up. The copies take 10 s to make, on a worker.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Viewport: see-through and faded parts drawn as instances (#4888)

See-through (an appearance's opacity) and faded (context around an active
component) bodies in one colour are drawn as instances like opaque ones,
their alpha in each placement's colour: one copy of the triangles, levels
of detail and culling, no per-placement copies built on the window thread.
Their placements are ordered back to front each frame (not their
triangles one by one) for devices that sort see-through surfaces.

The GPU benchmark gains `SOLIDCRAFT_BENCH_OPACITY` and the CPU share of
each frame, and its pictures each placement's own triangles too.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Viewport: order-independent transparency for see-through surfaces (#4734)

The model's see-through surfaces (glass, faded context; instanced or not)
were sorted back to front triangle by triangle on the window thread
whenever the view turned. They are now drawn with weighted blended order-
independent transparency (McGuire and Bavoil 2013): when there are any,
the frame is drawn into the renderer's own textures (as supersampling
does), the see-through surfaces gathered into a weighted colour and a
revealage target against the frame's depth, laid over the frame in one
full-screen pass, then the overlays drawn over them. Nothing is sorted.
A device that cannot blend into float targets keeps the sorted way.

25 placements of a 2 M-triangle STL at opacity 0.5, M2 Max, 3600 x 2190 px:
about 10 s a frame sorted, 21.6 ms (17.2 ms instanced, 2.8 ms with levels
of detail).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-22 16:17:39 +00:00
Jackson OurslandandClaude Opus 5.5 f843a3f83c Picking: remove the GPU pick buffer, the per-body BVHs answer alone (#4665) (#4666)
The pick buffer drew the model's surfaces and edge ids for every still
view to narrow which edges a pick tests. Measured against the BVHs alone
(release, M2 Max, 1400 x 900 view, 600 cursor positions), a pick costs
2-51 us through the BVHs and the tile saved 0-4 us of it, while each pass
cost 0.7-12.5 ms of GPU time per still view, plus gathering its scene on
the window thread, read-backs a frame or two late and repaints to wait for
them (and, before, a copy of every triangle, which made it refuse models
past 6 M triangles). Picks are unchanged: the BVHs give the same answers.

`pick_timing_on_curved_models` gains eight layers of bodies seen from above
(the pick buffer's best case) and `SOLIDCRAFT_BENCH_MESH` placements.

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-22 12:03:37 +00:00
Jackson OurslandandClaude Opus 5.5 981d23d1d2 Viewport: weld a large mesh's shared vertices quicker, and ahead of its first frame (#4679) (#4680)
`SharedTris::new` used the standard SipHash hasher and collected each
triangle's corners into a Vec: 451 ms for a 2 M-triangle STL on the window
thread. A quick hasher for the coordinates' bits and a fixed array make it
278 ms, and the background mesher that meshes an opened or inserted
file's bodies now welds their large meshes too (`warm_shared_tris`), so the
first scene that shows them takes the shared vertices ready-made.

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-22 12:03:28 +00:00
Jackson OurslandandClaude Opus 5.5 3d3dd27439 Viewport: leave out placements of an instanced mesh that are out of view (#4646) (#4647)
* claim: #4616

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Viewport: draw a part placed many times from one copy of its triangles (#4616)

A part placed by many occurrences was drawn, framed and picked from a moved
copy of its mesh per placement: on the CPU (the moved mesh and its scene
bytes) and on the GPU (its vertices uploaded once per placement).

- The model scene draws bodies that share a mesh as GPU instances: the
  shared triangles once, a rigid move and colour per placement
  (`InstancedMesh`, `vs_tri_inst`), in the solid, section and AO passes.
  `GpuScene::merged` expands them, so scene tests see world triangles.
- `ModelBody::mesh_in_place`, `plan_mesh_in_place_at`, `mesh_bounds` and
  `display_identity` give the shared mesh, its placement, the exact box and
  an identity without making the moved copy.
- Display refinement's body ids, Fit/Home bounds, the viewport's body boxes,
  the background mesher and picking use them (picking moves the ray into
  the part's frame and shares one pick index).

25 placements of a 2 M-triangle STL on an M2 Max: orbit 45 -> 37 ms a
frame, footprint 41 -> 16 GB, 24 copies 39 -> 18 s.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Viewport: draw large meshes from shared, indexed vertices (#4637)

An instanced mesh's vertices are shared between its triangles and drawn
indexed (`SharedTris`). A flat-faceted mesh (an STL's facets) shares them by
position alone and lights each fragment with its triangle's own plane
(`vs_tri_inst_flat`, `fs_*_flat`), taken from the fragment's position
relative to the eye so it stays exact at any zoom; its vertex outputs are
kept few, since a tiled GPU stores every vertex's outputs. A body alone is
drawn this way from 50 k triangles. A mesh too large for one buffer falls
back to unshared vertices.

`instanced_scene_gpu_benchmark` (manual, ignored): 25 placements of a 2 M-
triangle STL at 3600 x 2190 px, 4x MSAA, M2 Max: 21.3 ms a frame for the
placements' own triangles, 16.1 ms shared and indexed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

* Viewport: leave out placements of an instanced mesh that are out of view (#4646)

Each frame `Batches::cull` keeps the placements whose box may be seen
(`placed_box_in_view`: all eight corners past one side of the view in clip
space, so it holds behind the eye too) and writes them again only when
that set changes. The manual benchmark gains `SOLIDCRAFT_BENCH_CLOSE=1`:
close up on the middle of 25 placements of a 2 M-triangle STL (13 in
view), 3600 x 2190 px, 4x MSAA, M2 Max: 28.2 ms a frame for the
placements' own triangles, 17.2 ms instanced and culled.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-22 04:04:59 +00:00
93a2922c5e Display plans and the model cache for large STEP assemblies (#2814) (#4021)
* STEP import: mesh band faces that wrap around a closed surface correctly

Fixes two related bugs that caused SolidCraft to fail to mesh a face wrapping
all the way around a closed, non-periodic surface (a revolved ridge band, or a
tube-shaped B-spline surface): a missing face, a folded/self-overlapping
triangulation, or a hole in the rendered body.

1. Ring edges that run slightly off their surface. STEP files write a surface
   and the curves bounding it to the file's own declared uncertainty, so a
   ring edge can legitimately sit a few micrometres off the surface it
   bounds. The code finding where a ring edge crosses the planned split line
   compared against the edge's raw (off-surface) point using a hardcoded
   tolerance far stricter than the file's own precision, rejecting a valid
   cut. Now matches against the point on the surface directly below the ring
   crossing, respecting the file's declared tolerance.

2. Closed B-spline surfaces, their slits and their slivers. truck only treats
   the rotation of a revolved surface as periodic; a face on a B-spline
   surface whose control net closes on itself (a tube) wasn't recognized as
   wrapping around at all. This produced a folded face, a hole (a boundary
   loop landing a full period away from where it should cut), and one
   near-zero-width sliver face costing the entire body its tessellation
   plan. Adds periodicity detection for closed B-spline surfaces, period-
   aware hole-loop alignment, dropping of zero-width slit boundaries, and
   isolating a too-thin-to-triangulate face so it no longer takes the rest
   of the body down with it. The GPU shader's periodic wrap mirrors the CPU
   math so the two don't drift apart.

Regression tests: band_faces_split_when_ring_edges_run_off_their_surface,
band_face_with_a_zero_width_slit_is_still_meshed,
a_sliver_face_too_thin_to_triangulate_does_not_cost_the_body_its_plan.
Ablation-tested: disabling either mechanism in (2) reproduces its target
test's original failure message while leaving the other test unaffected.

* STEP import: mesh the band's split faces, not just split them

jyviko asked whether a band wall missing from the interactive app's render
was a shading issue or a genuinely missing face. It was missing: the ring-
edge split fix made the band's topology correct, but plan_face still
produced zero triangles for the two new sub-faces.

Two causes, both in plan_face (crates/kernel/src/tess/mod.rs):

1. The tolerance for how far an imported boundary may lie off its surface
   was tied to the requested mesh's chord tolerance instead of to the
   body's own size. A finer display tolerance could then reject a boundary
   a coarser one had just accepted, so the face would drop out exactly when
   zooming in on it.

2. Near a sharp corner of this flat-panelled band, several boundary points
   that are legitimately a few micrometres apart (within the STEP file's
   own declared precision) all projected onto the same surface point. The
   boundary loop then repeated a point in (u, v), which a triangulator
   can't follow, so the face's plan failed outright. One face without a
   plan costs the whole body its plan, and the fallback mesher can't
   handle a closed B-spline surface either, leaving both halves of the
   split band with no triangles at all: present in the topology, invisible
   in every renderer, and absent from the body's volume.

Fix: scale the off-surface tolerance to the body's own extent (new
Rules::extent), and collapse a run of boundary nodes that land on the same
surface point to one (drop_collapsed) before triangulating.

Regression test additions to band_faces_split_when_ring_edges_run_off_their_surface:
the body's volume now matches the analytic pentagon prism (was a third of
it - only the top cap - before this fix); display_plan leaves no face
without triangles from coarse to fine tolerance; and the actual tolerance
and facet angle the viewport draws with gives the split band its full
lateral area on both halves.

Confirmed on review_band_big.step, the fixture behind the review screenshots:
volume's relative error against the analytic pentagon prism drops from
1.4e-4 to 4.1e-6, and the band wall now occludes the floor grid behind it
in a snapshot the same way the top and bottom caps always did.

Full kernel suite: 739 tests, 0 failed.

* STEP import: hold an imported boundary to the file's own tolerance, not the part's size

jyviko: the off-surface acceptance added in the last commit
(rules.extent * OFF_SURFACE) scaled with the part, so a 2 m part
silently accepted a boundary 2 mm off its surface while a 20 mm copy
of the same part accepted only 0.02 mm. A STEP file's own declared
uncertainty does not grow with the part, so the bound shouldn't either.

Body gains boundary_tol: Option<f64>, Some(file's declared tolerance)
for a body read from STEP (crate::step_in::Units::tol, already used
by the ring-split fix in split.rs), None for anything this kernel
built. It's set before the import's own unmeshed-face check runs, and
carried over by the two face-split operations that re-cut an imported
surface (splitface.rs, outline.rs), since the file's uncertainty still
describes those boundaries afterward.

plan_face's accept now reads: the file's own tolerance when the body
declares one, or the old chord-tolerance/extent bound otherwise. The
old bound's tol * 10.0 term had to go too, not just the extent term:
at a 2 m part's typical display tolerance that term alone is ~20 mm,
nearly ten times the extent bound jyviko objected to, so keeping it
would have left the exact 2 m/2 mm example still silently accepted.

Two new tests in step_in/tests.rs:
- band_faces_split_when_a_scaled_copy_runs_off_its_surface: the same
  band fixture at a realistic 2 m scale with the same absolute slop
  and declared uncertainty as the small one, meshes the same way at
  every chord tolerance from coarse to fine.
- off_surface_budget_follows_the_file_not_the_part_size: the actual
  regression. The same off-surface error is accepted or refused the
  same way whether the part is 20 mm or 2 m; a 2 mm error against a
  1 µm declared uncertainty is refused at both sizes now (accepted at
  2 m under the old bound, refused at 20 mm - that asymmetry was the
  bug). Checked at both a coarse and a fine display tolerance.

Known gap, not fixed here: a STEP-imported body saved into a
.solidcraft design and reopened loses boundary_tol (native.rs's
decoder has no field to read it from) and falls back to the old
extent-based bound on reload. Today's design format has no concept of
this value at all; persisting it is a real format change, out of
scope for this fix.

Full kernel suite: 776 tests, 0 failed. clippy -D warnings clean.

* Display plans for STEP threads and pointed ends (#3847)

Three faces cost a threaded or pointed STEP part its whole display plan, so it
was drawn from the reference mesh: cracked, faceted, and never refined up close.

- A loop through a pole (a cone's apex on the boundary) seemed to wind around
  the surface: the jump at the apex, unwrapped to the nearest copy, was
  arbitrary. The loop is turned to end where it leaves the pole, so the step
  that closes it is an ordinary one (`seam::close_through_pole`).
- Projection onto a spline with hundreds of spans (a helical flank) restarted
  from a 16 x 16 grid and settled on another turn; it now starts from two
  points per span, capped.
- Sag refinement no longer inserts a point exactly on a boundary side, which
  split the side and failed the face.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* tess: a refused face's edges are divided finer and it is planned again (#2814)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* tess: sliver faces plan empty; a face no method plans is filled flat from its nodes (#2814)

A face whose boundary runs out and back along itself (two edges between the
same vertices, a rounding apart, as STEP files write slivers) encloses nothing
and is planned with no triangles. A face still refused after its edges are
divided finer (a cusp, a sliver whose sides sag across each other, a small cap
at a pole) is filled flat from its shared boundary nodes, so the rest of the
body keeps its display plan instead of falling back to the reference mesh.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* display mesh tests: a rim the chord tolerance leaves a line is divided again (#2814)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* portable B-rep: up to 8192 control points along a spline direction (#2814)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* solidcraft-cli: --debug logs what the engine reports to stderr

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* STEP import: a spline written over a tiny parameter range is read over one (#2814)

Some writers parameterize splines over a few millionths (a humanoid robot's
edges span 5.4e-6). The kernel tells parameters apart only to 1e-6, so over
such a range it evaluated the curve's end wrongly, and trimming an edge to its
vertices dropped a control point: the bodies could not be stored exactly and
the whole model cache was left out of the design. The knots are now stretched
to span one; the curve is the same and edges are trimmed by their vertices, so
nothing read from the file refers to the old parameters.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* model cache: a body keeps its file's declared accuracy (#2814)

A body restored from the cache dropped the accuracy the file it was read from
declares, so the planner held its boundaries to the kernel's own bound and a
reopened design drew differently from the import. It is stored with the body's
runtime values, and only when there is one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* tess: a boundary within a hundredth of its body lies on its surface, at every quality (#2814)

#1721 held a body read from a file to the accuracy the file declares. A writer
declares what it aims for, not a bound its edges keep: a 251-body robot
assembly declaring 0.01 µm has boundaries 10 to 30 µm off their surfaces, and
95 of its bodies lost their display plans. The bound is a hundredth of the
body again, what Normal quality always allowed, and now the same at every
quality, so zooming in still never drops a face (#1721's point). A file
declaring more raises it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

* STEP import: keep faces whose edge vertices sit slightly off their curves

Some exporters write an edge's vertex beside the end of its curve, further off
than the accuracy the file declares (a surface export declares
0.01 mm and puts a vertex 0.19 mm off two spline ends). The reader refused
such edges and left the face, and with single-face bodies the body, out.

A vertex may now sit up to the file's tolerance or 1/1000 of the model's size
(the diagonal around its VERTEX_POINTs), whichever is larger, never more than
the existing 1 mm ceiling. Spline and open-conic edges have the end control
point moved onto the vertex, so every such edge meets its vertices exactly.
The file's declared accuracy still governs closed-curve, seam and circle
checks.

* tess, io: review clean-ups (#2814)

- A sliver face is one whose loops are all slit: drop_slits finds it in the
  retry after a failed plan, so the pre-check on every face (folded) is gone.
- The flat fill uses cdt::triangulate (a lenient fill that leaves out sides
  crossing one already kept) and its flood-fill inside test, and the plane
  from geom::Plane.
- Refinement only builds its indexes when a face was refused, takes the
  faces to plan again from edge_faces, and re-samples edges on the pool;
  Rules::finer for the finer rules.
- One PackageWriter::fits for the model cache, the preview and the thumbnail.
- The band split's ring tolerance lives in split.rs; comments say what the
  off-surface bound is now.
- The CLI logger prints Level::as_str().

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LGeCa5jWWdroMWvXxN3LLM

---------

Co-authored-by: David Lackovic <lackovicdavid1234@gmail.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
Co-authored-by: Ray <raykholo@users.noreply.github.com>
2026-10-21 10:37:12 +00:00
Christian Svensson 511faedcfa Fix OpenGL startup with portable AO depth reads (#2888) (#2933)
* claim: #2888

* fix: preserve AO depth in a portable integer texture (#2888)
2026-10-20 11:41:40 -04:00
26f42ae09b Snapshot: draw a see-through appearance see-through, and show a tool body inside its target (#2329) (#2330)
* Snapshot: draw a see-through appearance see-through, and give the CLI --body, --tool and --section

A headless picture dropped a body's opacity: the scene carried only a colour, so a design with
a translucent appearance was drawn opaque by `solidcraft-cli snapshot`, by MCP screenshots and
in a design's thumbnail, although the window has always drawn it as glass. There was then no
way to look at the body a cut would remove inside the material it meets: only a volume to
measure and reason about, which hides exactly the errors a picture makes obvious (a tool that
stops short, or lands in the wrong place).

- `SceneMesh` carries the body's opacity; `view::build_scene` takes it from the design
  (`model_body_opacity`), and a selected body stays solid so the selection colour reads.
- The GPU capture routes a see-through body into the scene's glass at its own alpha, as the
  viewport does; the CPU rasterizer draws the solid bodies first (they own the depth buffer),
  then the see-through ones blended over them, far sides before near sides and without writing
  depth, so what lies inside shows and a line between the two is not hidden.
- `snapshot --body NAME:opacity=0.3,color=#rrggbb`, `--tool NAME` (that body solid in the tool
  colour, the rest see-through) and `--section z=27` run the commands the app itself runs
  (`appearance.assign`, `inspect.section`): the options are shortcuts, not a second renderer.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* Tests: measure the tool colour as it is drawn through a see-through target

The tool-colour filter was derived on paper and demanded a red more than twice either other
channel, which the target in front of it tints away (measured: (200, 108, 85) through opacity
0.3); and the colour check painted the tool, which a solid target hides.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* fmt

* Snapshot options: strip the colour's # rather than slicing it; README

Slicing `value[1..]` on an input-derived string is what the house rules ask us not to write,
even where a length check has already made it safe. The error a failed assignment reports now
names the body rather than claiming it came from --body, since --tool reaches it too.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* --tool: assign the bodies around the tool in groups, not one command at a time

Each appearance.assign is an undo step with a document snapshot behind it, so one per body
made --tool cost grow with the size of the design for no reason: the bodies around the tool
all get the same opacity, and the ones that share a colour can be named in one call. A design
whose bodies all wear the default grey now costs two commands whatever its size.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

* Rasterizer: say where its glass and the GPU's differ

Splitting each see-through body by facing orders its own two layers; it does not sort across
bodies, which the GPU does. Worth stating rather than leaving a reader to find it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
2026-10-19 16:10:45 -04:00
DuBearteandClaude Opus 5.5 a9c7616e7f Lines thinner than a pixel no longer vanish without multisampling (grid lines, #2823) (#2871)
* claim: #2823

* fix(view): lines thinner than a pixel no longer vanish without multisampling (#2823)

A line quad narrower than a pixel covers no pixel centre when it falls between
two rows, so the grid's 0.7 px lines came and went as the view zoomed where the
viewport draws without 4x MSAA (some browsers and devices). Such lines are now
drawn a pixel wide, faded by the width they lack.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-19 16:05:56 -04:00
William Dunn d8d04c9009 Tool modifiers: shared resolve, classic defaults, preview-safe Ctrl (#1616)
* Centralise tool modifiers with classic defaults and preview-safe Ctrl.

Route sketch/selection chords through a shared resolver. Keep Alt snap
suppress and dual toggle as defaults; optional Shift add-only and Ctrl
snap mapping. Skip preview rollback while a manipulator drag is active.

* Address #1616 review: classic snap suppress, single viewport owner, docs.

Default suppress to Ctrl/Cmd or Alt; drop navigation set_suppressed; remove unused label_key; add preview rollback gate test; document sketch modifiers.

* fix(#1616): correct Modifiers primary in tests; simplify rollback fixture.

* Treat Ctrl as primary off macOS, and rustfmt the modifier files.
2026-10-19 16:04:28 -04:00
CypherNaugh_0x 8e762fee57 Add optional ambient occlusion and graphics quality presets (#1135)
Thanks @CypherNaught-0x.
2026-10-18 21:29:03 -04:00
CamWight c5f3515073 Viewport: hidden edges no longer show through faces, and thin or coplanar faces stop z-fighting (#1434, #1739) (#1768)
Thanks @CamWight.
2026-10-16 16:26:19 -04:00
Jackson Oursland 8907da5e86 Viewport picking: a BVH per body mesh and a GPU pick buffer for visible edges (#1776)
Thanks @oursland.
2026-10-16 14:26:38 -04:00
Pablo López 741265f214 Create Sketch shows the origin planes through the model (#1755) (#1757)
Thanks @pablopeza.
2026-10-16 12:26:17 -04:00
Irakli Kourtis 1a0b33fcf3 Headless pictures draw on the GPU, with the viewport's pipelines (CPU rasterizer as last tier) (#1555)
Thanks @jyviko.
2026-10-15 22:19:26 -04:00
bherbruckandClaude Opus 5.5 ec64564f21 Fix the startup crash on D3D12 with FXC: CPU fallback for the tessellation pipeline, FXC-friendly shader (#1098) (#1099)
* GPU tessellation: a pipeline the device cannot build falls back to CPU meshing instead of crashing (#1098)

Build the tessellation compute pipeline inside error scopes; TessGpu::new returns an error,
and the viewport logs it and meshes bodies on the CPU. A Windows CI job builds the pipeline
with D3D12's FXC on the WARP adapter.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* Tessellation shader: basis functions in one-dimensional local arrays, so D3D12's FXC compiles it (#1098)

FXC has to unroll loops that write an array inside a struct (Basis.n0/n1) or a 2-D array
(ndu) by a counter, and cannot with run-time trip counts (X3511). A2.2 basis functions in a
local 1-D array returned whole; derivatives from the degree p-1 functions; the pole step
offsets computed rather than indexed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-14 10:40:24 -04:00
Irakli Kourtisandjyviko 587cf865b8 Display tessellation plans (4/4): one GPU dispatch for the scene, intersection edges by Newton (#806)
Packs the scene's planned bodies into one GPU upload and dispatch, and evaluates intersection edges by seeded Newton refinement from a fitted B-spline through exactly solved leader points, falling back to the previous evaluation near tangency. Intersection nodes carry their faces' parameters. This changes speed, not what the app draws.

Part 4 of 4 of #303.

Co-authored-by: jyviko <2501648+jyviko@users.noreply.github.com>
2026-10-13 22:57:08 -04:00
Irakli Kourtis ff7a347013 Display tessellation plans (3/4): display meshes from plans, GPU evaluation in the viewport (#805)
Display meshes are built from tessellation plans, with GPU evaluation in the viewport.
2026-10-13 16:27:16 -04:00