Port of snaporca 9fa304c77a. Parity OK: 17 files identical, 8 diverging at their
expected counts.
The everyday dimension in SolidWorks and Onshape, and this kernel had no form of
it. Distance constrains the straight-line gap; LockX/LockY pin one point's
ABSOLUTE coordinate. Neither relates two points along an axis.
DistanceX/DistanceY emit SLVS_C_PROJ_PT_DISTANCE against two unit direction
lines built in the solver's G_FIXED group, so they add no degrees of freedom.
THE DIRECTION IS SIGNED, and getting it backwards is silent. libslvs defines a
LINE_SEGMENT's direction as point[0] - point[1] (entity.cpp) and
PROJ_PT_DISTANCE constrains (pB - pA).dot(dir) (constrainteq.cpp:234), so the
reference lines are built head-first to mean +X and +Y.
The same signedness was a real defect in the GUI: the inline editor was
pre-filled with |delta|, so when the closest endpoint pair ran right-to-left,
opening the dimension and accepting the number shown would flip the point to the
other side of its anchor. Opening a dimension and accepting its own value must
be a no-op. The refs are now ordered so the shown value is positive.
On the CAD-1000-hours corpus, dimensioning and constraining is 31.9% of all
observed CAD time -- the largest single class, 7.6x feature operations. This is
the item in the constraint epic that lands most directly on it.
VERIFICATION LIMIT, as with the previous commit: this fork's kernel suite still
cannot run (find_package(assimp) fails at configure time, snaporca-w80c). The
shared sources are byte-identical to snaporca's, where kernel is 2603 assertions
/ 200 cases and ALL LADDERS HELD across all seven rungs.
Port of snaporca 9ec6405e2d. Parity OK: 17 files identical, 8 diverging at their
expected counts (DesignPanel.cpp 32, test_slvs_constraints.cpp 3).
Measured on the CAD-1000-hours corpus: 51.5% of observed CAD time is 2D sketch
work, and dimensioning/constraining alone is 31.9% -- the largest single class.
Two constraints every industrial sketcher has were missing here.
EqualRadius fixes a dead end rather than adding a feature. Picking two circles
and pressing Equal emitted EqualLength, which maps to SLVS_C_EQUAL_LENGTH_LINES
and constrains nothing on a curve: a silent no-op with no error. Equal is now
one button with two meanings, as in Onshape and SolidWorks.
Collinear emits PARALLEL plus PT_LINE_DISTANCE=0 rather than PT_ON_LINE, whose
internal valP param this libslvs port leaves at 0, drifting an already-collinear
pair.
Both types are appended at the END of SketchConstraintType: cereal serializes it
positionally, so inserting elsewhere reinterprets every saved recipe.
VERIFICATION LIMIT, stated rather than implied: this fork's kernel suite could
NOT be run. scripts/CAD/run-kernel-tests.sh fails at CMake configure time on
find_package(assimp), before any source compiles -- a pre-existing deps gap
(snaporca-w80c), not this change. The shared sources are byte-identical to
snaporca's, where the full gate passed: kernel 2588/195 and ALL LADDERS HELD
across all seven rungs.
Also fixes two defects in this fork's scripts/CAD/run-all-checks.sh:
- `cd $(dirname $0)/..` landed in scripts/ instead of the repo root, so every
rung looked for itself under scripts/scripts/. Broken since the script moved
into scripts/CAD/; the three sibling scripts were fixed then and this was
missed, so the gate has not run since.
- C defaulted to snaporca-gui, the OTHER fork's rig container, so this fork's
gate would drive snaporca's app and report green about the wrong binary.
run-kernel-tests.sh:31 documents the identical defect being fixed once
already for the build volume; this is the third instance.
The recipe's 3MF entry moved from Metadata/SnapOrca_cad.bin to
Metadata/orca_cad.bin, so projects saved by earlier builds opened with an
empty Design tab. Both 3MF backends now read either name and write only the
new one; the recipe version advances to 6 to mark the move.
Requested by SoftFever on PR #15238: ten of these had accumulated loose in
scripts/ next to ~20 unrelated upstream ones, with names that only meant
something to whoever wrote them. They now sit in scripts/CAD/, mirroring the
src/libslic3r/CAD/ and src/slic3r/GUI/CAD/ split, and the verb in the name is
the role: build- produces a binary, start- brings something up, run- runs a
suite, check- asserts one thing against a live app.
kernel-test.sh -> CAD/run-kernel-tests.sh
ladder-all.sh -> CAD/run-all-checks.sh
sketch-ladder.py -> CAD/check-sketch-engine.py
ladder-corpus.py -> CAD/check-sketch-engine-corpus.py
gui-ladder.py -> CAD/check-gui-sketching.py
offer-ladder.py -> CAD/check-gui-context-menu.py
mcp-sketch-smoke.py -> CAD/check-mcp-sketch.py
rig-build.sh -> CAD/build-gui.sh
docker-iter-build.sh -> CAD/build-gui-incremental.sh
gui-session.sh -> CAD/start-headless-gui.sh
"Ladder" was the worst of them: it named the shape of the test (rungs of
increasing difficulty) rather than what the test proves, so nothing in the
directory listing told you which one needed a GPU and which was pure kernel.
Every reference rewritten -- the docs, the cross-calls between the scripts,
Dockerfile.deps, and the container-side /OrcaSlicer/scripts paths. The three
shell scripts resolve REPO relative to themselves and now sit one level
deeper, so that walk went from /.. to /../.. . The copies these push into a
container's /tmp were renamed to match, or the container would have kept the
old names alive.
Two runtime paths deliberately NOT renamed. /tmp/orca-rig-build.lock is a
cross-fork contract -- both forks take the same lock so two concurrent builds
serialise instead of OOMing the box, and renaming it on one side silently
removes that guard. /tmp/gui-session.log is a runtime artefact, not a script.
Added scripts/CAD/README.md: what each script proves, what it needs, and the
two constraints that have each cost a session (never build inside the GUI
container; a window manager is required or synthetic keys are ignored).
On CI, which was the other half of the request: the kernel suite is already
there and always has been. The cases are registered in
tests/libslic3r/CMakeLists.txt under if (SLIC3R_CAD), which defaults ON and no
workflow turns off, so they build into libslic3r_tests and run under ctest on
every platform via unit_tests.yml -- like any other unit test, needing no new
job. They have simply never been seen to run, because the workflows on this PR
are still awaiting maintainer approval. run-kernel-tests.sh is the local loop
over the same cases, and it is the only script here CI could run: the other
six need an OpenGL canvas and synthetic input.
Verified: scripts/CAD/run-kernel-tests.sh from its new location, all tests
passed, 2562 assertions in 190 test cases.
* Keep mixed-color filaments intact when the extruder count changes
The extruder-count spinner resized the filament arrays in bulk at the tail,
which is where mixed-color slots live, so a new filament landed behind the
mix and the sidebar skipped a slot number. It now adds and removes one slot
at a time through the same calls the sidebar's +/- buttons use, so a new
slot opens ahead of the mixed tail and a removal renumbers object filament
ids, painted facets, custom g-code and mixed components rather than
clamping them away.
Drops the vector overload of set_num_filaments(), which this leaves without
callers.
* Skip straight-run splits in corner smoothing
Teach `CornerSmoother` to treat vertices that only continue a straight segment as part of the same leg instead of rounding them as corners. The smoother now keeps a three-point window so it can emit a corner only once both adjoining legs are known, which avoids unnecessary corner processing while preserving real turns such as hairpins.
* Add regression test for split-leg smoothing
Adds a FillCornerSmoothing regression test covering polylines with an extra collinear vertex in a straight run. The test ensures corner smoothing treats split and unsplit geometry identically, preventing inconsistent rounding radii in triangular/grid infill paths.
Port of snaporca 1bb9825db0. Four defects from an independent 20-agent audit,
each verified in the code first; two further findings from the same report were
verified OUT and are not in this commit.
remove_feature()/move_feature() remapped Extrude::sketch_ref and a Mate's two
connectors and nothing else, leaving seven of the nine index-bearing fields —
sweep_path_ref, loft_profile_refs[], pattern_curve_sketch, rib_sketch_ref, and
sketch_ref on Revolve, Sweep, Rib and the Surface* family — pointing at whatever
slid into the slot. Quiet by construction: the shifted index still names a real
feature, recompute() succeeds, the solid is built from the wrong profile. The
comment above the loop already required "EVERY field holding a feature index"; the
code under it handled two, because a type switch is only correct on the day it is
written. for_each_feature_ref() visits the FIELDS instead, so a feature type added
later is covered the moment it reuses one. plane_base and axis_plane_a/b are
excluded on purpose and documented at the helper — they encode an ordinal into the
datum-plane list, not an index into features[], and are filed separately. The
delete cascade got the same field-based treatment.
The regression test was run against the pre-fix code to prove it bites: all three
sections fail there, and move_feature returns TRUE while leaving sketch_ref == 1
where it must be 0 — success with the wrong answer, which is what makes this class
expensive.
apply_constraint, commit_entity_constraints and delete_constraint mutated the
recipe with no checkpoint() and no sync_recipe_to_model(), alone among seventeen
mutation sites in that file: Ctrl+Z reached past the constraint edit and discarded
unrelated work, and saving persisted the pre-constraint blob. A rejected constraint
now calls abandon_checkpoint() rather than leaving an undo step that does nothing.
MCP: params["generation"].get<uint64_t>() sat outside the try inside a bare
CallAfter lambda, so one malformed string terminated the process through the wx
event loop; it is type-checked now and the lambda lets nothing escape. The socket
bound with no mode of its own in a world-writable directory — umask around bind()
plus chmod, and it refuses to listen rather than listen wide. The reply write is no
longer a bare write(), which could SIGPIPE the app when a client hung up.
Kernel suite on this fork: 190 cases / 2562 assertions, green. GUI target compiles.
The full ladder gate ran on snaporca (ALL LADDERS HELD — gestures 98/98, offer
108/108, corpus and corpus-scale green) and fork-check parity holds at 17 identical
/ 8 diverging as expected, which is what makes that gate transferable here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
User report 2026-08-23, and it is right: "you have renamed the feature extrusion,
not the body. i clicked rename on the body feature tree and the feature extrude
changed name. this means that you consider the extrusion = the body. this is very
far from truth as a body can contain several extrusions."
That is exactly what the previous commit did. It resolved a selected body to
CadBody::source_feature and renamed THAT feature, on the reasoning that a body has
no name of its own. The reasoning described an implementation detail — CadBody::
name is derived and restamped on every recompute — and mistook it for the user's
model. An Extrude, a Cut and a Fillet all land on the same body: the maker is one
operation in its history, and renaming it renames the wrong object.
A body now has a name of its own. CadBody::user_name, set only by a rename, is:
- carried across recompute() by body index, next to the per-body colour override
and under the same index contract the GUI already relies on for visibility and
Move — without which a name would survive exactly until the next feature;
- written into the recipe, because bodies are recomputed and never serialised, so
a name has nowhere else to live and would otherwise vanish on reopen;
- shown on the Bodies row ahead of the derived maker name, as "Body N — name",
with the number still leading because every status line, the interference
report and the mate errors identify a body that way.
The recipe block is APPENDED after the variables block rather than given a version
bump. A build that predates it reads features and variables, returns, and never
looks at the trailing bytes — so yesterday's projects open here and today's
projects still open there. A bump would have cost every project written today its
readability by the previous build, for one optional field.
Renaming a FEATURE is unchanged. The feature tree renames features; the Bodies
list renames bodies; neither reaches into the other.
WHAT THIS COST, and why it is written down. Getting here took two wrong turns
inside one fix, both mine:
1. UnselectAll() -> Unselect(). Both trees are wxTR_SINGLE, where UnselectAll()
— the MULTI-selection call — does nothing. That was the one-word reason the
rename had been vetoed everywhere (BEGIN_LABEL_EDIT refuses while
tree_body_selection() >= 0). Fixing it turned a pair of harmless no-ops into
a real loop: the two lists clear each other so "the target" is unambiguous,
so clicking a body row ran apply_body_row -> m_tree->Unselect() -> the feature
tree's SEL_CHANGED -> m_parts->Unselect(), which cleared the row just clicked.
The handler now clears the other list only when it actually holds a selection.
2. Trusting a screenshot taken after a polluted run. A leftover Rib card had
shifted the whole panel, so a click measured against it landed nowhere near
the row. Relaunch, then measure.
VERIFIED, on the rig and in the kernel:
body renamed ('Extrude', user_name=False) -> ('Bracket', user_name=True)
features untouched ['Sketch', 'Extrude'] before and after
survives a recompute add a Hole to the same body: features become
['Sketch', 'Extrude', 'Hole'], body stays 'Bracket'
survives the recipe serialize -> deserialize -> 'Bracket'
New kernel test "a body carries its own name, through recompute and the recipe"
pins all three properties; the suite is 189 cases / 2547 assertions.
Gate green: ALL LADDERS HELD — offer table matches the atlas, kernel suite,
engine rungs 1-8, 977-sheet corpus + the heaviest sheets, gesture ladder 98/98,
offer ladder 108/108.
The sidebar Mixed Filament list, the extruder icons, the color painting
gizmo and the canvas filament bar now show the same bottom-to-top fade the
Edit Mixed Filament preview shows, custom gradient curves included, instead
of a horizontal fade between the two component colours. Ordinary and vendor
multi-colour filaments are drawn exactly as before.
Two commits carried across (snaporca 579a9a9162, f68613cfc5).
Past about 480 entities a sketch had NO constraints at all and said nothing: libslvs
declares MAX_UNKNOWNS = 1024 and is handed every entity in the sketch at two params
per point, so the whole system came back TOO_MANY_UNKNOWNS and try_add_constraints
rolled the entire inferred batch back. From there no dimension could ever be applied.
Constraints only couple entities that share a point, so the solver now falls back —
only on TOO_MANY_UNKNOWNS — to solving connected components separately and committing
all-or-nothing. The auto-constraint pass batches its Horizontal/Vertical constraints
instead of one solve each, which is what kept the bulk path fast once solves started
succeeding: a 1204-entity load went 1585 ms -> 562 ms.
Plus the scale rungs (a thousand-entity plate drawn on by hand; the heaviest real
drawings graded and timed), the --step 1 fix that used to select nothing while
reporting a clean run, and scripts/ladder-all.sh as the one-command gate.
Parity 17 identical / 8 diverging as expected. Kernel suite here: 188 cases /
2532 assertions, including "a sketch past the solver's unknown limit still solves".
snaporca-yww4, snaporca-x6v7, snaporca-j6sr
and a ladder that draws with the mouse
Three commits carried across (snaporca 4ffd60eacb, 421055c2ec, b71216ce0b):
1. A design made only of sketches must survive being saved. CadDocument::recompute
returned false with "no solid-producing features" for a document that has no
solid, and two callers read that as "unusable": the GUI syncs the 3MF recipe
only after a successful recompute, so a sketch-only design was saved with no
recipe at all, and deserialize_recipe ends with `return recompute()`, so even a
project that carried one was refused on load. Having nothing to build is now a
success; a feature that MEANT to build a solid and produced none still fails.
DesignPanel::refresh_tree syncs the recipe too, for the paths that call
m_doc.recompute() directly.
2. Scripted geometry arrives exact. The Horizontal/Vertical inference window and
the endpoint weld window both close to zero for add_entities_scripted; void
attribution probes from a point strictly inside each loop instead of from its
first vertex. Corpus rung 39 graded / 39 fully clean, was 35 with 6 failures.
3. scripts/gui-ladder.py — 17 rungs, 84 properties, all driven by synthetic clicks
and typed values rather than through the socket.
Parity 17 identical / 8 diverging as expected. Kernel suite here: 188 cases /
2532 assertions.
snaporca-mtav, snaporca-8xg1, snaporca-5hvl, snaporca-730j
Carries snaporca 0231bd5b68. Parity holds: 17 files identical, 8 diverging as expected.
A reflection reverses orientation, so mirror_entities now hands the reflected half back reversed
in ORDER and flipped per ENTITY — an arc swapping its angles as well as its ends, a spline
reversing its control points. Appending it to the source then yields one walkable chain instead
of two halves meeting head-to-head, and a mirrored CCW loop stays CCW.
This is the producer half of the confusion that cost three defects; the consumers (offset, and
the exact loop area) keep their defensive handling, because that is what makes them correct for
hand-built and imported sketches rather than only for geometry this function produced.
Contract change, carried with the reason: a mirrored line's p0 is the reflection of the SOURCE's
p1, and a mirrored CCW arc keeps a POSITIVE sweep — the reflection negates it, walking it the
other way negates it again. Both [SketchEdit] cases updated, and a new [SketchProfile] case
"a mirrored half continues the original chain" pins the property directly.
Kernel here: all tests passed, 2687 assertions in 232 test cases. GUI target builds and links.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Carries snaporca 572f794c84, d0f9a0052a, 9f7e4e3627 and 3974f8a170. Parity holds: 17 files
identical, 8 diverging by their expected counts.
EXACT AREA. A loop's area is now integrated entity by entity in traversal order — Green's
theorem — instead of being shoelaced over the render polyline, which faceted every arc into 24
chords and lost 2.02 mm2 on a 3706.86 mm2 stadium. 0.054%, invisible on screen, and wrong in a
number reported as "the area".
OFFSET FOLLOWS THE TRAVERSAL. Offsetting a mirrored profile put one half on the wrong side and
split the loop in two, because the chainer only followed p1->p0 links and each entity's offset
side was taken from its stored direction. Chains are now orientation-aware, seeded at a free end,
offset by `reversed ? -d : d`, and normalised head-to-tail on the way out — so offset is correct
for any input ordering and its own output cannot reintroduce the problem.
Both are the same underlying lesson, which has now cost three separate defects: an entity's
STORED direction is not its direction of TRAVEL around the loop.
THE LADDER. scripts/sketch-ladder.py is a graded suite of 2D sketches judged the way a person
judges them — VERTEX, LENGTH, ARC, TANGENT, SYMMETRY, CLOSED — with area only as a cross-check,
because area is derived and nobody can confirm it by eye. Eight rungs from a rectangle up to
MPD5 from the StudyCadCam corpus, a dia 27 x 95 pin reproduced as its revolve half-profile with
the R5 fillet tangency solved exactly. Entirely 2D: no extrude or any solid feature.
Kernel here: all tests passed, 2681 assertions in 231 test cases, including the new
"profile: a mirrored half offsets as one loop, not two". GUI target builds and links.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Carries snaporca 971320e129, 6b049f0dc6, 4aae782029, 444d59f212, 74cf3d7e54 and the build
guards from 597557a6e4. Parity re-verified after every hunk: 17 files identical, 8 diverging by
their expected counts — DesignPanel.cpp still 32, DesignCanvas.cpp still 16, which is the proof
each hunk landed on the right side rather than being copied over a real divergence.
OFFSET OFFSETS THE CHAIN. Per-entity offsetting returned a closed rectangle as four parallel
segments that no longer touch, so entities_to_wires gave four OPEN wires and nothing could be
extruded. offset_entities now chains by shared endpoints and repairs each seam by mitering the
neighbours to their intersection. Second bug, invisible to any single-entity test: +d meant
"left of travel" for a line but "radius + d" for an arc regardless of sweep, so a slot outline
offset with its straights going one way and its caps the other. The convention is now written on
the declaration and pinned by a test.
tests/libslic3r/test_sketchprofile.cpp is new and asserts the LOOP rather than coordinates —
the property that decides whether a profile can be built, and the one the existing single-entity
[SketchEdit] cases cannot see. Its include is catch2/catch_all.hpp here: this fork ships Catch2
v3 while snaporca is on v2, which is why the test files are a tolerated divergence.
RIGHT-CLICK PICKS WHAT YOU POINTED AT, so a line's own verbs are offered instead of the
empty-selection vocabulary; sk_delete stops sharing btn:delete with the feature tree; and an
element's defining number (length / radius / diameter / angle / distance) can be typed, from the
menu or from V.
TWELVE MCP SKETCH VERBS. The socket had ~40 verbs and none touched a sketch, so the 2D layer
could only be exercised by driving a GUI with synthetic clicks. sketch_describe reports each
closed loop, the loops it encloses as voids, exact areas, and where a chain is still open;
sketch_validate/sketch_heal are FreeCAD's ValidateSketch — find vertices that overlap within a
tolerance but carry no coincidence, then weld them AND record the constraint, so a loop closed
by floating-point luck becomes one closed by construction. scripts/mcp-sketch-smoke.py is the
loop that asserts all of it.
Kernel suite on this fork: all tests passed, 2677 assertions in 230 test cases. The GUI target
links against the rebuilt deps image (the wxInspector blockage is gone) and the binary carries
the new verbs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* Add caching system for presets
* Removing user\bundle serialization and keeping it only for system presets
* Integrate caching into WebGuideDialog which speeds up time of SetupWizzard and PrinterSelection dialog
* Add CI\CD step to prepare cache file in ahead of time so user does not need to wait
* Add partial cache generation when only one of the vendros is changed to speed up recalculation time
* Handle corrupted files
* Add cache to GuideDialog as previos version didn't work as expected
* Add inspecting tool and fix CI cache generation
* Generate cache per vendor
* Simplify code by mergin it in PresetBundle
* Simplify code a bit more
* Add cereal serialize() to VendorProfile, PrinterModel, Preset, and Semver
* Remove CachedPrinterModel/VendorProfile/Preset mirror structs from VendorCache
* Fix use-after-free in CallAfter lambda; replace raw thread pointer with unique_ptr
* Use get_vendor_cache_key() to match cache keys written by the app
* Remove BOM added by VSC
* Skip invalid vendors
* Remove leftover cache file
* Fix build for windows arm64
* Revert json cache back
* Update check for stale cache
* Serealize all value fields for Preset class to minimize regression later
* Minimize field duplication by moving Cache thing into PresetBundle
* Add tests for Cache system
* Add a bit more tests
* Merge branch 'main' into feature/cache_profiles_and_optimize_loading_speed
* Rvert from per-verndor to single cache file
Replace N per-vendor .cache files with a single system_presets.cache
that holds all vendors and presets in one serialized blob.
Cache load is now all-or-nothing: on hit all vendors are applied from
the bundle (sub-second); on miss all vendors are parsed from JSON and
a fresh bundle is written to the user cache dir.
Invalidation is driven by bundle_key - a sorted concatenation of all
vendor JSON version strings. Any vendor update invalidates the whole
cache and triggers re-parse on next launch.
Guide wizard (WebGuideDialog) loads the bundled cache into a plain
PresetBundle instead of a separate VendorGuideData struct, removing
the duplicate data model.
generate_system_cache simplified from a per-vendor loop to a single
save_system_presets_cache() call producing one output file.
* Transfer all Preset fields from cache via move assignmet
apply_vendor_preset_group was copying fields manually and missed
bundle_id, user_id, base_id, sync_info, updated_time, key_values,
ini_str. Replace field-by-field copy with move assignment of the
fully-deserialized Preset, then restore the vendor pointer which
is excluded from serialization.
* Ignore cache for future
* Remove not used files
* Ship one preset cache per vendor in place of the profile JSONs
Each vendor's system presets serialize into a single <vendor>.opc built at
package time, and a shipped build carries that file alone — the profile JSON
and its sub-file tree are pruned. The vendor loader, the setup wizard's profile
list and the resource installer all read a vendor through its cache, falling
back to parsing whenever one is absent, stale or unreadable, so the cache stays
an optimization and never a source of truth. Caches hold presets in source form
and resolve inheritance at load, through the same code the JSON path uses.
* Make the preset cache self-describing and load each vendor from the system folder alone
The cached DynamicPrintConfig is keyed by name, through a per-file dictionary of the
distinct opt_keys, the type each was written as, and the distinct enum value names,
instead of by serialization_key_ordinal — a position assigned by declaration order at
static init, where inserting one option shifts every later ordinal and the lookup then
succeeds on the wrong option. Because a name-keyed payload drops the options this build
cannot place rather than being rejected wholesale, the schema fingerprint goes, and with
it the two fallbacks that existed only because an installed cache died on every app
upgrade: the second lookup tier into resources/profiles and the parse fallback to the
same place. A vendor is loaded from <data_dir>/system/ and nowhere else, as on main —
which is what makes the app write its .opc files there again.
* Simplify the preset cache internals after review
* Use the shared temp-dir helper in the preset bundle loading test
* Bound stamp string reads in the preset cache
* Speed up the setup wizard with a profile-data cache
The wizard's per-vendor fast path threw on vendors present only in
resources, falling back to a ~29 s raw JSON scan on every open. Each
vendor now loads from the directory it was found in, and the derived
model/machine/filament/process catalog is cached whole in
<data_dir>/cache/wizard_profile_data.json, stamped by each vendor's
name and version - a fresh cache makes an open one file read, with no
bundle built and no presets installed (~0.2 s vs ~2 s).
* Remove debug SVG dump from a geometry test
* Move the per-vendor cache file format into PresetCacheFormat
* Move the vendor install helpers from PresetBundle into Utils
* rename
* fix flatpak
* change cache version to 1
---------
Co-authored-by: SoftFever <softfeverever@gmail.com>
* Normalize the junction direction vector over XYZE
calc_vmax_junction_deviation() treats the dot product of two jd_unit_vec as a
cosine, but the vectors were scaled by 1 / block.distance, which is the XYZ
length. On an extruding move the E component then pushes the 4D norm above 1 and
the dot product below -1, so the corner reads as straighter than it is and is
planned too fast -- the more so the higher the flow. Measured on a 6 degree
corner at scv 5: 86.9mm/s with no extrusion, 94.4mm/s at 0.029mm/mm, 150.0mm/s
at 0.1mm/mm.
Neither firmware does that. Marlin normalizes over XYZE for any extruding move
(planner.cpp: `if (... || esteps > 0) normalize_junction_vector(unit_vec)`) and
Klipper leaves E out of the cosine entirely, dotting only axes_r[0..2]
(toolhead.py::Move.calc_junction). Normalizing satisfies both: with E normalized
in, the cosine differs from the XYZ-only one by ~1e-5 at printing flow rates.
This is a deliberate divergence from PrusaSlicer, which still scales by
1 / distance -- it carries an older Marlin's behaviour.
Travel moves are unaffected, their vector was already unit length.
Reported by Copilot in review of #15304.
* Test that extrusion rate does not change corner planning
The junction deviation tests were all travel-only, which is exactly why the E
component of the junction vector went unchecked. Cover it: the same corner has
to be planned the same whether nothing, an ordinary 0.42 x 0.2 line, or a fat
large-nozzle line is extruded through it, on both Klipper and Marlin 2.
Reported by Copilot in review of #15304.
Review request on PR #15238: "Place CAD-related files (e.g. CadDocument/
GeometryEngine) into a separate folder."
src/libslic3r/CAD/ the kernel — CadDocument, GeometryEngine, the four
Sketch* units, SketchSolver, ThreadStandards
src/slic3r/GUI/CAD/ the tab — DesignPanel, DesignCanvas, DesignSketchTool,
SketchInlineEditor, McpControl, generated DesignOffer
Pure relocation: no line of logic changes. Two include rewrites follow from it —
files that moved re-spell their own neighbours against src/ (already on the
include path), and files that did not move pick up the new folder. docs and
docs/ux/mockups/gen_offer_table.py follow the same paths.
Verified: libslic3r, libslic3r_gui and libslic3r_tests all build, CAD suite green
at 2518 assertions in 194 test cases, and the sibling fork builds identically —
17 shared sources still byte-identical, 8 diverging by their expected counts.
* Plan corners with junction deviation where the firmware uses it
The time estimator only ever had the classic per-axis jerk model, which limits a
corner by the largest single-axis component of the velocity change. That is
anisotropic: the same corner is allowed sqrt(2) more speed on a diagonal than on
an axis, which paints a four-lobed ripple around every circular wall in the
actual speed and actual flow views, worst on small parts whose walls are made of
short segments.
Klipper has no classic jerk at all and Marlin 2 has none while M205 J is in use;
both plan corners with junction deviation, which sees only the corner angle. Add
that model and use it for those machines:
- Klipper: derived from the square corner velocity, as the firmware does
(jd = scv^2 * (sqrt(2) - 1) / max_accel), reading the scv from
machine_max_jerk_x, where process_SET_VELOCITY_LIMIT() already stores
SQUARE_CORNER_VELOCITY.
- Marlin 2: machine_max_junction_deviation, which was already loaded into the
machine limits but never reached the planner.
- Every other flavor keeps the classic jerk path unchanged.
The model has no per-axis jerk floor, so this also drops the hard slow spot the
estimator drew at the start of every loop from machine_max_jerk_e.
Toolpaths are unaffected: on a full export the only lines that change are M73.
The junction deviation maths, including Marlin's JD_HANDLE_SMALL_SEGMENTS arc
approximation, is ported from PrusaSlicer's src/libslic3r/GCode/GCodeProcessor.cpp.
The Klipper mapping is not in PrusaSlicer, which ignores SET_VELOCITY_LIMIT.
* Add tests for junction deviation corner planning
Cover the three properties the change rests on:
- a right angle on Klipper is planned at exactly the square corner velocity,
the identity that makes the scv to junction deviation mapping correct, and a
shallow corner is planned far faster than per-axis jerk allows;
- junction deviation gives the same speed whatever the corner's orientation,
while classic jerk keeps its sqrt(2) spread, which is the four-lobed ripple;
- machines that do not plan with junction deviation are provably untouched,
including a Marlin 2 printer that has it disabled.
snaporca-wm4s. Thickening the 4-walled open box (60x60 in plan, 40 tall, no caps)
by 5 produced volume 29648.15 where the geometry requires (60^2-50^2)*40 = 44000
— about 67% of it. The corner material at the four vertical edges was simply
absent.
CAUSE. MakeThickSolidBySimple offsets each face along its own normal and sews;
it never extends neighbours to meet, so wherever two faces join at an angle the
corner is empty. A flat sheet has no such join and was always exact (18000.000),
which is why the defect looked like a measurement artefact.
WHY THE TWO EARLIER ATTEMPTS COULD NOT HAVE WORKED. Both switched to ByJoin —
plain, then with Intersection/GeomAbs_Intersection — and both returned a shell,
not a solid, so the body lost its volume entirely and both were reverted. That is
not a parameter problem: in OCCT, BRepOffset_MakeOffset::MakeThickSolid builds a
solid only inside `if (!myFaces.IsEmpty())` (BRepOffset_MakeOffset.cxx:1115).
Handed an open sheet with no closing faces, it stops after the offset shell and
returns it, reporting IsDone() with a non-null shape containing no TopAbs_SOLID.
ByJoin hollows a CLOSED solid by removing faces; an open sheet is outside its
contract.
FIX. Close the sheet, then use the call that mitres: cap the free rims
(ShapeAnalysis_FreeBounds -> MakeFace), sew shell+caps into a closed shell, make
a solid, and hollow it inward passing the caps as the faces to remove — the caps
come back off and leave the wall. Two details, each found by measurement rather
than reasoning:
* A shell sewn from an extruded sheet carries no guarantee of outward
orientation, and MakeSolid does not fix it. Inside-out, the inward offset goes
OUTWARD: measured bbox 70x70x40 and volume 339141.59, larger than its own
bounding box because the result overlaps itself. A negative GProp mass is
exactly that inversion, so it is the test; Reverse() on it.
* A SINGLE face has no neighbour to mitre and must keep the BySimple path. It
does have a free boundary, so "has free wires" is the wrong question — capping
a lone face with its own rim sews a zero-thickness shell and measures 6000
against 18000.
Also: IsDone() is not a success test here, since both failed attempts had it
true. The code now explores for TopAbs_SOLID and refuses a shell.
Tests: new case asserts 44000 with the wall's bbox at 60x60x40 (catching the
inverted-orientation shape, which has the right volume nowhere near the right
place), plus the flat-sheet control at 18000 that must not regress. Full kernel
suite green: 2502 assertions in 187 test cases.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The mate palette called all five kinds viable on connectors created as
Point(world), which belong to no body. Clicking one built the Mate feature and
the RECOMPUTE then failed with "mate: mate_cs_b has no associated body" — the
refusal arrived one step too late, after the feature was already in the tree,
and the user is told about it by an error line rather than by the palette that
offered the thing.
mate_options now checks the same two conditions the apply path throws on: B must
have a body (B is the connector whose body moves; A is the fixed reference and
needs none), and that body must still resolve. Either way all five kinds go
non-viable with a reason, so the offer and the kernel cannot disagree.
Verified on the rig: with two Point(world) connectors, every row is now dimmed
and reads "connector B is not attached to a body — a mate moves B's body". That
also exercises the dimmed-with-a-reason presentation for the first time, which
until now had nothing to show because every kind was always viable.
Tests: a new [mate] case covering no body, a body that no longer resolves, and
the revival once B is given one. One existing case needed its setup widened
rather than its assertion weakened: "an unrecorded fingerprint does not make a
type non-viable" built two body-less connectors, so it was asserting a side
effect of the old permissiveness instead of the property it is named for. Its
connectors now have a body, leaving the missing fingerprint as the only variable.
Full [CadDocument] suite: 2458 assertions in 185 cases.
A rectangle with a circle inside it, drawn in ONE sketch, could not be extruded
to a plate with a bore from the GUI. Five defects were in the way. Each was
found by driving the app on a headless rig and measuring the result — the code
reads correctly at every one of these points, which is why they survived.
1. Wire orientation (kernel). SketchEngine::wires_to_face added every hole as
wires[i].Reversed(), which is only right when the sketch happens to wind both
loops the same way. A circle drawn clockwise inside a counter-clockwise
rectangle came out matching the outer boundary, OCCT swept it as a SECOND
contour, and the prism was the plate with its bore filled and the disc's
volume counted twice. Measured: bbox 67.17 x 219.67 x 10 with volume
152088 mm3 against a solid box of 147542 — a body larger than its own
bounding box, which is the signature. Holes are now added as-is and
ShapeFix_Face::FixOrientation() classifies them; that is winding-independent
and is the idiom make_extrude_regions already used for imported glyphs, which
is why holed TEXT always extruded correctly while a holed SKETCH never did.
After the fix: 142996 mm3, implied bore radius 12.03 mm against the circle
drawn.
2. The live-sketch click threw the picked region away. region_at() served only
as a yes/no gate and on_face_selected() carried no argument, so Extrude fell
back to whichever loop the resolver found first — clicking the material of a
plate-with-a-hole extruded the disc. The region is now carried through, and
DesignPanel also hands it to the tool with set_loop_pick(), AFTER open_tool()
because that re-derives selection state, since extrude_uses_loop() reads
selected_loop_entities() and that lives on the tool.
3. region_at() had no hole awareness and no innermost preference: it returned
the first polygon containing the point. It now skips a region when the point
lies inside one of that region's holes, and picks the smallest containing
loop, so a click in the bore selects the disc and a click on the material
selects the plate.
4. Startup segfault. DesignCanvas::request_repaint probed the GL backend via
OpenGLManager::get_gl_info().get_renderer() before the canvas had initialised
GL — glGetString with no context current and, before init_opengl(), no loaded
function pointers. Anything that asked for a repaint while the panel was
still being built landed there, with no window and nothing in the log. It now
bails at the top on !is_initialized() and asks for a Refresh instead. Note
the crash was in the PROBE, not in render(), which already guards itself.
5. A holed sketch on a plane whose normal points -Z came out as the full box PLUS
a disc — 220274 mm3 where 163726 was due (192000 + 28274). wires_to_face took
a SketchPlane parameter it never used and let OCCT infer a surface from the
outer wire; when the inferred normal disagreed with the sketch's, the hole
classification produced no hole. Every face is now built on the sketch's own
gp_Pln.
Also in this change, from the same rig session:
- A right-click that only clears the sketch selection no longer reports itself
as consumed, so it stops suppressing the offer menu. With any geometry in a
live sketch there was no menu route left to add a second entity.
- Escape no longer discards a live sketch that holds drawn geometry; it says so
and keeps the work (live_sketch_has_work()).
- The holed-region fill is an even-odd scanline instead of a keyhole bridge, so
no corridor triangle leaks from the bore to the nearest corner.
- Cyan is reserved for the selection: an unselected region no longer wears a
shade one step off the selected one.
- The origin planes follow the mode, so pressing Sketch on a document that
already has a body offers them again instead of naming a plane you cannot see.
Tests: three [holes] cases over add_extrude_entities asserting the plate-with-bore
volume, solid and face counts on both a +Z and a -Z sketch plane, and the
by-name refusal of two disjoint regions. Full [CadDocument] suite green.
He merged upstream main into the PR branch himself on 2026-08-13. Taking it into
the local branch rather than force-pushing over it: the fork copy is what PR
#15238 shows, and discarding a maintainer's merge to make my own push
fast-forward would be both rude and a loss of 130 upstream commits.
Brings the branch far closer to main than the 2026-07-24 merge-base the PR body
describes, which is most of what snaporca-36u9 was filed for.
Every version bump so far has permanently orphaned every project saved before
it. deserialize_recipe refused anything that was not exactly the current
version, and with no migration path v2 and v3 projects are unopenable today —
the 3MF still carries the mesh, so the user gets a frozen solid and no feature
history, which is the whole point of the subsystem silently absent.
The cause was the shape of the data, not the gate. save/load is one flat
symmetric list of ~90 fields with no framing, so a reader has no way to know
where a feature ends unless it agrees on every field.
Each feature is now written as its own cereal stream behind a length prefix, and
the same few lines handle both directions of mismatch. Older file, newer build:
the sub-stream ends early, the read throws, and the fields already assigned are
kept while the rest default — cereal assigns sequentially, so a mid-list throw
leaves the earlier fields set, and that is what makes this work. Newer file,
older build: the sub-stream holds more bytes than the reader knows; it reads what
it knows and stops, and the outer stream is untouched because the length prefix
was consumed in full. A field a project predates is not a corrupt project, so
neither case is an error.
v4 keeps its own pre-framing flat path and opens exactly as before —
cad_recipe_v4.bin is untouched and now serves as the witness for that. v2 and v3
stay refused, by name: their field lists no longer exist in this code. This fixes
the future, not the past, and the comment says so rather than implying otherwise.
From here a new field only needs appending to save/load — no bump, no orphaned
projects. That removes the cost that had blocked snaporca-44m and snaporca-dgv.
The helix round-trip test was reading the blob back flat, reaching into the
format instead of through it; framing necessarily breaks that, so it now goes
through deserialize_recipe, which is a stronger assertion than it made before.
Every field check it carried is unchanged.
Tests: four new [CadDocument][recipe] cases, including the one the change exists
for — a deliberately truncated feature blob must LOAD, keeping what it could read.
Suite 177 -> 181 cases, 2366 -> 2411 assertions.
snaporca-2txy.
Tommaso Bianchi