The bodies now always draw over the preview on every face a feature leaves unchanged, for
all features. Since a hole's cut sits inside the body, the Hole preview now hides the
bodies and shows the result alone, as Fillet/Chamfer and Draft do.
# Description
Follow-up to #15238. A review of the Design tab, and then testing its
Linux AppImage on desktops, turned up UX inconsistencies, kernel bugs
and a crash. This PR fixes them and fills the gaps found along the way.
Everything stays behind `SLIC3R_CAD` and the `enable_cad_feature`
preference. With the preference off, Prepare and Preview behave and
render exactly as before.
The first four commits are Design-tab fixes that landed on the fork
after #15238 and were never sent upstream: the value-field label, the
`GUI.hpp` include, sketch tool batches 9–14 and honest tool messages.
The later commits build on them.
**Interaction: one rule for mouse, Enter and Esc** (charter
`docs/CAD/cad_ux_guidelines.md` §4.2)
- Enter does what ✓ does; Esc does what ✗ does. Esc steps back one level
(value field → gesture → tool → selection) and never deletes anything.
- A click on empty space clears the selection and never applies a
pending operation.
- A right-click with nothing pending opens the offer. The click is
recognised by drift alone; the 200 ms timing rule is gone, since §6.2
forbids timing-dependent gestures.
- The panel's CHAR_HOOK owns Delete, Esc, Ctrl+Z/Y and F, so the
duplicate handlers in `GLCanvas3D` are removed.
- Ctrl+Z/Y work inside a sketch, and Edit ▸ Undo/Redo drive the Design
history while that tab is shown.
- A value the geometry cannot take (zero length, sweep > 360°, …) is
refused inside the field, and the field says why.
**Kernel**
- Solver: tangency and point-on-line pick the correct side;
circle–circle tangency is handled; the partitioned solve maps sentinel
references; constraints that could not be applied are reported instead
of silently dropped.
- Trim, extend, offset and mirror are fixed for arcs and ellipses.
Negative scale is handled, and zero-radius circles no longer produce
geometry.
- Feature → body references resolve by identity, so hiding, reordering
or deleting a feature no longer re-targets later features.
- Hole standards table corrected: inch countersinks are 82°. New threads
use the nominal diameter; new patterns use inclusive spacing. Both are
gated by flags, so existing recipes rebuild unchanged.
- **A closed loop that crosses or folds back is not a region.** Such a
loop passed the closed-loop check and MakeFace, and extruded into an
invalid solid with no caps. `SketchEngine::wires_to_face` now checks the
face (BRepCheck) and fails with a reason; `sketch_loop_defect()` finds
the crossing or cusp exactly, and the sketch tints the loop red and
marks the point.
- Recipe: saves origin, body colours and the loop auto-close setting in
a trailing block; the v4 reader is frozen as `load_flat_v4`; the 3MF
reader caps the entry at 1 GiB and the legacy entry name never overrides
the current one. New fields (`dressup_edges`, the Text parameters,
`revolve_axis_entity`) are appended at the end of the framed recipe, so
existing projects load and rebuild unchanged.
- OCCT failures are caught before `std::exception` in the new code (OCCT
≥ 8).
**Modelling**
- **Several solid edges in one Fillet/Chamfer.** Shift/Ctrl+click adds
or removes edges of the picked body; one feature dresses them all at one
size, every id resolved against the same body. The card and the offer
name the count.
- **Sketch Offset takes the whole outline** the picked curve belongs to,
with a live preview, instead of one segment (on a text outline a segment
is a fraction of a millimetre, so nothing seemed to happen).
- **Revolve about a line of the sketch.** The axis was only the sketch
plane's X or Y axis. It can now be any line of the profile sketch: a
construction centerline (preselected when the sketch has exactly one) or
an edge of the profile. The gizmo draws the axis dashed. A profile on
both sides of the axis is refused with the reason (MakeRevol failed
there with none). Surface Revolve takes the same axes.
**Panel, offer, text**
- The offer menu: its strings are translatable (the generator emits
`L()` markers, and the files are added to `list.txt`); it opens with a
title naming the selection; blocked verbs stay in place, greyed, with
their reason; each verb has one refusal wording.
- Extrude defaults to Join when the profile touches a solid, otherwise
New body. All result menus use one vocabulary: New body / Join / Cut /
Intersect.
- Interference and volume/area reports go to the status line, in
mm³/mm². Delete Body no longer asks for confirmation, since it is
undoable.
- **Text is its own feature** ("Text N" in the tree). Its dialog offers
any installed font (bold, italic) and the height in mm; it is modeless
and opens at the top right of the window, and the outline is drawn in
the view where it will go while typing. Editing the feature reopens the
dialog with its string, font and height. The outlines are saved too, so
the project opens the same on a machine without that font.
- New feature names match the card header (`Extrude 3`) and are
translated. Terminology and units are unified ("Coordinate system",
"Angle (°)").
- The Prepare Sketch/Primitive gizmos can be selected only when the CAD
feature is on.
- MCP: write methods are refused while the tab is busy (including while
the Text dialog is open); a timed-out command never runs; only a socket
is ever unlinked; `fillet`/`chamfer` accept an array of edges.
**Rendering in the Design view**
- Once extruded, a part was hard to read: both lights of the object
shaders sit near the camera, so the sides of a part came out in nearly
the same tone, and nothing marked where one face ends.
- The phong shader gains a studio lighting model, selected by a new
`lighting_model` uniform: a sky/ground hemisphere in world space, a key
light from the upper left and a weak fill, a plastic-like highlight, and
a darker base with a faint sheen toward the silhouette.
`GLCanvas3D::set_studio_lighting()` enables it per canvas; only the
Design canvas does. Every canvas sets the uniform on each use (0 for the
slicer's canvases), so Prepare and Preview render exactly as before.
- Every B-rep edge of a body is drawn as a 2 px dark line, depth tested
and pulled a few pixels toward the eye, so it hides behind the faces in
front. Seams of closed surfaces and degenerate edges are left out
(`GeometryEngine::display_edges`).
**Crash fix (Linux AppImage)**
- Drawing anything in a sketch crashed the AppImage: the constraint list
labels were formatted from narrow literals holding `—`, `·` and `°`;
`AppRun` sets `LC_ALL=C`, so wx's conversion returned NULL and
`wxString::Format` dereferenced it. They now go through
`wxString::FromUTF8`, as the panel's other non-ASCII literals already
do.
Docs updated to describe the design as it now stands: `design_tab.md`,
`interaction-model.md` and the portability note.
No change to slicing, profiles or presets. Existing 3MF recipes load and
rebuild as before: the v3/v4/v5 fixture tests pass.
# Screenshots/Recordings/Graphs
None attached. The behaviour was checked on the fork's Linux tester
AppImages, driven under Xvfb (see Tests).
## Tests
- `libslic3r_tests` on this branch, rebased on `main` (dc0e26918): 943
of 944 cases pass (1 skipped), 79,733 assertions. Linux, system OCCT.
- New or updated cases cover: the solver side and tangency fixes;
trim/offset/mirror on arcs and ellipses; body identity across history
edits; datum remapping, expressions, bindable fields; circular pattern,
hole standards, threads; the round trip of colours, origin and
auto-close; `body_touching_sketch`; multi-edge fillet/chamfer and its
save/load; the crossing/cusp loop analysis and the extrude refusal; the
Text parameters' save/load; `display_edges` on box, cylinder and cone;
revolve about a centerline, about the profile's own edge, refused
through the profile, stale axis index, axis save/load. The
truncated-recipe test accounts for every new tail field.
- All touched GUI translation units compile with GCC, and with clang
`-fsyntax-only` under the macOS job's warning flags. The modified
shaders pass `glslangValidator`.
- `gen_offer_table.py --check` passes, and `xgettext` + `msgfmt
--check-format` are clean on the CAD sources.
- On the fork's Linux tester AppImages (same code on `cad-mainline`),
driven headless:
- drawing lines and rounded rectangles no longer crashes;
- Offset on one side of a rectangle selects and previews the whole
outline; Enter adds it;
- Text: the dialog opens at the top right, "Text N" appears in the tree
and the outline in the view while typing; double-clicking the feature
reopens it and a new string redraws in place;
- a block with a through hole and fillets: faces separate by
orientation, fillets shade round, edges are drawn and hidden behind the
part;
- a half-profile beside a construction centerline: Revolve preselects
the centerline, shows the axis dashed and builds the expected tube
(4000π mm³).
- **Not verified by hand:** multi-edge selection and the red loop
marking; Windows and macOS were not run.
[How to Download Pull Requests Artifacts for
Testing](https://www.orcaslicer.com/wiki/how_to_download_pr_artifacts)
## Summary
Adds the remaining OrcaPrinterAgent work on top of feature/OrcaSonar:
connector capability gating, AMS filament mapping, and the printer
Storage tab. Contains no changes to the other printer agents (Moonraker,
Qidi, Snapmaker, BBL, Creality) — that refactor belongs to its own
branch.
## Changes
Capability gating (IPrinterAgent, OrcaPrinterAgent)
• Parse supported_features, capabilities.protocol.features, and
supported_commands from get_capabilities, including replies without a
flags block.
• Add supports_command, supports_feature, and uses_filament_mapping
queries with defaults that preserve existing agent behavior.
• Gate macro-backed AMS commands on fms plus the advertised command, and
slot metadata writes on filament_slots. Unknown capability means not
supported.
## AMS filament mapping
• Serialize ams_mapping2 into print.gcode_file.filament_mapping, keyed
by array position, dropping {255,255}; omit when empty so unmapped
prints stay byte-identical.
• Refuse a mapped print visibly until the correlation flag and
filament_mapping capability are satisfied, and refuse partially mapped
prints with an unmapped used filament. The GUI predicate shares the
serializer's definition.
• Resolve printer-set trays to the matching Generic preset by material
type and color; skip forced external-spool selection on printers without
an AMS.
• start_local_print_with_record returns non-success so PrintJob falls
back to start_print.
## AMS device UI on non-Bambu agents
• Resolve the printer model from the agent model id or selected profile
so the AMS Materials Setting dropdown and Confirm button work;
third-party trays keep editable temps.
• Render explicitly-empty trays as "Empty" via a wire-derived is_empty,
keeping the slot editable.
• Hide flow-dynamics K/N controls for non-Bambu agents, leaving Bambu
behavior unchanged.
## Storage tab (printer G-code files)
• New IPrinterAgent file APIs — list_printer_files,
get_printer_file_thumbnail, delete_printer_file,
get_printer_file_metadata — with ORCA_NETWORK_ERR_CMD_NOT_SUPPORTED
defaults.
• NetworkAgent forwards to the bound agent; OrcaPrinterAgent implements
them against /server/files.
• New StoragePanel / OrcaFilesPanel and a Printer/ file-grid abstraction
(IFileGridModel, RemoteFileGridModel, BambuFileGridModel).
## Tests
Covers serializer keying/omission, GUI/agent gate agreement, payload
shape, capability parsing and per-command AMS gating, model resolution,
empty-tray classification, and AMS sync.
The only use of Slic3r::create_thread in this file sits behind
!BBL_RELEASE_TO_PUBLIC && __WINDOWS__, so the include looked unused in
every configuration CI builds. Windows Debug and RelWithDebInfo builds
failed without it.
A Linux job configures without the precompiled header and runs clang-tidy over the C++ lines a pull request changes. The only check for now is misc-include-cleaner for missing includes; .clang-tidy is where further checks get enabled.
scripts/run_clang_tidy.sh (Linux, macOS) and scripts/run_clang_tidy.ps1 (Windows) run the same check locally: the same configure, the clang-tidy version pinned in scripts/clang_tidy_requirements.txt, and the same comparison against OrcaSlicer's main. They offer to install what is missing, or print the command to do it by hand.
Resolve the filament list's printer model from the agent-reported model id or the selected printer profile, so the AMS Materials Setting dropdown and its Confirm button work on OrcaSonar. Third-party trays pre-fill and allow editing their temps, and the popup's initial selection no longer overwrites them.
Render explicitly-empty trays as "Empty" instead of "?" using a wire-derived DevAmsTray::is_empty, while keeping the slot editable.
Hide the flow-dynamics K/N controls for non-Bambu agents — tile, dialog, validation and sends — with Bambu behavior unchanged.
Tests cover model resolution, the K/N gate and empty-tray classification.
Drop the index-correlation gate so mapped prints are gated only on the filament_mapping capability
Resolve AMS trays set on the printer UI to the matching Generic preset using their material type and color
Skip the forced external-spool selection on printers without an AMS and add coverage for the mapping gates
Parse connector-scope capabilities from get_capabilities: supported_features, capabilities.protocol.features and supported_commands are read for orca devices even when the reply has no flags block, reset before each parse, and left untouched for Bambu.
Gate macro-backed AMS commands on fms plus the advertised command (change filament, select tray, control, user settings, RFID, drying stop); gate slot metadata writes on filament_slots. Unknown capability is not support.
Serialize PrintParams::ams_mapping2 into print.gcode_file.filament_mapping with the array position as filament_index and {255,255} dropped. The field is omitted when empty, so an unmapped print stays byte-identical. A mapped print is refused visibly while ORCA_FILAMENT_MAPPING_CORRELATION_VERIFIED is false or the connector did not advertise filament_mapping; the GUI predicate shares the serializer's definition so both gates agree. Refuse a partially mapped print that leaves a used filament without a target.
Return a non-success result from start_local_print_with_record so PrintJob falls back to start_print instead of treating an unsent print as success.
Tests cover serializer keying and omission, GUI/agent gate agreement, payload shape, the with_record result, capability parsing, and the per-command AMS gate.
* Remove Unused Project Includes and Forward-Declare Where a Type Is Only Referenced
Generated with include-what-you-use and applied conservatively. Only OrcaSlicer's own headers, the ones under src/ and tests/, are removed or forward-declared; standard-library and third-party includes are left alone. An include is removed only when both the Release and the Debug configuration leave it unused, never from inside a conditional block, and never from a file with platform-specific blocks, which only gain includes. Files whose only use of a header sits behind a feature or debug macro (libvgcode's OpenGL ES and marker code, the ARACHNE/TESTS_EXPORT_SVGS debug output) keep their includes.
clonable_ptr.hpp gains #pragma once; it had no include guard and was only safe while Config.hpp was its sole includer.
* Remove Unused Project Includes From Files With Platform-Specific Code
A Linux include-what-you-use run cannot see the code inside _WIN32, __APPLE__ or __linux__ blocks, so its verdict is only taken where nothing the removed header declares, directly or through what it includes, is named inside those blocks. Removals also have to hold in both the Release and Debug configuration and never touch a line inside a conditional block.
* Restore the libslic3r Precompiled Header and Direct Includes Lost in the Platform Pass
The platform-file pass treated pchheader.hpp as an ordinary header and
emptied it, and left GUI_Preview.hpp and 14 other files relying on
headers they no longer reached directly.
* Restore MainFrame.hpp in ParamsDialog.cpp for the Windows-Only Reparent Call
* Include Headers That Files Reached Through Ones the Cleanup Removed
* Drop Includes Duplicated by the Cleanup or by Main's Own Additions
* Leave PreciseSeam.cpp as Main Has It After the Precise Seam Rework
The Design view drew sinking outlines by looking up its bodies in the
plate's model, reading past the end of an object's volumes once a body
was committed. The Design canvas no longer draws sinking outlines.
Merged by /bot merge on behalf of @peachismomo (id 52488812).
Grants: resources/profiles/OrcaFilamentLibrary/filament/Elegoo, resources/profiles/OrcaFilamentLibrary.json, resources/profiles/Elegoo, resources/profiles/Elegoo.json
Head: 00276bf8ee
Commit to Plate now sends all visible bodies to Prepare as one object with a
part per body, so their relative placement survives. A dropdown beside the
button switches to Commit to Plate (as bodies), the previous one-object-per-body
behaviour, and the choice is remembered.
GLGizmoSlaSupports, GLGizmoHollow, GLGizmoFaceDetector, GLGizmoText and GLGizmoAdvancedCut were already left out of the build, and GLGizmos.hpp, the only header including some of them, had no includers. VoxelizeCSGMesh.hpp uses types that no longer exist, SLA/bicubic.h does not compile, and Utils/ProfileDescription.hpp is included nowhere. Their CMake and gettext source-list entries go with them.
Brings `belt-printer` up to date with `main` (4b4a261787) so that #14394
merges cleanly again, and adds the follow-up fixes the merge needs. This
PR targets `belt-printer`, not `main`.
## Commits
1. **GCode: hold the writer by value again.**
- Belt printing had turned `GCode::m_writer` into a `unique_ptr`, so
that `BeltGCode` could swap in a new writer carrying the belt
kinematics.
- Nothing subclasses `GCodeWriter`, and `set_kinematics()` can install
the belt mapping on the existing writer. This commit removes the swap.
- About 190 `m_writer->` edits revert, which takes `GCode.cpp` from 29
conflict hunks with main down to 3.
- The G-code is identical to the current `belt-printer` head on two
BabyBelt projects (see Verification).
2. **Merge upstream/main.** The resolutions are listed in the merge
commit. The ones that needed a decision:
- `write_belt_header()` follows main's relocated header block (#15897,
#15915).
- First-layer acceleration keeps the per-path first-layer plane test,
now with main's cached nozzle index (#16028).
- The arc-to-polyline fallback moves into the out-param
`extrude_arc_to_xy`, which is the overload `GCode` now calls (#16108).
- The belt fields join `GCodeProcessorResult`'s forwarding assign.
- The Clipper2 renames (#15969).
- Belt printers still reserve no CLI wipe tower (#15837).
- The new sparse-layer tower options are hidden for belt printers
(#15841).
3. **Belt: register the raw G-code toggle as a Preview shortcut.**
- Main's assignable shortcuts (#15706) replaced the key switch that
carried **B**.
- The toggle is now `ToggleBeltRawGcode`, bound to B in the Preview,
where B was free. It is listed in the shortcuts dialog and can be
rebound, and the legend shows whichever key is bound.
4. **Precise Seam: slice modifiers in the belt slicing frame.**
- The new `slice_single_volume_regions()` (#16072) sliced modifiers with
`trafo_centered()`, so on belt prints the modifier regions landed in the
unrotated frame.
- It now uses `trafo_sliced()`, as the seam enforcers and support
volumes already do. On non-belt printers the two transforms are the
same.
- A regression test is included.
5. **Belt profiles: inherit what they repeat and pass main's profile
checks.**
- The BabyBelt Pro and IR3 V2 filaments name their single extruder
variant, as the library-based filaments of other vendors do, and drop
overrides that repeat the library value.
- The Custom belt base inherits `printer_extruder_id` from its parent.
- `normalize` drops the obsolete keys. `fix-variant` gives the machine
limits their silent-mode entry, which they previously read from the
single value.
- All three vendor versions are bumped.
- Printcepts and IdeaFormer keep their own machine and process bases,
because only filaments can inherit across vendor bundles.
6. **Profile validator: accept a belt printer's tool change without a
tower.**
- The slice sweep (185cfe4323) forces a prime tower and requires its `CP
TOOLCHANGE START` block.
- Belt printers have no wipe tower: they purge into a prism object.
Their filament change is the plain `T` command, which they emit
throughout the slice.
- On belt printers the validator now looks for that `T1` line instead.
## Verification
All on Linux.
- **`m_writer` equivalence.** The current `belt-printer` head
(b22384a559) and commit 1 were each built and used to CLI-slice two
BabyBelt benchy projects (389k and 804k lines of G-code). The output is
identical apart from the per-run object ids in the `printing object …
id:` comments.
- **Merged branch, belt-specific checks.**
- The belt header is written, no `;_BELT_BAND` markers leak, and the
belt axis never steps back.
- Against the Oct 3 belt + main merge, the only G-code differences are
fill ordering on a few layers and time estimates. Both come from main's
changes since then.
- Against the pre-merge output the differences are much larger. That is
expected: main's CLI now refreshes a project's settings from its system
presets (#15953, #16038), so for example `z_hop` follows the belt
profiles' 0.
- **Tests.**
- `fff_print_tests`: 345/345 test cases pass, including 30 `[belt]`
cases.
- `libslic3r_tests`: 1081 passed, 2 skipped.
- The new Precise Seam test fails 266 of its 284 assertions with the fix
reverted.
- **Profiles.**
- `scripts/orca_profile_tool.py check` passes for all 69 vendors.
- `OrcaSlicer_profile_validator -s -l 2`: all 1271 slices succeed.
Before commit 6, the six belt printers failed.
- A flattened before/after snapshot of every belt preset shows no value
a belt printer reads has changed.
- **GUI** (BabyBelt Pro, clean datadir):
- B switches the Preview between the designed and the raw machine-frame
G-code, and pressing it again restores the view exactly.
- The legend reads "Show raw G-code (belt only) [B]".
- The shortcuts dialog lists the toggle under Preview → Display as
rebindable.
- The sparse-layer tower options stay hidden in Advanced and Expert
modes.
Main has since gained one CI-only commit (f3d0b8a553), which merges
cleanly on top.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
https://claude.ai/code/session_01AJzy1xeQV3FePh5HfahDyn
The validator slices every printer with two filaments and the prime tower
forced on, then requires the tower's CP TOOLCHANGE START block as proof
that change_filament_gcode ran. A belt printer has no wipe tower: it purges
into a prism object on the belt, so Print::has_wipe_tower() is false and
the change is the plain T command set_extruder() emits. All six belt
printers failed the sweep on that alone, although each one changes
filament throughout the slice. Look for the T1 line on belt printers.
main's profile checks reject the belt bundles: their filaments override
variant keys with one value under the library's six-variant presets, the
copied vendor commons carry keys the slicer no longer reads, the IR3 V2
and BabyBelt Pro machine limits miss the silent-mode entry, and the Custom
belt base pins printer_extruder_id to one entry under a three-variant list.
- The BabyBelt Pro and IR3 V2 filaments name the one extruder variant
their printers have, as the other vendors' filaments built on the
library do, and drop every override that only repeats the library value
(diameter, density, temperature range, most of the fan settings...).
The eSUN filaments inherit the narrowed list.
- The Custom belt base inherits printer_extruder_id from its base.
- normalize drops silent_mode, adaptive_layer_height and
tree_support_with_infill; fix-variant gives the machine limits their
silent-mode entry, which they read from the single value before.
- The IR3 V2 drops two limits equal to its base in both modes.
- Custom's index is regenerated and all three vendor versions bumped.
Flattening every belt preset before and after, nothing a belt printer
reads changes: printer_extruder_id is one id per variant, all 1.
Printcepts and IdeaFormer keep their own machine and process bases: only
filaments can inherit across vendor bundles (from OrcaFilamentLibrary), so
they cannot build on the Custom belt printer.
slice_single_volume_regions() sliced Precise Seam modifiers with
trafo_centered(), but a belt printer slices its layers with
trafo_sliced(): the belt rotation, any pre-slice remap and the lift off
the plate on top. On a belt print the modifier regions landed in the
unrotated frame, away from the walls they were meant to place the seam
on. Slice them with trafo_sliced(), as the support volumes and the seam
enforcers already are. It equals trafo_centered() off a belt printer.
Main's assignable shortcuts replaced the canvas key switch that carried
the belt "show designed / show raw G-code" toggle on B. Register it as
ToggleBeltRawGcode, bound to B in the Preview (B is only taken on the
Plater, by the mesh boolean gizmo), so it can be rebound and is listed in
the shortcuts dialog. The legend checkbox shows whatever key is bound.
Brings belt-printer up to main 4b4a261787. Resolutions:
- G-code header (#15897, #15915): main moved the header, config and
thumbnail block later in _do_export; write_belt_header() moves with it,
still after the thumbnails and outside the BTT_TFT gate.
- _extrude: first-layer acceleration keeps the per-path first-layer plane
test with main's cached nozzle index (#16028); main's set_speed out-param
form (#16108) everywhere else.
- GCodeWriter (#16108): the arc-to-polyline fallback for machine mappings
that cannot express G2/G3 now runs in the out-param extrude_arc_to_xy,
which is the overload GCode calls, and appends to the caller's string.
- GCodeProcessorResult: the belt fields join main's forwarding assign.
- Clipper2 (#15969): belt arrange helpers take Slic3r::Point; the tree
support join types lose their ClipperLib qualifier.
- CLI arrange (#15837): belt printers still reserve no wipe tower.
- Wipe tower options (#15841): the two new sparse-layer toggles are hidden
for belt printers like the rest of the tower options.
- Keyboard shortcuts (#15706): main's registry replaces the old key switch;
the belt view toggle is re-registered in the next commit.
- Print::process: the belt purge-plan undo runs before main's SliceStarted
event.
- scripts/filament_id_snapshot.json: deleted on main (a77209af8f).
- Includes and appended tests: union of both sides.
Belt printing turned GCode::m_writer into a unique_ptr so BeltGCode could
swap in a freshly built writer carrying the belt kinematics. Nothing
subclasses GCodeWriter: the machine mapping lives in its MachineKinematics,
which set_kinematics() installs on an existing writer. A GCode is built for
every export and the only state on the writer when init_belt_writer() runs
is the plate offset, which the swap had to copy across by hand.
Install the belt kinematics on the writer in place, drop the copied offset,
and drop the virtual markers on GCodeWriter that the old subclass needed.
Every m_writer-> in GCode.cpp goes back to m_writer., which is most of the
belt diff in that file and most of its conflicts with main.
The pressure-advance pattern keeps its shared_ptr writer: the unique_ptr
kinematics make GCodeWriter move-only and that class must stay copyable.
yw4z