Two decisions from the user, 2026-08-23, after the first pass at making rename
reachable.
ONE SURFACE CARRIES THE VOCABULARY, and it is the row's own menu: Rename (F2),
Edit, then Move up, Move down, Show / hide, then Delete. Offered as a choice
between completing the menu or completing the header icons; the menu won because
the element you click answering with what applies to it is this fork's charter,
and because a menu grows without spending an icon nobody recognises. The header
icons stay exactly as they are — a quick bar for the common three — so nothing
that worked yesterday moved.
The menu is grouped rather than listed: what the row IS (name, contents), where
it SITS (order, visibility), and what removes it. Right-click SELECTS what it
points at before opening, so the menu can never act on a row other than the one
under the cursor.
MOVE BODY LEAVES THE FEATURE-TREE HEADER. It never belonged there: that header
sits over the FEATURE tree, and the button had to guess its subject from
whatever happened to be selected — a feature row got answered with "Select a
body to move it", an instruction about a different kind of object in a list that
does not contain one. A body now has its own action row on the Bodies card:
Move, Show / hide, Delete, Colour. The last three are deliberate COPIES of
feature-tree actions, not a reorganisation: a body row is a different subject,
and someone working in the Bodies list should not travel to another card to hide
or recolour what they have just selected. Each handler already resolves the body
row itself (on_toggle_visibility, on_delete_body, on_set_body_color), so the
card gives them a home and decides nothing.
The offer already carried Move for a selected body (verb `transform`, accepts
body_solid), so that half of the requirement was in place and is untouched.
One piece of the old button was function, not clutter: it also scaled imported
Text/SVG artwork, which IS a feature-row action. That moved to the row's own
menu as "Scale artwork", shown only on a row that has imported regions — so the
capability survives the split instead of disappearing with the button.
Verified on the rig, by the gestures themselves: right-click a row -> the six
items in their three groups; choosing Move down turned ['Sketch1', 'Sketch2']
into ['Sketch2', 'Sketch1']; and with a body present the Bodies card shows its
four actions while the feature header no longer shows Move.
Gate green: ALL LADDERS HELD — offer table OK, kernel 188 cases / 2532
assertions, engine rungs 1-8, 977-sheet corpus + the heaviest sheets, gesture
ladder 98/98, offer ladder 108/108.
RIG DISCIPLINE, learned the expensive way in this session: ladder-all.sh does
NOT relaunch the app, so hand-driving the rig immediately before it leaves state
the ladder's reset_document() does not clear — here the first rung drew nothing
at all and reported "sides []", which reads exactly like a regression in the
rectangle tool. Relaunch the app before a gate, and re-run before believing a
failure that appears in rung one.
User report 2026-08-23: "on feature tree, i cannot rename sketch name".
The rename was not broken. Verified on the rig before changing anything: select
the row, press F2, type, Enter — Sketch1 becomes Base, the name reaches
m_doc.features[idx].name and sync_recipe_to_model() persists it. The offer's
btn:rename verb does the same. What was missing was any way to find that out.
Everything a person would try did something else:
- the pencil in the section header is EDIT (on_edit_feature)
- a double-click fires wxEVT_TREE_ITEM_ACTIVATED, which is also Edit
- none of the seven header icons renames
- right-clicking a row did nothing at all
and the comment above the label-edit handlers claimed a "slow double-click"
renames, which does not survive wxGTK: the activation wins and the sketch opens
for editing instead. So the only route was an undocumented function key, and the
report is exactly right from where the user stands.
The row now answers the gesture people actually use on a named row: right-click
gives Rename (F2) / Edit / Delete, with Rename opening the in-place editor on
that row. Right-click SELECTS what it points at first, so the menu can never act
on a different row than the one under the cursor.
And selecting a row now says what the row can do: "Sketch1 selected — F2 or
right-click renames it, double-click edits it". Cheaper than a tooltip nobody
hovers, and it uses the status line that already exists for exactly this.
A note on the surface, since it is a design call: the CANVAS right-click is the
offer, this fork's single adaptive menu, and this is not that. A tree row is a
different surface, and its menu is three items about the row. Routing tree rows
through the offer would mean teaching the offer a selection kind that is not
geometry, which is a larger change and not what this report needed.
Verified on the rig by the gesture it names: right-click the row, choose Rename,
type, Enter -> ['Sketch1'] becomes ['Base profile'] in describe_scene.
Gate green: ALL LADDERS HELD — offer table OK, kernel 188 cases / 2532
assertions, 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.
Mixed-colour slots are virtual and never flushed. Guard auto_calc_flushing_volumes_internal against them as BambuStudio does, and make the flushing dialog's default matrix and the sidebar 'modified' comparison physical-only so the untouched mixed rows no longer count as a user edit and the Re-calculate result matches the physical-only table.
Two user reports from the same session on the deployed build, 2026-08-23.
FIRST: "if I select a shape (es a circle draw in construction lines) and then I
try to toggle contruction to obtain a full line, does not work". Reproduced: Q
converts the selection and so does the offer's Reference > Construction row —
both run m_keys_sketch['Q'] — but the CHECKBOX, the one control actually
labelled Construction, only ever called set_sketch_construction(), which arms
the mode for the NEXT entity. So the obvious control was the one route that
could not convert existing geometry, and it failed silently while also flipping
the draw mode behind the user's back. It now carries Q's meaning.
Scoped to Select mode, and that scoping is not cosmetic: drawing AUTO-SELECTS
what was just drawn (draw-then-edit), so with a draw tool armed "there is a
selection" does not mean the user picked anything — it means they finished a
line. The first version converted there and turned the box into a trap: arm
construction, draw the axis, click the box to go back to real geometry, and
instead of disarming the mode it converted the axis just drawn. The gesture
ladder's C4 rung does exactly that and reported three construction entities
where it wanted one. In Select mode the intent is unambiguous.
SECOND, and this one destroyed work: "after creation of a circle, a round angled
rectangle and a slot, and mirror of those shapes on a vertical line inside a
outer rectangle, preview is ok but application creates errors: the circle is
mirrored, but rectangle and slot are redrawn as pieces of circles screwing both
the original shapes and the copies." A screenshot came with it, and it showed
more than the words did: the ORIGINALS were wrecked too — the rounded rectangle
was drawn as a four-lobed cloud, each corner fillet having gone the long way
round, and the slot had ballooned into two near-full circles.
Measured on the rig, a slot mirrored about a vertical line:
rails 62.873 / 62.873 -> 2.082 / 62.913
caps r=21.554 sweep=-180.00 -> r=32.214 sweep=-237.66
and all four sources moved, the axis line with them
Cause: confirm_op's Mirror branch bound an Arc copy to its source with a
Symmetric constraint on the CENTRE ALONE. An arc has five degrees of freedom;
pinning two of them leaves the endpoints and the sweep free while the shape's
own coincidences still pull on them, and the solver answers with a different,
internally consistent sketch — which is what a reflex cap and a 2 mm rail are.
A circle came through the same code untouched because a circle HAS no endpoints
to leave free, which is exactly why the failure reads as "circles fine, rounded
rectangles and slots destroyed".
Three parts, and each one is here because the measurement caught the previous
one being half a fix:
1. Arcs are bound by BOTH ENDPOINTS. Endpoints before centre in the ladder:
{p0, p1} is four equations against five DoF and pins the sweep, while
{centre, p0, p1} is six and is refused — the refusal is what silently
degraded the batch to a set that left the sweep free.
2. Every copy is reflected from the PRE-BATCH source, so a batch that disturbs
the sketch cannot hand the next copy already-moved geometry.
3. THE APPLIED RESULT IS THE PREVIEW — checked on the sources AND the copies,
and on violation the whole constraint web is dropped and both halves are
restored to the reflection the preview drew. try_add_constraints rolls back
only when a solve FAILS, and every failure here came from a solve that
succeeded at something else. Guarding only the sources fixed the slot and
left the rounded rectangle's copies at a 13.8 mm rail and a 308 degree cap:
the original was safe and the copy was still wrong, which is half a fix.
The parametric link is kept whenever it provably holds the geometry, and dropped
when it does not. A wrong shape is worse than an unlinked one.
WHY NOTHING CAUGHT THIS: the gesture ladder's mirror rung reflects three
straight LINES. It sat green through the whole defect. C4b now mirrors a slot,
so the reflection has arcs in it, and grades the property the user actually
stated: the copy is the source reflected, the source does not move, and no cap
comes back reflex.
VERIFIED against the user's own scene, rebuilt gesture by gesture on the rig —
outer rectangle, circle, rounded rectangle and slot, a vertical CONSTRUCTION
line as the axis, all 17 entities mirrored in one gesture:
ok the mirror axis is a construction line
ok picked the axis and all 17 entities
ok originals unchanged (moved: [])
ok every copy is the exact reflection (worst 0.000000000)
ok no source arc turned reflex — the 'cloud' failure
ok no copied arc turned reflex (6 arcs checked)
One grader correction worth recording, because it cost a round and would cost
the next one too: a reflection REVERSES ORIENTATION, so a copy legitimately
stores p0/p1 the other way round. Comparing p0 to p0 grades the storage order,
not the geometry, and reported a perfect mirror as an 8.98 mm error. Endpoints
are compared as an unordered pair.
User report, 2026-08-23, after using the freshly deployed build: "selection and
removal of existing elements of the 2d sketch is not intuitive, and the bottom ui
text does not illustrate what the user has to do to properly use the selected
tools. Mirror, for example, does not indicate: first select mirror line then
entities to be selected, and there is no UI indication of what is being selected.
normally, costruction lines are dotted." Four defects, all of them in the 2D
vocabulary this fork's charter puts at the centre, and all four fixed here.
snaporca-1c0c (P1) — the prompt was written ONCE, when the tool was armed.
DesignPanel's select_tool lambda set a sentence and nothing ever revised it, so
every step after the first was unguided: Mirror said "pick axis, then entities"
and then never said which of the two you were on; Escape silently downgraded an
armed tool to Select (request_exit is layered: anchors, then tool, then session)
while the line still named the tool you had left; and nothing ever mentioned that
Del removes a selection. The fix moves the line off the arm event and onto the
tool's LIVE state. DesignSketchTool::emit_step_hint() reports (mode, step, picks)
whenever that triple moves, from render() — the one place every state change in
this tool passes through. Putting it there instead of in the thirty-odd branches
of on_mouse is the whole point: a per-call-site notification is a thing the next
tool forgets to add, and it costs three int comparisons a frame. DesignPanel owns
the words, in ONE table (sketch_step_prompt), whose step numbers are the same ones
render() previews and on_mouse consumes, so the description cannot drift from the
code that reads the clicks. Every tool now names the gesture that ENDS it, because
none of them was discoverable: an empty click applies an edit-op or a transform,
right-click cancels it, Esc goes back to Select.
snaporca-vd6v (P2) — Mirror mirrors its axis pick and its target picks into
m_selection, so both painted white and the picture could not answer "what did I
select as what". The edit-op's first pick — Mirror's axis, Fillet/Chamfer's first
line — now paints violet. Violet and not cyan: cyan means SELECTED in this canvas
and nothing else may wear it, a rule this file already carries in writing.
snaporca-imlq (P2) — construction geometry drew as a solid grey line. Every CAD
dashes it, and grey alone does not read as "reference" against the
under-constrained orange. dash_polyline() chops the polyline before it reaches
draw_quad_strip, with the dash and gap in world units scaled by units-per-pixel,
so a dash keeps its size on screen instead of becoming a solid line when you zoom
out and three dashes when you zoom in.
snaporca-oql1 (P2) — Backspace now deletes as Del does. On every laptop this runs
on, Del is a chord and Backspace is what a hand reaches for. The Select-mode
prompt states the rest (Shift-click adds, double-click takes the loop, Del
removes), and the first step of every armed tool names the Esc route back to
Select, which was the invisible half of "selection is not intuitive".
Also, on the same report: the sketch stroke half-width goes 0.6 -> 0.3 mm. At 1.2
mm wide the orange line swallowed a short segment and hid which of two near
parallel lines the cursor was on. One constant, because all twenty call sites of
draw_quad_strip are sketch strokes.
Retired on the way: the on_sketch_selection_changed status writer. It said "N
selected — Delete removes them" while an edit-op mirrored its picks into the
selection, i.e. in the middle of a Mirror gesture, where Delete does nothing of
the sort. on_sketch_step says the true thing for Select and says nothing false
anywhere else. And the live length/angle readout is now APPENDED to the step
guidance rather than replacing it: it fires on every mouse move, so it used to
erase the instruction for the step in progress one move after the click that
started it.
Two false trails, recorded so the next session does not walk them again:
- DesignPanel.hpp deliberately does not include DesignSketchTool.hpp, so the
panel's handler takes the mode as an int and the .cpp casts it back. The
first attempt put Mode in the header signature and the build said only
"expected ',' or '...' before 'mode'".
- The offer ladder failed six properties against a perfectly good binary
because I had relaunched the rig myself without SNAPORCA_KEYTRACE=1, and its
[OFFER] trace lines ARE its instrument. A ladder with no instrument reports
"None", which reads exactly like a regression in the offer. ladder-all.sh
launches it correctly; a hand relaunch must too.
VERIFIED, not merely compiled. Driven on the headless rig with synthetic mouse
and keyboard, and photographed at each step: "Mirror — first click the LINE to
mirror about (a construction line works) · Esc goes back to Select" ->
"Mirror — axis set · now click the entities to mirror · right-click cancels"
-> "Mirror — axis set · 1 to mirror · click another to add or remove it ·
click empty space to apply", with the axis violet, its target white, and the
construction lines dashed while real geometry stays solid.
Full gate green afterwards (scripts/ladder-all.sh, ALL LADDERS HELD):
offer table vs the atlas OK
kernel suite 188 test cases, 2532 assertions
engine ladder rungs 1-8, ALL RUNGS HELD
corpus rung 977-sheet drawing corpus, every 20th -> 49
sampled, 39 gradeable, 39 clean
corpus scale rung the 6 heaviest sheets, all clean
gesture ladder 93/93 properties, real mouse and keyboard
offer ladder 108/108 properties, through the right-click
menu and the verbs behind it
That harness is the reason a UX change of this size can be made in one pass and
believed: 93 + 108 properties are driven the way a person drives the app, and the
977-sheet corpus keeps the engine underneath them honest against real drawings
rather than against my own arithmetic.
It passed alone and failed inside scripts/ladder-all.sh, twice, on the same property: "right-click
with two picked -> None". The diagnostic it now prints said what a screenshot could not — editing=True
with an empty selection after two clicks that should have picked two entities. Three rig facts, all
about the DRIVER, none of them a product defect:
A LEFT-CLICK ON A LINE'S MIDDLE OPENS ITS LENGTH FIELD. The Select branch tests m_live_quotes before
it picks, with a ~24 px label tolerance, and a line's Length quote sits at its midpoint — so the
click that was meant to select it promoted a dimension and froze the canvas instead. Everything
after it landed on nothing. It bit only when the previous step had left that line selected, because
live quotes are drawn for the SELECTION: hence passing alone and failing in the gate. Lines are now
picked at 0.3 along, clear of the label.
A FIELD THAT HAS NOT OPENED YET READS LIKE ONE THAT NEVER WILL. The queue opens each field from a
CallAfter, so a driver that looks once, sees nothing and moves on gets frozen by the field that
arrives a moment later. keep_as_drawn() now waits for QUIET — two consecutive clear readings — and
draw_line_at drains stragglers before handing back.
A KEY CANNOT CANCEL A SESSION WHOSE CANVAS IS FROZEN. gui-ladder's enter_sketch dismisses the old
sketch with Escapes, which an open field swallows, so the session survived and the four calibration
probes landed in it on top of what was already there: "calibration expected 4 points, got 7". Every
rung now enters through fresh_sketch(), which cancels through the socket first — that cannot be
swallowed.
Measured after the fix, in the sequence that failed: gesture ladder 93/93 then offer ladder 108/108,
back to back on the same app.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
snaporca-ziam said gen_offer_table.py would silently delete the model-mode "Constrain sketch"
row, because that row lived in the generated header and not in tool_atlas.json. Running it found
more than that: FOUR rows existed only in the header — constrain, rename, and the three typed-
value rows sk_length / sk_radius / sk_angdist — and sk_delete's action had drifted, pointing the
sketch row at btn:delete, the FEATURE delete.
All five are now in the atlas, so the header regenerates byte-identically from it. Verbs may carry
a `note`, emitted as a C++ comment above the row: a rationale written into a generated file is
deleted by the next regeneration, which is how this started.
snaporca-z8rs (P1), found by making that true: after the atlas held all 92 verbs, the regenerated
header differed from the checked-in one by EXACTLY ONE LINE — kOfferVerbCount, 91 against 92.
Every consumer loops i < kOfferVerbCount, so the last row of the table was invisible: never listed
by show_offer_menu, never findable by mcp_run_verb. The verb that fell off the end is sk_angdist,
"Angle / distance…" — the typed-value row for a two-entity selection. On the one selection where
you would ask for the angle between two lines, the row that types it was not in the menu.
It survived because nothing compared the Sk2Ent menu against the table: sk_angdist accepts Sk2Ent
and nothing else, so an off-by-one that dropped the LAST verb was invisible from every other
selection. The vocabulary rung now covers Sk2Ent too, and picking the pair taught it one more
rig fact — shift-clicking a circle at its +X point grabs the RADIUS GRIP, which replaces the
selection with that one entity, so the pair silently collapsed to one and the offer answered
SkLine. Correctly, for the selection that actually existed.
gen_offer_table.py --check proves header == atlas and changes nothing; it is now the first step of
scripts/ladder-all.sh, and the only one that needs no rig.
Offer ladder 107/107, gesture ladder 93/93, both on the rig.
snaporca-ziam snaporca-z8rs
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
The 2D vocabulary is 46 verbs: 22 have a shortcut and the gesture ladder drives them, 24 have
none and nothing had ever exercised those. They are reachable only from the right-click offer, so
a key-driven ladder could not have touched them whatever it did. Four new rungs drive all 24, and
the coverage claim itself is now arithmetic against DesignOffer.hpp (rung O8) rather than a
sentence in a comment that rots when a verb is added.
The assertions are CONSTRUCTION invariants wherever a click cannot be exact — a regular polygon's
sides are equal to 1e-9 and its vertices lie on one circle; a tangent arc's radius at the shared
endpoint is perpendicular to the line to 1e-9 (measured cos 5.97e-17); the three clicks of a
3-point circle all lie on it; a circumscribed pentagon's circumradius is the inscribed one's over
cos(pi/5), 1.236067977 against 1.236067977. Where a value field opens, the typed value is graded
exactly: a moved line travels +25.000000000 in X and 0 in Y, a rotation turns 30.000000000 deg
and leaves the length alone, a scale multiplies it by exactly 3, a linear array's pitch is
[20.0, 20.0, 20.0] and a polar one's spokes are 60 deg apart all the way round.
Three defects found doing it, all fixed here:
snaporca-ua9g (P1) — delete_selected left three things behind. The AUTO-EDIT QUEUE, so a queued
field opened on a deleted entity and its commit went nowhere: draw a rounded rectangle, delete
everything, draw a 2-point circle, type 30 — the field opens, the digits are accepted, and the
radius stays 32.992020763. reset_autoedit() exists for exactly this and its own comment says so;
it was simply never called from here. The FEATURE GROUPS, whose [begin,end) ranges all shift on a
delete, so feature_of() answered with a group the user never drew — survivors are now remapped
and any group that lost a member is dropped, the rule the placed quotes already followed. And the
SOLVER STATE: no re-solve, so sketch_describe reported dof=16 for a document holding one circle.
snaporca-ekt9 (P2) — the read-back could not see three of its seven entity types. Ellipse,
EllipseArc and BSpline serialised as a bare type name: no centre, no semi-axes, no rotation, no
sweep, no poles. gui-ladder's ellipse rung had to grade the faceted area of the loop at 2e-2 —
that tolerance IS the faceting error — and its spline rung could only count entities. Now they
carry their parameters, and the ellipse arc's ends are asserted to satisfy (x/a)^2+(y/b)^2 = 1 to
1e-9.
Also read-only, and the reason the other two were found at all: sketch_describe now reports the
armed TOOL, the count of PENDING anchors, and whether a value field is EDITING. A menu walk that
lands one row off arms a neighbouring tool and then draws something plausible — the first run of
the authoring rung drew a circle of area 45238.93 and graded it as a rectangle. Every menu pick
now asserts which tool it armed, and the polyline rung (a per-segment Length field freezes the
canvas after every click) could only be written once the driver could ask whether a field was open.
Offer ladder 102/102 -> 105/105 with coverage. Gesture ladder 93/93 and the kernel suite
188 cases / 2532 assertions, both unchanged.
snaporca-ua9g snaporca-ekt9
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
The gesture ladder proved the TARGET — a complex closed profile, exact in vertices, lengths, arcs
and symmetry, voids correctly attributed. It proved it by arming every tool with a letter key,
which leaves the goal's own MECHANISM untested: the design logic pivots on right-click, and the
verbs offered are supposed to adapt to the element under the cursor. 47 of 86 Design-tab verbs
have a GUI action and no shortcut, so a key-driven ladder cannot reach more than half of them.
scripts/offer-ladder.py drives the menu. It asserts nothing from pixels: show_offer_menu emits an
[OFFER] trace from the same loop that builds the rows (behind the existing SNAPORCA_KEYTRACE), so
what the ladder reads cannot drift from what the user is shown, and the expected row set is
predicted by parsing DesignOffer.hpp rather than transcribed by hand. 25 properties, four rungs:
what each element type offers, that the menu equals the table for four selections AND that the
four differ, a 120 x 80 profile authored entirely through the menu, and a tool with no keyboard
route at all driven from the only door it has.
Two real defects, both found by it, both fixed here:
snaporca-ghcz (P1) — right-click was a black hole while any draw tool was armed. Every draw case
ended with `if (evt.RightDown()) { m_points.clear(); return true; }` and returned true even with
nothing to abandon; on_mouse records that in m_right_consumed and DesignCanvas suppresses the
offer whenever it is set. Measured: with Line armed, two right-clicks in a row produced no menu
and no tool change; only Escape freed it. Same rule snaporca-xmh6 wrote for the selection —
clearing nothing is not a gesture terminator. One shared right_abandon() now consumes the click
only when an anchor was really down; 16 sites, plus Polyline/BSpline (which end a chain, correct
only when there IS one) and Point (which has no anchor at all).
snaporca-lnri (P2) — right-clicking a sketch point offered the empty vocabulary. select_at_screen
tests hit_test_point first and records the hit in m_point_sel, but the offer counts m_selection
only, so a Point entity could never reach the entity branch and SkPoint was unreachable by
construction. A Point IS its own handle, so it is selected as an entity; other entities keep the
handle pick, since a line's endpoint is a drag target, not a vocabulary.
Offer ladder 25/25, gesture ladder 93/93 (no regression), both on the rig. The offer ladder joins
scripts/ladder-all.sh as the fifth rung.
snaporca-ghcz snaporca-lnri
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
Draw-then-edit is armed from a jump in the entity count: render() sees n > m_autoedit_seen,
selects the last entity and schedules open_primary_autoedit. A scripted add made while a
creation tool was armed looked exactly like a drawn gesture, so it opened that tool's value
field — and an open field freezes the canvas (on_mouse_impl returns early on m_awaiting_length)
and swallows every letter (in_text includes inline_busy()). Measured on the rig: after
sketch_add, 'p' + click added nothing (4 entities before, 4 after); one Escape and the identical
sequence gave 5. It also explains the selection = [last index] that sketch_describe reported
although action_sketch_add never selects anything — the render pass wrote it.
Escape worked because it sequences two set_tool calls: the pending CallAfter fires between them,
so the second one commits the field it finds open. Arming a tool directly is one call, and the
CallAfter fires after it.
Fix: resync m_autoedit_seen at the end of add_entities_scripted, so a scripted add is not read as
something the user just drew. An already-open field is left alone. Covers sketch_add,
sketch_mirror and sketch_offset — the three callers.
The scale rung's Escape workaround is deleted, which is the issue's acceptance criterion; it is
now the regression test. Gesture ladder 93/93 on the rig.
snaporca-j7gc
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MrMzTpAf78U4NG2M8jfvHY
Over the whole 977-sheet corpus: 767 graded, 767 fully clean, 0 failures. The 210 skips are sheets whose part outline is not a closed stroked path at all (largest loop 5 to 132 mm2, measured), and one of them — MPD133 — is password-protected, which was being reported as an engine ERROR. Both now say what they are.
snaporca-j6sr
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
Tommaso Bianchi