Two turned-shape programs of the frontend replay differed from the
composable model:
- Setting a turned variant to fill in its hugging variant board: the
sidebar fixes the board's width first, the model's property edit did
not, and a layout pass of a fill child in a hugging row collapses a
turned child before growing it, which left it a singular transform.
The model's edit of a layout child now applies the same sizing fixes
(`cls/fix-child-sizing`, `cls/fix-parent-sizing`).
- A swapped head turned with its main, filling a row: the frontend's
layout update settles the placement of the copy shapes it moves, which
gives the head its main's unskewed turn; the model's reflow did not,
and kept the layout's skew. The model's reflow settles them too
(`cll/generate-sync-transformed-placement`).
Both programs now pass in the frontend replay with WASM off (frontend
generated tests).
Harness only; the benchmark is not run.
Also in this commit:
- Log each shape's parent id in the WASM replay's debug output. The debug log
named each shape's parent, which copies and their mains share; it now gives
the parent's id too, so a subtree can be told apart from its copies'.
AI-assisted-by: claude-opus-5-5
`fix-child-sizing` and `fix-parent-sizing` give a layout's children and
parent the sizing a hugging layout allows when the sidebar edits a
layout child. They move from the frontend's shape layout events to
`app.common.logic.shapes`, unchanged, so the composable model can apply
them as the sidebar does.
No behavior change; the benchmark covers sync and is not run.
AI-assisted-by: claude-opus-5-5
A masked group takes all its geometry from its mask. When a copy turned
the mask of a masked group, a sync gave the copy's group the geometry of
its main's group, square to the page, and fitted only the group's
parents, so the group no longer matched its turned mask; a reflow then
fitted it again, and settling and a full sync gave different files.
After a sync writes a masked group's attributes, the group fits its mask
again along with its parents.
Test: generated `a-synced-masked-group-fits-the-mask-its-copy-turned`
(fails before the fix).
Benchmark (bench-components, alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+--------+--------|
| deepest-edit-through-three-levels | 35.215 | 22.245 | 23.118 | 0.63 |
| edit-through-a-chain-of-dependents | 68.931 | 43.743 | 44.839 | 0.63 |
| library-update-into-copies | 83.083 | 54.447 | 57.814 | 0.66 |
| main-edit-among-many-shapes | 4.837 | 2.742 | 2.876 | 0.57 |
| main-edit-into-10-copies | 1.79 | 1.077 | 1.135 | 0.6 |
| main-edit-into-100-copies | 15.415 | 10.488 | 11.093 | 0.68 |
| main-edit-into-1000-copies | 153.086 | 106.222 | 112.37 | 0.69 |
| variant-switch-into-copies | 32.052 | 23.155 | 23.929 | 0.72 |
AI-assisted-by: claude-opus-5-5
Since F570 a flip moves a shape's radius tokens with the corners they
apply to. The composable model kept the tokens of a copy's authored
radius on the corners they named before the flip, so a flipped copy
that had set one corner by hand failed I9.
After each step and each settle, the model's authored radius takes the
tokens in the orientation the shape ends with
(`shadow/radius-orientation`). The token property passes at 300
programs.
Test: generated `a-flip-moves-the-radius-tokens-a-copy-authored`.
Harness only; the benchmark is not run.
Also in this commit:
- Block a shape the model moves into a blocked board. A move relocates a shape
as a nest does, and a shape relocated into a blocked parent is blocked, an
override on a copy. The composable model applied that to nests only, so moving
a copy into a board blocked by hand failed I2.
- Predict a mask's move through the group it moves. The composable model
predicts where a transformed shape lands. The mask of a masked group passes
its change to the group, so the prediction applied the change to the mask
alone and skipped the modifier tree. A flex that places the group puts it back
after the move, so moving the mask of such a group left it where it was, and
the prediction failed I9.
- Compare a pasted copy with the main its file accepted. The composable model
expects a copy pasted on the page to author the geometry it shows when that
differs from its main's. It read the main as loaded, so in a stale library
world, after the library resized a main the file had not accepted, a pasted
copy showing the accepted size was expected to author it, and the property
failed I3.
AI-assisted-by: claude-opus-5-5
Converting a rect to a path gave an outline that went back to its start
with a line and stayed open, so a stroke missed its join at the first
corner. With the WASM engine on, the engine converts the rect and closes
the outline, so the two engines gave different paths for one shape.
The rect's outline now ends with `close-path`, as the WASM engine's does
(user decision). The path editor closes subpaths its own way on entry,
which its lifecycle test now expects.
Found by the WASM replay of the pinned generated programs (program 329),
which passes with the change.
Tests: `path-data-test/shape-to-path-rect-closes-its-outline`.
The conversion is outside the paths the component benchmark measures;
it is not run.
AI-assisted-by: claude-opus-5-5
A bool whose only child was a group kept, after ungrouping that group,
the content it had made of the group: the outlines of its shapes one
after the other, not the boolean result of them. The next refit of the
bool, such as the undo of a later edit of one of its shapes, recomputed
it, so that undo did not give back the state before the edit.
An ungroup now refits the group or bool around the ungrouped shape, with
the undo values of the refit.
Tests: `groups-test/ungrouping-a-group-in-a-bool-combines-its-shapes-again`
(fails before the fix), generated
`an-ungroup-in-a-bool-combines-its-shapes-again`.
The ungroup is outside the paths the component benchmark measures; it is
not run.
AI-assisted-by: claude-opus-5-5
The composable model carried a switched copy's geometry onto the new
head as a plain value, while the other carried shapes keep the parent
whose later resizes size them. Resizing the copy that holds the head then
scaled it through its constraint in production, and the model expected
its old size.
The head now carries its geometry with its parent too, its size only:
its place is its position's.
Test: generated `a-switched-head-in-a-copy-scales-with-its-parent`.
Harness only; the benchmark is not run.
Also in this commit:
- Draw 100 programs per random property on the JVM. The coverage check of a
random composition test asks each kind for 0.5% of the run's changes. At 50
programs a seed could draw a kind too few times: the stale library property
drew its child removal 4 times, all changing state, and failed its check while
its property passed. At 100 programs that seed passes, and the shares vary
less.
AI-assisted-by: claude-opus-5-5
Update main of a copy that swapped a nested copy, pushed inside a copy of
the main, puts the swapped shapes in the position's place and points them
at the position's subtree. Their differing content became overrides, but
their constraints did not: a pushed icon kept the copy's fixed scroll
while its new reference scrolls fixed, untouched and different from its
main.
A shape pushed below the head now marks its constraints group when its
constraints differ from those of its new reference, keeping what the copy
showed, as the head already does. The composable model expects that
override.
Test: generated
`an-update-main-of-a-swap-keeps-the-constraints-its-children-showed`.
Benchmark (bench-components, alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 21.399 | 22.081 | 0.61 |
| edit-through-a-chain-of-dependents | 68.931 | 41.679 | 43.499 | 0.6 |
| library-update-into-copies | 83.083 | 50.272 | 54.817 | 0.61 |
| main-edit-among-many-shapes | 4.837 | 2.258 | 2.591 | 0.47 |
| main-edit-into-10-copies | 1.79 | 1.034 | 1.071 | 0.58 |
| main-edit-into-100-copies | 15.415 | 10.439 | 11.139 | 0.68 |
| main-edit-into-1000-copies | 153.086 | 101.545 | 105.155 | 0.66 |
| variant-switch-into-copies | 32.052 | 23.541 | 24.772 | 0.73 |
AI-assisted-by: claude-opus-5-5
A radius token names a corner as its shape shows it, but a copy mirrored
unlike its main inherited the main's corners as they were. A token value
edit then rounded the corner the main keeps square: the copy's tokens
named its other corners.
A flip now moves a shape's radius tokens with the corners they apply
to, and a transform writes them without making them an override. Sync,
reset, update main and a switch carry tokens into the orientation the
destination ends with (`ctsr/tokens-in-orientation`,
`cll/oriented-by`). The composable oracle compares a copy's radius
group with its main's in the copy's orientation.
Tests: `token-apply-test/a-flip-moves-a-radius-token-with-its-corner`,
generated `a-token-edit-keeps-the-square-corner-of-a-mirrored-copy`.
Benchmark (bench-components, alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 22.352 | 23.145 | 0.63 |
| edit-through-a-chain-of-dependents | 68.931 | 44.184 | 45.036 | 0.64 |
| library-update-into-copies | 83.083 | 52.481 | 54.229 | 0.63 |
| main-edit-among-many-shapes | 4.837 | 2.354 | 2.498 | 0.49 |
| main-edit-into-10-copies | 1.79 | 1.031 | 1.1 | 0.58 |
| main-edit-into-100-copies | 15.415 | 10.729 | 10.946 | 0.7 |
| main-edit-into-1000-copies | 153.086 | 103.496 | 105.968 | 0.68 |
| variant-switch-into-copies | 32.052 | 23.207 | 23.7 | 0.72 |
AI-assisted-by: claude-opus-5-5
The frontend interpreter recorded three library steps unlike the pure
runner, so three replayed programs differed from the model with no
difference in production:
- A detach leaves a nested copy's child plain when the loaded library no
longer has its main; the interpreter now records those children, as the
detach node does, and the oracle reads each as its main's child.
- A duplicate takes the reference inputs of the shapes it copies, so a
duplicate of a copy back on an older library holds that version.
- A library sync remembers its copies against the synced library as
loaded and the other libraries as accepted.
The detach node and the interpreter share `n/record-unlinked-children`.
Frontend generated tests replay the three programs.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
Two dashboard templates, Avataaars and Tutorial for beginners, hold the
filter attribute of some shapes as a map an old reader split out of its
text, as `{:url(#filter_2) nil}` for `url(#filter_2)`. The svg attribute
schema takes only scalars outside `:style`, so their import failed the
file validation after migrating.
Migration 0039 turns such a map back into the reference to one of the
shape's defs it names, matched to the defs id as stored though 0035 may
have kebab-cased the map, and drops any other map value. Both templates
then import valid, with their filters.
The benchmark is not run: the change is a file migration.
AI-assisted-by: claude-opus-5-5
The model kept the geometry a shape authored relative to the parent it
authored it in. A move keeps where and how the shape shows, so moved into
a turned board it was expected turned with it, where production keeps
its turn on the page. The model records the authored geometry again in
the shape's new parent after a move.
Harness only; the benchmark is not run.
Also in this commit:
- Draw fewer random programs in JS and check coverage on the JVM. The random
composition properties drew 50 programs in JS as on the JVM, which made the JS
run of the common tests take about two and a half hours. JS runs the same
model and production code, so it now draws 15 programs per property as a
parity check, and the coverage shares, which need the JVM's 50 programs to
tell, are asserted on the JVM only.
AI-assisted-by: claude-opus-5-5
An update main that adds a copy's new shape to the main makes the
differences between the pushed shape and its new reference authored. A
pasted shape holds empty fill and stroke lists where its main holds none,
so pushing it marked its fills and strokes overridden though nothing
shows differently. No fills or strokes and an empty list of them are the
same there.
Benchmark (bench-components, run alone):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 24.875 | 25.599 | 0.71 |
| edit-through-a-chain-of-dependents | 68.931 | 47.991 | 52.684 | 0.7 |
| library-update-into-copies | 83.083 | 60.901 | 62.895 | 0.73 |
| main-edit-among-many-shapes | 4.837 | 3.193 | 3.451 | 0.66 |
| main-edit-into-10-copies | 1.79 | 1.199 | 1.252 | 0.67 |
| main-edit-into-100-copies | 15.415 | 10.375 | 11.472 | 0.67 |
| main-edit-into-1000-copies | 153.086 | 113.348 | 121.731 | 0.74 |
| variant-switch-into-copies | 32.052 | 23.113 | 24.522 | 0.72 |
AI-assisted-by: claude-opus-5-5
With `PENPOT_REPLAY_PROGRAMS` naming an EDN file of programs, a frontend
generated test replays each of them and logs its own results, which
compares the frontend with the model on many programs without a browser.
Without it the test does nothing.
Harness only; the benchmark is not run.
Also in this commit:
- Give the generated text component a measurable size. The generated worlds'
text component had a board and a text of a hundredth of a pixel, where turning
or moving them left the transform to float noise and the checks reported
differences no design shows. The board is 100 by 100 and the text 100 by 20;
`add-frame-with-text` sets up its text with the size its child params give.
AI-assisted-by: claude-opus-5-5
`oracle/inequivalence` compared a shape's values on their own, so the
geometry tolerance of `equivalent?` did not apply and float noise was
named as the first difference. Each key is compared within the shape it
belongs to.
Harness only; the benchmark is not run.
Also in this commit:
- Replay layout, structure and grid track edits as the UI does. The frontend
interpreter set a layout's gap, padding and item alignment and a layout
child's margin with a plain shape update, where the sidebar goes through the
layout events that reflow the layout; a copy then kept its children where the
old gap put them. It dispatches those events now.
- Watch the targets of a replayed program in the browser. With `watch`, the
browser replay logs the events the store handles and each move of the shapes
the program's operations target.
AI-assisted-by: claude-opus-5-5
A fill child of a grid takes the size of its content, which the grid's
fill then fits to its cell. An absolute child is not filled by the grid,
and the content of a grid whose flex columns span its width is that width
and its padding, so an absolute fill board of a grid grew by its padding
at every layout pass and never settled.
An absolute child keeps its size.
Benchmark (bench-components, run alone):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 23.593 | 26.323 | 0.67 |
| edit-through-a-chain-of-dependents | 68.931 | 50.077 | 55.477 | 0.73 |
| library-update-into-copies | 83.083 | 50.157 | 52.815 | 0.6 |
| main-edit-among-many-shapes | 4.837 | 2.511 | 3.077 | 0.52 |
| main-edit-into-10-copies | 1.79 | 1.172 | 1.215 | 0.65 |
| main-edit-into-100-copies | 15.415 | 9.878 | 12.423 | 0.64 |
| main-edit-into-1000-copies | 153.086 | 123.768 | 175.446 | 0.81 |
| variant-switch-into-copies | 32.052 | 21.933 | 22.474 | 0.68 |
AI-assisted-by: claude-opus-5-5
A text style edit that applies the generated typography in a world whose
file holds none named a typography that resolves to nothing, so the model
wrote a dangling reference the copies then had to match. Without the
typography, that edit clears the text's typography instead.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
Ungrouping a board in a grid main ungroups its copies alike. A copy with
its own cells placed the shapes the ungroup brought in by its own flow,
which adds tracks along its direction: a copy flowing in columns grew a
row its main did not have, so tracks it never overrode stopped following
its main's.
A copy with its own cells now first takes the tracks the ungroup left in
the grid it refers to, on the axes it does not override, then gives its
free cells to the new shapes in its main's layer order.
Benchmark (bench-components, run alone):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 22.032 | 22.764 | 0.63 |
| edit-through-a-chain-of-dependents | 68.931 | 43.547 | 51.026 | 0.63 |
| library-update-into-copies | 83.083 | 52.297 | 57.441 | 0.63 |
| main-edit-among-many-shapes | 4.837 | 2.474 | 4.362 | 0.51 |
| main-edit-into-10-copies | 1.79 | 1.091 | 1.162 | 0.61 |
| main-edit-into-100-copies | 15.415 | 11.465 | 12.265 | 0.74 |
| main-edit-into-1000-copies | 153.086 | 111.337 | 115.898 | 0.73 |
| variant-switch-into-copies | 32.052 | 22.186 | 27.148 | 0.69 |
AI-assisted-by: claude-opus-5-5
With `debug`, the browser replay logs each target and the shapes every
program ends with whether the WASM engine is on or off, so a WASM failure
can be compared with the same program on the plain engine.
Harness only; the benchmark is not run.
Also in this commit:
- Apply the generated typography only where the file holds it. A text style edit
picks the generated typography for one of its values, but only worlds with
text styles add that typography to the file. In the others the model applied
the edit's styles while the frontend applied a missing typography, which sets
no style, so the two diverged. The edit names the typography only when the
file holds it.
- Replay programs in the browser as the frontend tests do. The browser replay
mocks the backend and the timers as the frontend generated tests do, turns the
WASM engine off for the plain engine even when the page's configuration turns
it on, and in debug logs the content a text edit commits and each text's
characters and font weights.
- Name the boards a deleted component's main was held by. The canonical design
data names shapes and components by their place in the tree, so the model and
the frontend compare although each allocates its own ids. A deleted component
keeps the ids of the board that held its main, as a variant's component keeps
its container's once the variant is undone; with that board gone the ids
stayed raw and differed between the two runs. They are named after the
component when no live shape is.
- Report where the model and the frontend first differ. A backend mismatch in
the frontend replay printed the whole diff of the two canonical designs, which
lists float noise the comparison accepts next to the difference it rejects.
`oracle/inequivalence` walks the two designs as `equivalent?` compares them
and returns the path of the first rejected value with both sides; the mismatch
message leads with it.
- Keep text sizes and take inline programs in the browser replay. With the WASM
engine on, the replay keeps each text's stored size in place of the engine's
text resizes: the model keeps stored sizes, and the generated texts are too
narrow to measure, so a new copy's measured text differed from its main's
unmeasured one. The page also takes the programs inline in an `edn` query
parameter, which a shrink loop drives.
AI-assisted-by: claude-opus-5-5
A `:simple` padding or margin type repeats the top and right values on
the bottom and left, as `ctl/paddings` and `ctl/child-margins` read them,
but the WASM engine received the four stored values. A board whose type
was switched to simple after editing its sides laid out its children with
the stale bottom and left values in the WASM engine only, so moving one
child shifted all of them and marked copies as touched.
Both the shape payload and the per-attribute calls now send the values
the layout functions return.
The benchmark covers JVM code; this change is ClojureScript only, so it
is not run.
AI-assisted-by: claude-opus-5-5
A bool shows its segments. Its transform and the order of its corners
say nothing its flags and segments do not: sync orients its children
from its turn and flips, and fitting it resets its transform. A turned
bool in a flipped board could store a transform that does not mirror,
or its corners in another order, depending on which step last fitted
it. Design snapshots now leave its transform out and take its corners
in any order.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
A bool keeps its turn and flips when fitting it resets its transform, so
sync orients its children from them, choosing between the turn before
the flips and the flips before the turn by which is nearer the bool's
axes. A fitted bool's axes are the identity, which tells neither apart,
and the choice fell to the flips first: a flipped, turned copy bool gave
its child a half-turned orientation, which a sync of the settled state
did not. The turn now goes first on a tie, as a flip of a turned shape
stores it.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 22.783 | 25.354 | 0.65 |
| edit-through-a-chain-of-dependents | 68.931 | 51.070 | 52.568 | 0.74 |
| library-update-into-copies | 83.083 | 54.684 | 63.220 | 0.66 |
| main-edit-among-many-shapes | 4.837 | 2.978 | 3.782 | 0.62 |
| main-edit-into-10-copies | 1.79 | 1.212 | 1.281 | 0.68 |
| main-edit-into-100-copies | 15.415 | 12.907 | 13.628 | 0.84 |
| main-edit-into-1000-copies | 153.086 | 107.916 | 123.169 | 0.70 |
| variant-switch-into-copies | 32.052 | 25.136 | 29.877 | 0.78 |
AI-assisted-by: claude-opus-5-5
Run on the dashboard templates, the scan reported thousands of copies
that a sync would rewrite with the same geometry within a hundredth of a
pixel, the same path segments, or a text whose letters and styles show
the same in another spelling. It now compares geometry and path content
within a hundredth and text content by its override groups, so what it
reports is a difference a user can see.
The scan is a debugging tool outside sync; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
Resizing a copy board scales a swapped head in it through its
constraints. The next sync of the head, as the root of its own component,
gave it its component's size back. As a sync that resizes a board already
does, a transform that resizes one now marks the geometry of each
untouched swapped head it gives another size, unless a layout sizes it,
so later syncs keep that size. The harness model authors the same size.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 23.800 | 28.893 | 0.68 |
| edit-through-a-chain-of-dependents | 68.931 | 48.458 | 60.725 | 0.70 |
| library-update-into-copies | 83.083 | 61.325 | 63.038 | 0.74 |
| main-edit-among-many-shapes | 4.837 | 2.523 | 2.669 | 0.52 |
| main-edit-into-10-copies | 1.79 | 1.180 | 1.214 | 0.66 |
| main-edit-into-100-copies | 15.415 | 11.228 | 11.628 | 0.73 |
| main-edit-into-1000-copies | 153.086 | 110.756 | 120.456 | 0.72 |
| variant-switch-into-copies | 32.052 | 24.565 | 28.627 | 0.77 |
AI-assisted-by: claude-opus-5-5
A pasted shape gets empty fill and stroke lists, as the clipboard gives
them, while its main and a later sync of it have none. Both show nothing,
but the oracle and the model told them apart. Values and design
snapshots now take a missing list as an empty one.
Harness only; the benchmark is not run.
Also in this commit:
- The harness changes that check the next commit, "Author the size a resize
gives an untouched swapped head".
AI-assisted-by: claude-opus-5-5
The letters of a text keep, for a paragraph with no letters, the index of
its run among the runs. It was counted after the run was added to the
transient list, so it pointed one past the run: at the next paragraph's
run, or past the end when the empty paragraph came last. Comparing the
styles of such a copy with its main failed, and importing the Penpot
Design System template stopped in the migration that recomputes old
text overrides.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 21.783 | 24.854 | 0.62 |
| edit-through-a-chain-of-dependents | 68.931 | 47.467 | 53.332 | 0.69 |
| library-update-into-copies | 83.083 | 70.859 | 102.070 | 0.85 |
| main-edit-among-many-shapes | 4.837 | 3.322 | 3.654 | 0.69 |
| main-edit-into-10-copies | 1.79 | 1.200 | 1.572 | 0.67 |
| main-edit-into-100-copies | 15.415 | 11.457 | 18.418 | 0.74 |
| main-edit-into-1000-copies | 153.086 | 109.634 | 125.910 | 0.72 |
| variant-switch-into-copies | 32.052 | 30.110 | 35.890 | 0.94 |
AI-assisted-by: claude-opus-5-5
The clipboard's paste maps over the fills and strokes of every shape it
adds, so a shape without them, such as a group, gets empty lists. The
harness paste left them out, and the frontend replay of a generated
paste differed from the JVM run by them.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
A grid that hugs its content along an axis took its percent tracks
there as percentages of its current size, which the hug then changed,
so the grid settled only over several reflows, in both engines. Along an
axis the grid hugs, percent tracks are now sized by their content, as
CSS does for a container without a definite size, and the grid settles
in one reflow.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 22.420 | 23.267 | 0.64 |
| edit-through-a-chain-of-dependents | 68.931 | 42.715 | 52.079 | 0.62 |
| library-update-into-copies | 83.083 | 59.152 | 66.682 | 0.71 |
| main-edit-among-many-shapes | 4.837 | 3.469 | 3.887 | 0.72 |
| main-edit-into-10-copies | 1.79 | 1.119 | 1.416 | 0.63 |
| main-edit-into-100-copies | 15.415 | 10.790 | 12.449 | 0.70 |
| main-edit-into-1000-copies | 153.086 | 108.911 | 129.565 | 0.71 |
| variant-switch-into-copies | 32.052 | 23.968 | 26.842 | 0.75 |
AI-assisted-by: claude-opus-5-5
Ungrouping in a grid main brings the group's shapes into each copy as
well. A copy with its own cells gave them its free cells in the order
the main's grid flows them, while sync gives such a copy's new children
its free cells in the main's layer order, so the result depended on the
path. The ungroup now uses the main's layer order too.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 22.979 | 26.978 | 0.65 |
| edit-through-a-chain-of-dependents | 68.931 | 43.948 | 50.432 | 0.64 |
| library-update-into-copies | 83.083 | 51.875 | 58.852 | 0.62 |
| main-edit-among-many-shapes | 4.837 | 2.911 | 3.140 | 0.60 |
| main-edit-into-10-copies | 1.79 | 1.160 | 1.245 | 0.65 |
| main-edit-into-100-copies | 15.415 | 11.340 | 11.833 | 0.74 |
| main-edit-into-1000-copies | 153.086 | 110.971 | 121.066 | 0.72 |
| variant-switch-into-copies | 32.052 | 24.935 | 29.875 | 0.78 |
AI-assisted-by: claude-opus-5-5
A copy duplicated in place under a stale library differs from its
updated main, and the check that keeps a pasted copy's look marked its
geometry although its source had none marked. The check now applies only
to a copy that lands under a parent other than its source's. The harness
model authors a pasted copy's geometry where it lands when its source
authored it, and when another parent gave it its orientation or size.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 21.980 | 22.894 | 0.62 |
| edit-through-a-chain-of-dependents | 68.931 | 43.830 | 44.755 | 0.64 |
| library-update-into-copies | 83.083 | 53.028 | 54.143 | 0.64 |
| main-edit-among-many-shapes | 4.837 | 2.352 | 2.487 | 0.49 |
| main-edit-into-10-copies | 1.79 | 1.103 | 1.125 | 0.62 |
| main-edit-into-100-copies | 15.415 | 10.851 | 11.191 | 0.70 |
| main-edit-into-1000-copies | 153.086 | 108.036 | 112.301 | 0.71 |
| variant-switch-into-copies | 32.052 | 22.906 | 23.847 | 0.71 |
AI-assisted-by: claude-opus-5-5
A copy shape that authored its place in a plain board, then taken into a
grid board by an ungroup of its main, is placed by that grid. The model
still compared the place it had authored. A shape a layout places is no
longer held to an authored place.
Harness only; the benchmark is not run.
Also in this commit:
- The harness changes that check the next commit, "Mark a pasted copy's geometry
only when it changes parent".
AI-assisted-by: claude-opus-5-5
Switching a swapped head back to the variant its main holds syncs it from
the copy's root, which misses a move the copy's layout gave its parent:
a grid copy with its own gap moves its child boards, and the restored
head landed where the board would be without that gap. The restored head
now takes the place its main shape has in its actual parent, unless it
authors its geometry or a layout places it.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 23.030 | 27.413 | 0.65 |
| edit-through-a-chain-of-dependents | 68.931 | 42.850 | 57.976 | 0.62 |
| library-update-into-copies | 83.083 | 60.905 | 63.778 | 0.73 |
| main-edit-among-many-shapes | 4.837 | 3.209 | 4.225 | 0.66 |
| main-edit-into-10-copies | 1.79 | 1.162 | 1.511 | 0.65 |
| main-edit-into-100-copies | 15.415 | 11.296 | 12.963 | 0.73 |
| main-edit-into-1000-copies | 153.086 | 103.929 | 116.557 | 0.68 |
| variant-switch-into-copies | 32.052 | 22.613 | 25.676 | 0.71 |
AI-assisted-by: claude-opus-5-5
A nested copy inside a flipped or turned main shows the orientation its
parent gives it. Copied and pasted elsewhere, it kept that orientation
without marking its geometry, so the next sync turned it back to its
main's with no edit. Each new top copy of a paste or duplicate now goes
through the check a detached copy does: when it is not oriented and
sized as its main turned like its new parent, its geometry is marked as
its own. The harness model authors the same geometry.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 22.546 | 22.997 | 0.64 |
| edit-through-a-chain-of-dependents | 68.931 | 42.115 | 43.255 | 0.61 |
| library-update-into-copies | 83.083 | 52.064 | 54.230 | 0.63 |
| main-edit-among-many-shapes | 4.837 | 2.409 | 2.585 | 0.50 |
| main-edit-into-10-copies | 1.79 | 1.072 | 1.139 | 0.60 |
| main-edit-into-100-copies | 15.415 | 10.340 | 10.963 | 0.67 |
| main-edit-into-1000-copies | 153.086 | 105.617 | 109.281 | 0.69 |
| variant-switch-into-copies | 32.052 | 22.217 | 22.754 | 0.69 |
AI-assisted-by: claude-opus-5-5
The CLJS grid centered a child horizontally with half its margins, as
CSS and the flex layout do, but vertically it shifted the child by the
whole top and bottom margins. Both axes now center the margin box.
Found by laying out random grid boards in a WASM workspace and reflowing
the same objects with the CLJS engine.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 21.021 | 21.658 | 0.60 |
| edit-through-a-chain-of-dependents | 68.931 | 41.085 | 42.996 | 0.60 |
| library-update-into-copies | 83.083 | 51.898 | 52.585 | 0.62 |
| main-edit-among-many-shapes | 4.837 | 2.255 | 2.395 | 0.47 |
| main-edit-into-10-copies | 1.79 | 1.031 | 1.069 | 0.58 |
| main-edit-into-100-copies | 15.415 | 9.877 | 10.192 | 0.64 |
| main-edit-into-1000-copies | 153.086 | 100.151 | 105.021 | 0.65 |
| variant-switch-into-copies | 32.052 | 22.340 | 23.371 | 0.70 |
AI-assisted-by: claude-opus-5-5
A paste kind joins the structure programs: the clipboard's copy of a top
copy, or of a shape a main holds, pasted at the page root. Nothing in
them leaves a copy level, so no ref is advanced. The model carries the
overrides of the copied subtree to the pasted one, as for a duplicate.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
Sync adds the copies of a main's new children before it removes those of
the children the main lost. Grouping two children of a full grid main
gave each grid copy one flow child more than its cells for a moment, so
it grew a column; the removals left that column empty and the next sync
dropped it. A copy that does not override its cells now takes its main's
tracks on each axis it does not override (all of them on a reset) before
it takes its main's cells.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 20.772 | 21.537 | 0.59 |
| edit-through-a-chain-of-dependents | 68.931 | 41.197 | 42.556 | 0.60 |
| library-update-into-copies | 83.083 | 49.581 | 51.122 | 0.60 |
| main-edit-among-many-shapes | 4.837 | 2.181 | 2.405 | 0.45 |
| main-edit-into-10-copies | 1.79 | 1.108 | 1.158 | 0.62 |
| main-edit-into-100-copies | 15.415 | 10.273 | 10.948 | 0.67 |
| main-edit-into-1000-copies | 153.086 | 100.399 | 109.973 | 0.66 |
| variant-switch-into-copies | 32.052 | 21.866 | 22.788 | 0.68 |
AI-assisted-by: claude-opus-5-5
Three kinds join the structure programs, and the programs built on them:
- move: a shape a main holds, or a nested copy head, moves to the page,
into a board or group of its own main or of another. The model gives
an override of its constraints the defaults of its new parent, as the
move resets them.
- rename: the assets panel's rename, with names that carry a path. The
component must take the normalized path and name, and its main root
the whole name.
- annotate: a component's annotation is set or removed. Only that edit,
or its undo or redo, may change an annotation.
The frontend interpreter realises the rename and the annotation through
their workspace events.
Harness only; the benchmark is not run.
Also in this commit:
- Ignore the layout-given size of a hidden layout child in the oracle. A hidden
shape leaves its layout's flow, so it keeps the size the layout last gave it
along its fill or auto axes, while a sync from its main can give it the
main's. The layout gives it its size again once it is shown. The oracle
expected the main's size, and the design snapshots of a settled and a fully
synced world differed by it. Those sizes are now nobody's to keep, as the
place of a hidden layout child already is.
AI-assisted-by: claude-opus-5-5
Moving a shape of a copy group that a layout places changes the group's
bounds, and the layout moves the group back: the moved shape ends where
it was and its siblings move the other way. Nothing was marked, so the
next sync put the siblings back and the user's move was lost.
When a transform moves some copy shapes of the outermost group or bool a
layout places and leaves others, every copy shape in that group now
authors its geometry. A transform of the whole group or of its board
marks nothing more. The harness model authors the same shapes, and
transforms of copy shapes in groups inside layouts are generated again.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 23.506 | 23.987 | 0.67 |
| edit-through-a-chain-of-dependents | 68.931 | 40.749 | 42.565 | 0.59 |
| library-update-into-copies | 83.083 | 49.609 | 50.752 | 0.60 |
| main-edit-among-many-shapes | 4.837 | 2.344 | 2.551 | 0.48 |
| main-edit-into-10-copies | 1.79 | 1.103 | 1.179 | 0.62 |
| main-edit-into-100-copies | 15.415 | 10.109 | 10.492 | 0.66 |
| main-edit-into-1000-copies | 153.086 | 103.527 | 107.994 | 0.68 |
| variant-switch-into-copies | 32.052 | 21.691 | 22.453 | 0.68 |
AI-assisted-by: claude-opus-5-5
An update main pushes a copy's grid direction into its main. When that
main is itself a grid copy with its own cells, sync orders its children
as they flow in those cells, but the update ordered them as the pushing
copy lists them. The next sync of that copy then reordered them with no
edit. The update now gives such a main the order of the copy's cell
flow.
Benchmark (lower quantile, ms):
| world | baseline | now | mean | ratio |
|------------------------------------+----------+---------+---------+-------|
| deepest-edit-through-three-levels | 35.215 | 21.632 | 22.076 | 0.61 |
| edit-through-a-chain-of-dependents | 68.931 | 42.236 | 44.386 | 0.61 |
| library-update-into-copies | 83.083 | 48.557 | 51.934 | 0.58 |
| main-edit-among-many-shapes | 4.837 | 2.443 | 2.556 | 0.51 |
| main-edit-into-10-copies | 1.79 | 1.129 | 1.149 | 0.63 |
| main-edit-into-100-copies | 15.415 | 10.085 | 10.387 | 0.65 |
| main-edit-into-1000-copies | 153.086 | 101.224 | 103.687 | 0.66 |
| variant-switch-into-copies | 32.052 | 20.846 | 21.588 | 0.65 |
AI-assisted-by: claude-opus-5-5
A copy bool set to fill in a grid cell is sized by the grid, and its
shapes scale with it. The oracle expected them at their main's size and
place. The shapes of a group or bool whose size a layout owns are now
sized and placed by that layout too.
Test: the-shapes-of-a-copy-bool-a-grid-fills-are-sized-with-it; shape
structure programs pass (seed 9008, 300).
Harness only; the benchmark is not run.
Also in this commit:
- Compare the corners of a masked group in any order on undo. A masked group
takes the corners of its mask. When the mask is a bool, a sync that refits it
computes them again from its children in any order, and the undo of that sync
refits it the same way. The undo check already compares the corners of a bool
in any order; a masked group now follows its mask.
- Read the size of a copy path from its bounds in the harness model. A path
stores no width or height, so the model saw no change when a copy path that a
layout places was resized: its place is the layout's and its size values stay
nil. The model now reads a size change of a shape with no width from its
bounds, and expects the resize to mark it.
AI-assisted-by: claude-opus-5-5
A group keeps its transform when it fits its children but clears its
flip flags; a bool keeps its flags while fitting it resets its
transform. Sync read the orientation of both from their transform, so a
bool in a copy whose main was flipped, resized afterwards, looked
unflipped, and a full sync unflipped its shapes, apart from the settled
file.
A bool's orientation now comes from its turn and flips, a group's from
its transform. The harness generates resizes of copy bools again.
Test: a-copy-bool-and-its-shapes-are-compared-by-their-corners, now
reached; group, sync, switch, golden master and generated group and bool
tests pass (376).
Benchmark (alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+--------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 21.34 | 27.387 | 0.61 |
| edit-through-a-chain-of-dependents | 68.931 | 46.931 | 67.865 | 0.68 |
| library-update-into-copies | 83.083 | 63.966 | 74.531 | 0.77 |
| main-edit-among-many-shapes | 4.837 | 2.42 | 2.623 | 0.5 |
| main-edit-into-10-copies | 1.79 | 1.122 | 1.169 | 0.63 |
| main-edit-into-100-copies | 15.415 | 10.589 | 11.406 | 0.69 |
| main-edit-into-1000-copies | 153.086 | 101.69 | 106.961 | 0.66 |
| variant-switch-into-copies | 32.052 | 21.962 | 26.736 | 0.69 |
AI-assisted-by: claude-opus-5-5
The two programs deferred during the variant runs pass on the current
branch with every step executed, after the round 4 fixes. Generated
tests keep them passing.
Tests: a-flipped-turned-copy-switches-without-losing-its-geometry,
a-turned-pushed-copy-made-a-variant-keeps-a-valid-file.
Harness only; the benchmark is not run.
Also in this commit:
- The harness changes that check the next commit, "Read a bool's orientation
from its turn and flips when syncing".
AI-assisted-by: claude-opus-5-5
The coverage control requires each kind to supply 0.5% of a program
set's changes. Grid cell and grid direction steps exist only in grid
worlds, a share of the programs, so with 50 programs a seed with few
grid worlds left them under the share, and the swap and library
programs failed on random seeds. These kinds must still change state,
but need no share of the changes.
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
A nested copy that fills its flex parent while its main is turned gets
its size from the layout. Switching it to a variant without that layout
carried the size as an override, as both mains agree on their size and
its copy chain authors a turn, so it kept the layout's size for good.
A dimension the layout of the copy gave it before the switch is no
override: the copy takes the new variant's.
Test: a-switch-carries-no-size-the-old-layout-gave-a-child.
Switch, swap, sync and golden master tests pass (215); variant programs
pass (seed 9007, 300).
Benchmark (alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+--------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 20.013 | 20.948 | 0.57 |
| edit-through-a-chain-of-dependents | 68.931 | 39.641 | 40.924 | 0.58 |
| library-update-into-copies | 83.083 | 50.865 | 52.074 | 0.61 |
| main-edit-among-many-shapes | 4.837 | 2.536 | 3.259 | 0.52 |
| main-edit-into-10-copies | 1.79 | 1.04 | 1.067 | 0.58 |
| main-edit-into-100-copies | 15.415 | 9.946 | 10.414 | 0.65 |
| main-edit-into-1000-copies | 153.086 | 99.227 | 101.039 | 0.65 |
| variant-switch-into-copies | 32.052 | 21.451 | 24.531 | 0.67 |
AI-assisted-by: claude-opus-5-5
A nested copy in a main that authors its height and scales vertically
has its height scaled when an update main resizes that main. The model
kept only the size as authored, as the copy's place in a plain main is
not an override, and compared it raw. An authored size now follows its
parent's later resizes through its constraints, though its place stays
unauthored.
Test: an-authored-size-follows-a-resize-of-its-main-through-its-constraint.
Structure, swap and scale programs pass (seed 9006, 150 each).
Harness only; the benchmark is not run.
AI-assisted-by: claude-opus-5-5
A grid copy that authors its cells gives its children without a cell
free cells at sync. It took them in the copy's order of its children,
which grid reordering changes, so a child given back by a shrunk track
and a child new from the main could take each other's cells. They now
take free cells in the order of the main shapes they refer to, as sync
adds children.
Test: a-grid-copy-gives-free-cells-to-children-in-their-main-order.
Grid, sync and golden master tests pass (399).
Benchmark (alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 21.19 | 22.83 | 0.6 |
| edit-through-a-chain-of-dependents | 68.931 | 42.408 | 44.279 | 0.62 |
| library-update-into-copies | 83.083 | 53.897 | 55.394 | 0.65 |
| main-edit-among-many-shapes | 4.837 | 2.367 | 2.548 | 0.49 |
| main-edit-into-10-copies | 1.79 | 1.031 | 1.103 | 0.58 |
| main-edit-into-100-copies | 15.415 | 10.483 | 10.835 | 0.68 |
| main-edit-into-1000-copies | 153.086 | 100.626 | 107.239 | 0.66 |
| variant-switch-into-copies | 32.052 | 23.355 | 23.926 | 0.73 |
AI-assisted-by: claude-opus-5-5
A grid copy that authors its cells and has no free cell for a child its
main gains grew a column for it at the first sync. The next sync gave
the copy its main's columns again, took the grown column for a shrunk
inherited track, and fitted the child back into the inherited tracks,
stacking it in another child's cell.
When a copy's inherited tracks shrink, its default cells past them
(automatic, single, aligned by default) give their children back, which
take a free slot or a track the grid grows for them, as when they first
came. Cells with something of their own are fitted as before, and the
overflow is no longer fitted back into the inherited tracks. The grid
model does the same, and the oracle accepts the default tracks a grid
grows along its flow past its main's.
Tests: a-grid-copy-with-its-own-cells-grows-a-column-for-a-new-child-for-good,
a-grid-copy-keeps-an-aligned-cell-its-main-takes-a-column-from,
a-grid-copy-keeps-its-aligned-cell-when-its-main-adds-a-row.
Grid, sync and golden master tests pass (315); grid programs pass
(seed 9004, 300).
Benchmark (alone; a first run under load read 1.09 for the variant
switch world, which F526 does not reach):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+---------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 23.132 | 28.537 | 0.66 |
| edit-through-a-chain-of-dependents | 68.931 | 47.053 | 57.855 | 0.68 |
| library-update-into-copies | 83.083 | 55.445 | 64.964 | 0.67 |
| main-edit-among-many-shapes | 4.837 | 3.117 | 3.305 | 0.64 |
| main-edit-into-10-copies | 1.79 | 1.169 | 1.544 | 0.65 |
| main-edit-into-100-copies | 15.415 | 13.442 | 14.626 | 0.87 |
| main-edit-into-1000-copies | 153.086 | 119.198 | 141.449 | 0.78 |
| variant-switch-into-copies | 32.052 | 25.091 | 29.622 | 0.78 |
AI-assisted-by: claude-opus-5-5
Whitespace only, as the formatter wants it after the last changes to
same-state? and the transform op of the generator.
Harness only; the benchmark is not run.
Also in this commit:
- Record the frontend golden master of case AY again. The undo and redo checks
compare the tokens library by its exported content, so the golden master
writes it key by key; the JVM golden master of case AY was recorded again for
that, and the frontend one is now too. Same values, written differently.
- The harness changes that check the next commit, "Grow a track for a grid
copy's new child and keep it at sync".
AI-assisted-by: claude-opus-5-5
A bool's stored flip flags and transform drift apart, so neither sync
nor the model can read its orientation. The model and the oracle now
compare a group or bool, and the shapes it holds, by their corners in
any order, without their turn and flips, whether or not the authored
values record corners. Their corners are ordered by angle around their
centre: an order by position to the hundredth of a pixel swapped the
corners of an empty bool, 0.01 px wide, with float noise.
Resizing a copy bool after its main was flipped leaves settled and
fully synced files apart (F518, deferred until a bool keeps a reliable
orientation); the generator skips resizes of copy bools until then.
Tests: a-copy-bool-and-its-shapes-are-compared-by-their-corners,
an-empty-bool-in-a-copy-keeps-its-corners-in-one-order.
Harness only; the benchmark is not run.
Also in this commit:
- Leave a group's flip flags out of the harness undo check. Undoing a text style
edit fits the group around the text again, and fitting a group clears its flip
flags by design, though its corners and transform still show the flip. The
undo check now leaves the flip flags of groups and bools out and compares
their corners.
- Round corner angles before they wrap in the harness corner order. The harness
orders the corners of groups, bools and their shapes by angle around their
centre. A corner level with the centre, as in a square turned 45 degrees, fell
at 0 or just under a full turn as noise fell, and its order flipped. Angles
are now rounded before they wrap.
- Read the size of a bool from its selrect in the update main check. The harness
check of an update main compared the sizes of the parent of a pushed shape by
their width and height, and crashed when that parent was a bool, which has no
width or height of its own. Sizes are read from the selrect.
- Order the corners of groups in the harness design snapshot. The settled and
fully synced files of a bool made in a variant main whose copies are turned
differed only by noise: up to 5e-6 px in the bool, which computes its geometry
in single precision. But the design snapshot ordered only a bool's corners,
not a group's, and the angle order, to the millionth of a radian, flipped a
corner level with the bool's centre under that noise.
- Author no size for a swapped head a layout places in the harness. Switching a
nested head in a filling grid copy to a wider variant grows the grid. The
model took the grown copy for a sync resize and made the untouched swapped
head author the size the resize gives it, though the layout sizes it, not its
constraints. Heads a layout places are left out.
AI-assisted-by: claude-opus-5-5
The undo of a group's synced attributes fitted the group to its
children again, after its own undo, so a turned group would not keep
the turned size an earlier fit gave it. But fitting a group clears its
flip flags, so a flipped group came back unflipped.
The undo of a group or bool whose geometry a sync changes restores all
of it, its place, size, corners, transform, turn and flips, whatever
fits of it ran before.
Tests: a-copy-group-whose-main-turns-matches-its-corners-whatever-its-flips,
undoing-an-update-main-that-turns-a-group-gives-the-group-its-size-back.
Benchmark (alone):
| :world | :baseline | :now | :mean | :ratio |
|------------------------------------+-----------+--------+---------+--------|
| deepest-edit-through-three-levels | 35.215 | 20.394 | 21.255 | 0.58 |
| edit-through-a-chain-of-dependents | 68.931 | 43.583 | 44.351 | 0.63 |
| library-update-into-copies | 83.083 | 51.967 | 52.683 | 0.63 |
| main-edit-among-many-shapes | 4.837 | 2.366 | 2.526 | 0.49 |
| main-edit-into-10-copies | 1.79 | 1.044 | 1.132 | 0.58 |
| main-edit-into-100-copies | 15.415 | 9.805 | 10.146 | 0.64 |
| main-edit-into-1000-copies | 153.086 | 98.202 | 101.844 | 0.64 |
| variant-switch-into-copies | 32.052 | 21.832 | 22.268 | 0.68 |
AI-assisted-by: claude-opus-5-5