Thanks @el-j — merging.
**Summary:** Pathfinder Trim/Crop (and Merge) now build on `each_classified_face` (generalizing `each_face`) so classification and face construction share one batched pass, dropping the duplicate `batched_classified_contours` after #4375 as @bflatastic asked. Differential tests lock subpath order and the fully-covered shape case.
Floating panels are OS windows of their own, and craft-ui keeps a palette above the
app's other windows while it is active. A modal dialog or an open menu is drawn in
the main window, so a floating panel covered it: the Colour Picker's OK button and
fields, Recolor Artwork, and the menu bar's drop-downs (#5238, defect 1).
While a modal dialog or a menu of the in-window menu bar is open the docking renderer
hides each palette window, and when neither is open each palette shows again at its
palette level: topmost while the app is active, behind other applications' windows
while it is not, as craft-ui leaves a palette it creates. Hiding rather than lowering
covers both: a lowered window can be clicked back to the front, and craft-ui varies
the level through the viewport builder, which egui applies when the window opens and
not to one already on screen, so the commands go explicitly to the windows there are.
The menu bar tells the docking renderer whether one of its top-level menus is open,
the same state its hover switching uses; egui's own "any popup open" misses the
in-window menu bar.
Measured on Windows with the app focused and panels floating: before, a palette was at
the front of the z-order and topmost over the main window with the Colour Picker open;
after, the main window is in front and the palette is hidden until the dialog or menu
closes.
Tests: a_modal_lowers_the_native_palette_windows and
a_menu_of_the_menu_bar_lowers_the_palette_windows check the commands sent, and
a_palette_window_keeps_out_of_the_way_of_the_main_windows_dialog_and_menus the level
and visibility table. The wheel test
`wheel_over_a_floating_character_font_field_changes_the_font` fails on this machine
without this change too (font-order dependent), so it is not part of it.
Each gradient definition that no swatch names took its name by trying
`Gradient`, `Gradient 2`, `Gradient 3`, ... from the start, so a document
with many such gradients tried every earlier name again for each new
one. The writer now keeps the number the next search starts from, and
still skips names that a gradient, swatch or pattern already has.
- A test names 1,000 gradients around two taken names and checks that
each name is tried once (1,002 tries; 502,496 when each search starts
over).
- An export test writes 300 gradients next to a `Gradient 3` swatch and
a `Gradient 7` pattern, and checks the names in order and the same
output when written again. It passes with the old search too.
The editing copy of a PDF-compatible .ai wrote every image as its outline
filled with its average color. Embedded images are now written with their
pixels, and linked images as placed files with a preview:
- Embedded images are rasters in the form of the version 7 specification
(section 8, https://13thmonkey.org/documentation/fonts/AI7FileFormat.pdf):
%AI5_BeginRaster, the color space, the image matrix and the XI operands
of current files, then the samples in binary in a %%BeginData section.
Gray, RGB and CMYK images keep their samples (CMYK images their ink
amounts), and see-through images their alpha channel after the color
samples. The matrix is written to ten decimals.
- Linked images are placed files (section 16): the matrix, the box of the
file's pixels and its path, %%IncludeFile, the folder relative to the
document's (when the relative path ends in the path's file name), and a
preview of at most 256 pixels a side.
- Images that can't be decoded or have more than 33,554,432 pixels (the
most the reader reads) are written as their outlines filled mid gray,
images of 16 bits a channel (CMYK TIFFs included) at 8 bits, and linked
images in opacity masks and pattern tiles as embedded images, each with
a warning.
Each of these is written in the form VectorCraft's .ai reader reads, and
the tests read the written data back through that reader.
The reader now gives an image its name and opacity mask when the
operators that end its raster (XH, Xr and the N after %AI5_EndRaster) come
before its dictionary, reads an empty relative folder as the document's
own folder, and looks for the %%EndData that ends an image's samples
after the alpha channel its XI operands give it, preferring one whose next
line is XH, so samples that hold the marker's bytes are read whole.
Tests: the round trip of every kind of image and its state, a second write
of what is read back giving the same bytes, the raster and placed-file
forms, previews of large linked images, and the engine's editing copy with
every save option; a fuzz case for damaged images and placed files.
Docs: docs/file-format-parity.md.
In the legacy format, area type whose story flows through several linked
areas has one text path (`Tp` ... `TP`) for each area, each with its own
matrix and frame (version 7 specification, sections 15 and 15.5.3). The
reader kept only the first frame, placed it with the text object's last
valid matrix, and put the whole story in that one frame.
Each linked area with a frame now opens as its own area type object,
placed with its own matrix, and the frames are threaded in file order.
The first frame holds the story: the characters after every `TP`, and
the overflow text. When the file opens, the story is laid out again
through the thread (section 15.1 lists the flow from area to area as
final-form information). A linked area without a frame is left out.
Area type with one area opens as before.
The dictionary that follows a text object (its opacity mask, text wrap,
name and other keys) went to its one object. Its opacity mask and text
wrap now go on every frame, and its other keys on the first frame.
Point type is read with its first text path's matrix, and type on a
path with the matrix and start of the first text path that has a path.
Area type with no frame at all opens as point type with its first text
path's matrix. Before, these used the last valid matrix and, for type
on a path, the last start.
One text object reads at most 4,096 text paths. The number of frames
times the story's character count is kept to 4,194,304, so a longer
story keeps fewer frames. Frames past either limit are left out, with a
warning.
A ui.pointer request made before synthetic input was dropped counted
toward the per-frame request limit even when its caller had already
given it up and no reply was sent. A backlog of such requests could hold
live requests back for a frame. It now counts only when the app claims
it and replies with the error, as the other given-up requests are
skipped.
A legacy .ai in the version 7 form names a linked image's file with
`(path) modified XG` and draws the image with `XF`, which takes the
operands `XI` takes and has no samples after it (Adobe Illustrator File
Format Specification 7.0, sections 8.2 and 8.3). The reader ignored both
operators, so the image was dropped without a warning.
- `XF` after `XG` becomes a linked image of that file, its pixels placed
as an `XI` image of the same operands is placed. The file keeps no
pixels of it, so a placeholder stands in until the file is relinked in
the Links panel, with a note in the open result. No pixels are read,
so the image is not held to the pixel limit of `XI`'s samples.
- An `XF` without a file name before it, or with operands `XI` couldn't
read (no matrix, too few numbers, no pixels, or samples of other than
1 or 8 bits), is left out with the existing "an image VectorCraft
can't read was left out" warning. An `XG` applies only to the image
after it in the same raster object; `() 1 XG` before an embedded `XI`
keeps that image embedded.
- A PostScript page draws nothing for `XF`, so there the placeholder's
area is left out of the comparison of the layers with the page, and
the file keeps its layers. A PDF part draws the image's pixels, so a
file with one is compared with it there as before, and its layers are
not used without the PDF part.
- Linked images of different files get image keys of their own even
when they show the same pixels (the placeholder, or two identical
placed-file previews), so updating or exporting one file's image
reads that file only.
- `vectorcraft_doc::links::missing_placeholder` draws the placeholder: a
light tile, crossed out, as SVG import shows for a missing file.
Tests: reader tests for the placement, the image keys and the cases left
out, an end-to-end test of a legacy file with a linked image, a test of
a file with a PDF part, and a fuzz case for damaged files with `XG` and
`XF`. The new note has a row in each catalog that translates the
placed-file note.
First slice of https://github.com/storytold/vectorcraft/pull/1281, split
as the review asked. `--deep` gains rows for hovering, a hit test that
misses, marquee, smart guide targets and shift-click on a large document,
so the later slices have a baseline.
PathData::bounds built a kurbo path on every call. It now measures the
segments in place, with the same guards, and a proptest compares the
result bit for bit with what the path gave.
Co-Authored-By: arhammusheer <arhammusheer@users.noreply.github.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Type in the text document of a .ai or EPS file's editing data that names
no character or paragraph style, a story without style runs included, is
set in [Normal Character Style] and [Normal Paragraph Style]. The reader
took what such type leaves out from the text document's base attributes
instead, or built-in defaults for a story without paragraph runs, and it
did not read paragraph styles.
A Large Canvas file (1:10) holds the normal styles' lengths at the art's
scale (1.2 for 12 pt) and the base attributes' at the page's (12). The
reader scaled the base attributes' 12 by 10, so the type came in at
120 pt. In an EPS the page comparison then rejected the layers, and the
file opened as its printed page without its linked images.
In any file whose normal styles differ from the base attributes, such
type now takes what the normal styles set, such as font, size, colors,
stroke, alignment, indents, space before and after, and hyphenation.
`.ai` and EPS editing-data import now reads object tags
(`/identifier (string) XT`) and applies each one to the object before it,
whatever kind of object it is. The `AdobeURL` tag sets the object's URL.
Every other tag is added to the object's data in order, up to 64 per
object. A tag with no object right before it, with an empty identifier,
or with an identifier or string longer than 64 KiB is ignored, and so
are tags in a pattern's or brush's art. A clipping path keeps its own
tags; the copy that paints its fill gets none.
Before, a tag on art that shows was an operator the reader did not know,
so it stopped the editing data from being read, and brush art holding a
tag was not read, so the brushes using it were left out.
An image's tags follow the `XH` that ends its data, and the state written
after a group or a gradient mesh is still theirs when their tags come
first. When a drawn look stands for an object whose style is not found,
the look keeps the object's note, URL and data, and a live blend keeps
those of its drawn steps.
`sync_views` matched views to documents by uid only while the number of
documents was unchanged. A `command.batch` (or a recording) that closed a
document fell back to matching by position, so the documents after the closed
one took their neighbours' views, and a move later in the same batch was not
applied to the views. The views are now matched by uid whenever the documents
change: a closed document's view is dropped and a new document gets its saved
view. Closing from the UI no longer removes the view by hand.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The editing copy for other apps now writes linear and radial gradients and
grid patterns as themselves instead of as average colors and cut tiles.
- Gradients the art uses are defined in the setup (`n Bn`, then
`%AI5_BeginGradient` … `Bd`, the stops as `Bs` with their opacity and
midpoints, `BD`), and gradient swatches nothing uses in a non-printing
block. Stops keep gray, CMYK, RGB and spot colors at a tint. A
gradient that isn't a swatch gets a name that no swatch or pattern of
the document has.
- Each fill or stroke with a gradient is an instance around its painting
operator (`Bb`, `Bh`, `Bg`, `Xm`, `Bm`, `Bc`, `BB`), on a compound
path's last subpath and on clipping paths too. `Bg` places the gradient
relative to the path: a linear one by its angle and its start and length
as fractions of the path's extent along it, a radial one by the matrix
that takes the unit circle onto its extent ellipse, its center as
fractions of the path's bounds in that circle's space and its radius as
a multiple of the one the center's distances to the right side and to
the bottom give. In a pattern's tile the placement is mirrored top to
bottom.
- Patterns whose tiles make a grid and hold their art are defined with
`%AI3_BeginPattern` … `E`, the tile's box first, and fills and strokes
paint with `p` and `P`; gradient and pattern swatches are in the palette.
- Freeform gradients, gradients along or across a stroke and gradients
without stops, with more than 256 stops or without extent are written
as their average color, and other patterns as their tiles cut to the
shape, each with a warning. Gradients that dither or mix perceptually
are written as classic ones, stops linked to global process colors as
their colors, and pattern swatches without the placement that moves,
turns or scales their pattern, with warnings.
- Reading: an unpainted box that a pattern's definition marks as its
tile's box (`AIPattern_Editor_Backing_Tile_Rect`) is no longer read as
art, and a linear gradient whose stops all lie at one place keeps each
stop on its side.
The definitions, instances and patterns follow sections 3.4, 3.5 and 14
of the version 7 specification
(http://mercator.elte.hu/~saman/hu/okt/AI7FileFormat.pdf). The stop
opacity, `Bg`'s placement and the marked tile box follow the newer form
that current .ai files carry, which that specification predates.
Each of these is written in the form VectorCraft's .ai reader reads, and
the tests read the written data back through that reader.
Docs: docs/file-format-parity.md.
The editing copy of the art now writes a guide as its path followed by
`(N) *`, or `(n) *` when the path is closed, without the `(N) *` before
the path. The legacy version 7 .ai keeps both, as section 6 of the
version 7 specification shows them
(http://mercator.elte.hu/~saman/hu/okt/AI7FileFormat.pdf).
A ui.pointer request waits, ahead of the requests after it, until the
synthetic input sent before it has been applied. When its caller gave
it up during that wait, the request still waited there, and the
requests behind it waited for up to 1024 queued input events and could
time out.
A given-up pointer request is now dropped at once instead of waiting.
While Select All holds a threaded story, a paragraph whose text starts in one of the
story's locked or hidden frames is no longer converted, and the empty parts of a converted
paragraph in such a frame are no longer restyled. Before, a locked frame whose first
paragraph took in the previous frame's empty last part was converted, and so was a
paragraph that ran on from a locked first frame into an editable one.
After Select All in a frame of threaded text, Remove Line Breaks and
Remove Paragraph Breaks changed only the edited frame, and Convert to
Text over MCP converted only the selected frame. Both now act on the
text of every frame of the story as one undo step.
- addon.cleanup.removeBreaks reads the frames of the story as one text,
so a break at the end of a frame before a word in the next one becomes
a space. The same applies when every frame of a story is selected as
objects.
- text.convertListToText without ids, id or a range converts every
paragraph of the story. Locked and hidden frames of the selected story
are not converted, even when ids names them. When the call converts
the whole story, the story stays selected; before, the call reset the
Type tool's selection to the edited frame.
- A selection inside one frame, and unthreaded type, behave as before.
Docs: docs/mcp.md, docs/target-app-parity.md.
`the_fallback_glyph_of_a_consonant_after_its_vowel_sign_is_reported` expected
the fallback face's cmap glyph for the vowel sign I. On the Ubuntu runner the
fallback face is Noto Sans Devanagari, whose shaping draws the sign with
another glyph (545 instead of 32), so the test failed on `main`. It now checks
that two glyphs are drawn in Ka's cluster, the first is not Ka's and the
second is, and keeps the `text.glyphAlternates` checks.
Fixes#5750
push_num formatted every number with write!("{:.*}"), which runs the
exact decimal float formatter, then scanned and trimmed the text. That
formatter was about 95% of the time to export a path of 200,000
anchors.
At 9 decimals or fewer, push_num now scales the value by 10^decimals
and writes the digits of the rounded integer when that rounding is
certain: the scaled value is below 2^53 and further from a half than
the scaling's rounding error. Ties and larger values still go through
the formatter, so the output is the same byte for byte.
Exporting one path of 200,000 anchors takes a median of 43.4 ms
instead of 167.0 ms, and 20,000 rectangles 81.4 ms instead of 141.4 ms
(Windows 11, release build).
A `backquote` given as a string or a number, on the request or on an
event, was read as absent, so `"true"` moved the object's geometry
instead of its pattern. `pointer_gesture` and `ui.pointer`, headless
and desktop, now answer with an error before any event runs.
Where two open paths lie along the same stretch, the planar map gave
that stretch to one of them only, so a merge or an erase across it left
the other path in place there. Every other open path along an edge now
has its own piece in `lines_under`, under a closed path's outline or
along a line in front. `lines`, `edges` and Live Paint's edges stay one
per stretch, and an Alt-click there still takes the piece in front.
Two seeded tests run 300 random edits each on a long path and a long
compound path: anchors added and deleted at the start, middle and end of a
subpath and where a run of segments ends, subpaths closed and opened, anchors
and handles moved, and in the compound path members added and removed and
subpaths added to a member. The subpaths range from 2 anchors to 3,000, open
and closed. After each edit the kept path's elements, subpaths, bounds and
miter tips equal those of a new `KeptPath`, its runs hold each subpath's
segments, and each run's box and tips equal those measured again.
The index arithmetic in `Sub::spliced`, `Sub::resize` and `KeptPath::update`
uses `checked_sub` and `checked_add` instead of plain subtraction. When the
segments or runs kept for a subpath do not fit its anchors before the edit,
`resize` returns `None` without changing anything, and `update` writes the
subpath's elements again from the new anchors and measures it whole, as it
does for a subpath of fewer than two anchors. A compound path whose kept
members do not pair with the new ones returns false, and the renderer makes
the path again whole.
When anchors of a kept path (a path or compound path of 1,024 anchors or
more) are added or deleted, or a subpath is closed or opened, the next
frame writes again only the path elements from the first anchor that
changed to the last, and the closing segment and ClosePath when an end
anchor changed or the subpath was closed or opened. The elements after
the change move with it. Before, such an edit made the whole path again
and measured it whole.
- The runs of segments now carry their own ends. The runs that held the
segments written become one run, split into runs of 128 when it grows
past 256, and only those runs are measured again, with the miter tips
of the run of the segment with length before each span. When a
subpath is closed or opened and the stroke reaches farther from one
than from the other, all of that subpath's tips are measured again.
- On a subpath that is open before and after the edit, the handle before
the first anchor and the one after the last are not compared, so
Delete Anchor Point, which clears them, writes again only the segments
around the anchor deleted.
- A subpath that has fewer than two anchors before or after the edit is
written again whole. A compound path member that gains or loses
subpaths is made again. When members are added or removed, those
members and every member between the first and the last of them are
made again, and the members before and after them stay.
- An edit that adds or deletes anchors in, closes or opens more than 16
subpaths makes the whole path again, as before.
- An edit that leaves every anchor as it was, such as a new stroke,
leaves the kept path as it is.
The pixels and the cull bounds are the same as with a path made again.
Docs: docs/target-app-parity.md.
Doc comments for the Select add-on commands, the Select by Link Status
dialog body, the Links panel menu and the test helpers (review: docstring
coverage).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Selection Tools add-on: addon.select.byLinkStatus selects the image
objects and placed documents whose link is linked, embedded, missing or
modified (one status or several), as links.list reports it. Optional ids
limit the search, add adds to the selection, and zoom runs the existing
Zoom to Selection in the desktop app. Objects that are hidden, locked, on
a template layer or outside the isolated group are left out and returned
with the reason. Only the selection changes.
Select Linked / Embedded / Missing / Modified are in Object > Plug-ins >
Selection Tools and the Links panel menu. Each is off when no image has
that status, and Shift adds to the selection. Select by Link Status...
lists the images with name, status and format, with a name search, a
status filter, checkboxes with All / None and Zoom to Selection. The
status bar says what was selected and how many were left out, with
plural rows in every catalog.
Tests: engine (statuses, ids, add, zoom, skip reasons, bad params),
command sweep, and UI (message, list filtering, menu items, Links panel
menu, headless dialog frame). Docs: mcp, target-app-parity, ui-parity,
roadmap.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Layers and Links panels share one thumbnail thread across documents,
and it ran its jobs in the order they came. Rows of a document left with
a long queue held up the thumbnails of the next document shown.
The thumbnail thread can now drop the jobs it has not started for some
slots. Each frame the Layers and Links panels name the document they
show; when it changes, the queued jobs of every other document are
dropped, and when no document is open, all of them are. Those rows ask
for their thumbnails again when their document shows again, and the
thumbnails already made are kept. The Navigator and the arrange panes do
not use the new call.
With 150 rows of 20,000-anchor paths queued for one document, the 8 rows
of such paths of the next document shown get their thumbnails in 153 ms
instead of 3,265 ms (debug test build, median of three runs).
A move also moves where a stroke's gradient is placed, but the slices of a
gradient along or across the stroke take only its stops and kind. The
moved stroke is now compared with the stroke the slices were made for
with the gradient's placement left out, so a placed gradient no longer
makes its slices again after every move.
A path whose stroke has a gradient along or across it kept its slices under
the node's address, so every nudge made the slices again for the whole path,
although its stroke outline was already reused. The renderer now keeps the key
of a stroke's latest slices under the object's id, the stroke's place and the
tolerance level, and when the object has only moved (same stroke and gradient,
same fill rule, anchors moved by one offset) it draws those slices moved by the
offset, translating only the slices that reach the piece being clipped. Scale,
rotate, an anchor edit or a change to the stroke or its gradient makes the
slices again. The pixels are the same.
In remote mode, press_key queues its key for the next UI pass, while
pointer_gesture drives the tool as soon as the app reads the request. A
release sent right after a key could commit the drag before the key
reached the tool, so C pressed while dragging out an arc left it open.
A ui.pointer request now waits, ahead of the requests after it, until
the synthetic input sent before it has been applied.
Review of #5694 asked for two changes.
Add-on, not a core tool, as Enclosing Lasso is:
- The command is addon.contour.set, in an add-on module of its own
(crates/engine/src/cmd/addons/contour.rs) with its tests beside it.
- The tool is listed under Object > Plug-ins > Contour and under its own
"Contour" heading in All Tools. It has no Basic or Advanced toolbar
slot and no default shortcut. The Contour bar is unchanged.
- The add-on drawer test lists Contour with the other add-on tools.
- Docs (mcp.md, the parity documents, ROADMAP) say add-on and use the
new command name; mcp.md gains a Contour add-on section.
Translations: rows for "Contour", "Contour Tool", "Sharp", "Bake" and
"Turn the Contour into plain paths" in all 26 catalogs ("Round" and
"Square" already had them). Drafted from each catalog's own vocabulary
(for Bake, the verb it uses for Expand Appearance; a word other than the
catalog's Stroke or Path word for the tool's name where they collide);
they are not native-speaker reviewed.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Two fixes from review of #5694, each with a test that fails without it:
- Dragging a selected Contour sent the tool's own corner options, so a
Round or Square Contour came back Sharp and lost its miter limit, and
the Contour bar (which reads the effect) disagreed with the shape. The
press now reads the Contour's own corners and miter limit and every
preview of that drag carries them; a new Contour still takes the tool's
options.
- A drag that never moved the distance (along the edge, or rounded to
nothing by Shift) previewed offset 0 and committed an unchanged copy
and an undo step. Nothing is previewed until the distance leaves where
it started.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Add a Contour tool (Modify group). Dragging away from a selected path or
shape's edge makes a copy of it that carries a live Offset Path effect,
grown by the drag; dragging toward the inside shrinks it. The result is
one undo step and stays editable until baked with Expand Appearance. A
second drag on a selected Contour changes its distance instead of
nesting another effect.
- object.contour: the command behind it (offset, sharp/round/square
corners, miter limit), usable over MCP and the control channel too.
- Contour tool: live preview, Shift rounds to whole points, Escape
cancels, a dashed guide with a white halo shows the outline the shape
had, and a pill shows the distance.
- Contour bar: a floating bar above the selection with the offset box,
Sharp / Round / Square and Bake (Expand Appearance).
- Toolbar slot under Modify (Basic and Advanced) and dc-contour.svg, an
original icon with its ASSETS.md row.
- Tests for the command, the tool's pointer gestures and the bar; the
border-with-a-gap recipe (two Contours, bake, Minus Front) is covered
end to end.
- Docs: beyond-Illustrator row, revision histories and the ROADMAP log.
Beyond Illustrator: a VectorCraft design built on Offset Path. Scores and
estimates unchanged.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The Control bar asks whether Select > Start Global Edit can run every
frame it shows a one-object selection. For a path, compound path or
symbol instance that check built the object's form: a copy of every
anchor and handle, normalized into a unit box. On a 200,000-anchor path
it took about 2.2 to 3.0 ms a frame.
The check now walks the object the same way and measures only the box
of those points, then applies the same size test, so it gives the same
answer for every object without allocating. Starting Global Edit still
builds the forms it compares.
A document tab dragged sideways along the tab strip moves its document to the
place an insertion line marks; the current document stays current. Dragged
above or below the strip, the tab still tears off into a floating window. The
views Window > New Window opened move with their document, and each document
keeps its own zoom, centre and artboard.
The move is the engine command `document.move {from?, to}`, so the control
channel, MCP and recordings reach it, and a replay moves the document before
the steps recorded after it. `sync_views` keeps the views with their documents
by uid when the documents change order.
Behaviour: https://helpx.adobe.com/illustrator/using/customizing-workspace.html
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The check that reuses a moved path's stroke outline allowed each
coordinate to differ from the moved one by 4 x f64::EPSILON times the
sum of the compared coordinates and the largest coordinate of the first
point, about 2.7 pt near 1e15, so a 0.5 pt anchor or handle edit there
drew the outline made before the edit.
offset_between and anchors_moved now measure the offset on each axis at
the pair of points with the smallest coordinates on that axis, then
check every pair against that offset within 3 units in the last place
of the pair's own coordinates. A first anchor far from the origin no
longer sets the tolerance for points near it.
Tests: 0.5 pt anchor and handle edits at 1e9 and 1e15 (and 0.0005 pt
at 1e9); an edit near the origin of a path that starts at 1e16; a path
whose coordinates shrink with each point, moved by growing amounts.
dotCooCoo