From 1c42d79e1376e0cff511411321e0caf7c9314868 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 11 Oct 2026 11:26:07 +0000 Subject: [PATCH 1/3] web: the landing page's "What's new" is 0.3.0's Six cards from the 0.3.0 changelog, with its figures: fidelity against 0.2.7, JPEG halos and parameters on the noisy tier, faces completed behind their covers and counters cut out, strokes by default, nodes at extrema and --editability, and the Lean theory with its generated checkers. The 0.2.0 cards are in that release's changelog. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_016AZWMqqHhRQ337N5EWhNC6 --- web/index.html | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/web/index.html b/web/index.html index c5472c7c..fcb7efbc 100644 --- a/web/index.html +++ b/web/index.html @@ -209,15 +209,15 @@

Exact vectors from your logo, in your browser.

-

What's new in 0.2.0

-

The Studio grew from a tracer with a report into an editor for the whole job, and it now runs in the browser as well as on the desktop.

+

What's new in 0.3.0

+

Traces come out closer to what the artist drew, and much closer on the resized JPEGs and screenshots logos usually arrive as. Against 0.2.7 on the 246 icons of the regression gate, every gated measure is better or level, in every mode and condition.

-
Palette

Colour groups, gradients included

Drag one ink onto another and they are drawn as one: the shapes between them join instead of keeping a traced seam. A group can become a chosen colour or one of its members, and a gradient member is extended over the whole group. Suggestions for near-duplicate inks are offered, never applied unasked.

-
Opening an image

Auto, or a short Custom wizard

Every image is traced at once. A card offers to keep that, or to walk through what the image is, clean-up, colours, detail and shape, with a live preview of each choice beside Auto's trace.

-
Fabricate

Ready for cutters and lasers

Layered vinyl with registration marks, inlay, print-then-cut stickers, stencils with bridges, and single-line paths for pens, scoring blades and lasers, in millimetres. A preflight shows what will not cut before anything is cut. SVG, DXF with true arcs, and GRBL G-code.

-
Honesty

Certainty bands

Where each traced boundary could be, measured from the pixels, drawn over the trace: sure, soft and unsure edges at a glance. On a clean 128-px emoji the median edge uncertainty is 0.09 px; the same image as a JPEG at quality 20 reads 0.50 px.

-
Transparency

Clear ground stays clear

Transparent images are traced as they are, and the clear ground no longer takes a slot from a colour cap: on ten emoji at four colours, mean dE00 went from 0.457 to 0.186. Side-by-side fills no longer let a hairline of background through.

-
Speed

Faster, with identical output

The curve fit and the ring repair run in parallel blocks, the palette and the boundary solve do only the work they read, and the Studio measures in parallel: on large images the slowest stages take a third to 70% less time, every change checked byte for byte against the build before.

+
Fidelity

Closer to the artist's file

Quality's colour difference from the artist's own file (dE00) is 12.5% lower at 128 px, 12.1% at 512 px and 15.2% at 512 px on an opaque ground; Fast's is 9 to 12% lower. Quality also writes fewer parameters: 1.32 per artist parameter at 512 px opaque, down from 1.96.

+
Noisy input

JPEGs and screenshots

A JPEG's colour bleeds two or three pixels across every edge. Those halos now go back to the inks they lie between instead of becoming shapes of their own. On icons resized and saved at quality 80, Quality's dE00 is 10.4% lower and Fast writes 4.13 parameters per artist parameter, down from 7.29.

+
Shapes

Whole shapes, under what covers them

A shape partly hidden by another used to come back as the sliver that shows. Quality now draws it whole: the simplest shape that leaves every visible pixel as it was. A letter's counter is cut out of its glyph instead of painted over it, and each fill is one compound path.

+
Strokes

Strokes by default

Where a shape is a centreline and one width, Quality writes it as a stroked path whenever that is the shorter description at as good a fit. It had been opt-in; it now costs about a sixth more time on a 512 px trace.

+
Editing

Nodes where a designer puts them

Nodes sit at a curve's horizontal and vertical extremes, with the handles on the axis, where the fit prefers them. With editability on, a circular arc becomes cubics cut at its extremes and the joins are smooth, the way drawing tools write curves.

+
Proofs

Proved, then checked

A pixel's coverage is an area, so a column of pixels gives a boundary's exact mean height, whatever its slope. That theory is machine-checked in Lean, and results the theorems speak about are accepted only by checkers generated from the proofs.

Full list, with the measurements behind each figure: the changelog.

From e02b2dfee4764fdc03f3710783763db3d16e0498 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 11 Oct 2026 12:07:03 +0000 Subject: [PATCH 2/3] cases: a junction is read where it shows, not from each region's path `junction_tri` failed on main after the 0.3.0 merge (junction_err_px 0.48 against 0.2), with the junction exactly where 0.2.7 put it. The check measured how near each large region's path comes to the true junction, which assumed a path is what shows. Since 0.3.0, Quality completes a face beneath the faces painted over it: the Y junction's lower wedge comes back as a rectangle under the other two (``), the visible corner is where its covers' edges meet, and the rectangle's own edge is 0.48 px from the junction under paint that hides it. The junction is now read from the render. The four source pixels about the true junction are rendered at 64 samples per pixel (the document nested in a window viewBox, so both frames stay the document's own), each artist region is matched to the traced area of its ink that overlaps it most, and the furthest any of them comes to the junction, less the truth's own reach (0.01-0.03 px of blended samples), is the error. Moving the true geometry by 0.05, 0.1, 0.2, 0.3 and 0.5 px reads 0.06, 0.11, 0.20, 0.27-0.31 and 0.43-0.50 on both cases, so the 0.2 px limit means what it said. At the whole-image SCALE the nearest sample inside a 120-degree corner can sit 0.2 px from its apex, which is why the window. junction_tri reads 0.080 on this build and on 0.2.7 alike (the visible corner, 0.09 px off, has not moved); junction_quad 0.035 and 0.014. Quality 42/42 and Fast's 28 recorded cases pass their ratchets. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_016AZWMqqHhRQ337N5EWhNC6 --- bench/cases/junctions.py | 96 +++++++++++++++++++++++++++++++--------- 1 file changed, 76 insertions(+), 20 deletions(-) diff --git a/bench/cases/junctions.py b/bench/cases/junctions.py index 38e3ccae..439dac19 100644 --- a/bench/cases/junctions.py +++ b/bench/cases/junctions.py @@ -16,10 +16,11 @@ from __future__ import annotations import math +import re import numpy as np -from _common import Case, Check, dist_to_polygon, poly_attr, svg_doc +from _common import INK_TOL, Case, Check, poly_attr, render, svg_doc A, B, C = "#204080", "#c04030", "#3f8f4f" SIZE = 128 @@ -50,14 +51,6 @@ def measure(t): return measure -def _shape_area(s) -> float: - tot = 0.0 - for r in s.rings: - x, y = r[:, 0], r[:, 1] - tot += abs(float(np.dot(x, np.roll(y, -1)) - np.dot(y, np.roll(x, -1)))) / 2 - return tot - - def _abut(name, what, body, seam_len): measure = _seam_checks(seam_len) @@ -79,9 +72,8 @@ def _junction(name, what, body, seam_len): def check(t): gap, bleed = measure(t) # Every region that meets at the junction owns a corner there. The one whose - # boundary strays furthest is the one that moved the junction. - big = [s for s in t.out_shapes if _shape_area(s) >= 200.0] - worst = _junction_error(t, big) + # visible corner strays furthest is the one that moved the junction. + worst = _junction_error(_window(t.truth_svg), _window(t.out_svg)) return [ Check("junction_err_px", worst, 0.2, "<", "0.2 px = 2x the output coordinate grid; beyond it the corner is visibly cut"), @@ -94,15 +86,79 @@ def check(t): return Case(name=name, what=what, svg=svg_doc(SIZE, body), check=check) -def _junction_error(t, big) -> float: - """The furthest any large region's boundary sits from the true junction point.""" - if not big: - return float("nan") - q = np.asarray([[JX, JY]], dtype=float) +# The junction is read from a render of the four source pixels about it, at 64 samples +# per pixel: on the whole image at SCALE the nearest sample inside a 120-degree corner can +# sit 0.2 px from its apex, the size of the limit itself. +WIN, WIN_RES = 2.0, 64 + + +def _window(svg: str) -> np.ndarray: + """The square of side 2*WIN about the true junction, rendered at WIN_RES per pixel. + + The document is nested, as drawn on the image's pixel grid, in one whose viewBox is + the window, so the two frames (the truth's grid, inkvec's pixel centres) stay the + document's own business. + """ + n = int(2 * WIN * WIN_RES) + svg = re.sub(r"<\?xml[^>]*\?>", "", svg).strip() + m = re.search(r"]*>", svg) + if m is None: # no output: no region reaches the junction + return np.zeros((n, n, 4), dtype=np.float32) + tag = re.sub(r'\s(?:x|y|width|height)="[^"]*"', "", m.group(0)) + tag = tag[:-1] + f' x="0" y="0" width="{SIZE}" height="{SIZE}">' + outer = (f'' + f'{svg[:m.start()]}{tag}{svg[m.end():]}') + return render.render(outer, n, n) + + +def _regions(rgba: np.ndarray) -> list[tuple[int, np.ndarray]]: + """Each ink's 4-connected areas of its own colour, a quarter pixel or more: (ink, mask). + + Read from a render, not from the paths, because a path is not what shows. A face drawn + whole beneath the faces painted over it (its covered part completed) has no corner at + the junction of its own; its visible corner is where its covers' edges meet. Blended + pixels match no ink, so two areas of one colour meeting only at a point stay apart. + """ + from scipy import ndimage + + out = [] + for k, ink in enumerate((A, B, C)): + rgb = np.asarray([int(ink[i:i + 2], 16) / 255.0 for i in (1, 3, 5)], dtype=np.float32) + near = (np.abs(rgba[..., :3] - rgb).max(-1) <= INK_TOL) & (rgba[..., 3] >= 0.99) + lab, n = ndimage.label(near) + for i in range(1, n + 1): + m = lab == i + if m.sum() >= 0.25 * WIN_RES * WIN_RES: + out.append((k, m)) + return out + + +def _reach(mask: np.ndarray) -> float: + """How near, in source px, a region's samples come to the true junction.""" + ys, xs = np.nonzero(mask) + d2 = ((xs + 0.5) / WIN_RES - WIN) ** 2 + ((ys + 0.5) / WIN_RES - WIN) ** 2 + return float(np.sqrt(d2.min())) + + +def _junction_error(truth: np.ndarray, out: np.ndarray) -> float: + """The furthest any of the artist's regions, as traced, sits from the true junction. + + Each region of the artist's is matched to the traced area of its ink that overlaps it + most, and the truth's own reach (its blended samples about the junction, 0.01-0.03 + px) is taken off. A junction moved by d leaves none of the corners further than d + away, and one of three 120-degree corners at least 0.87 d away, or one of four right + angles d away, so this reads the displacement itself: 0.05, 0.1, 0.2 and 0.5 px moves + of the true geometry read 0.06, 0.11, 0.20 and 0.43-0.50. + """ + traced = _regions(out) worst = 0.0 - for s in big: - d = min((float(dist_to_polygon(q, r)[0]) for r in s.rings), default=float("inf")) - worst = max(worst, d) + for k, m in _regions(truth): + mine = [o for j, o in traced if j == k] + best = max(mine, key=lambda o: int((o & m).sum()), default=None) + if best is None or not (best & m).any(): + return float("inf") + worst = max(worst, _reach(best) - _reach(m)) return worst From f1b7d062fd3eed86e541523266bba19d8c0b36d5 Mon Sep 17 00:00:00 2001 From: Claude Date: Sun, 11 Oct 2026 12:09:03 +0000 Subject: [PATCH 3/3] contract: the x86_64-windows-msvc hashes, recorded by CI for 0.3.0's output Taken unchanged from the facade (windows-latest) job's INKVEC_BLESS=add artifact on main at 3d1841c (bindings run 38135195486): 48 lines added, nothing else differs from the committed file. The release's re-bless on Linux had dropped every Windows entry, so the Windows contract ran on everything but the exact bytes until now. Co-Authored-By: Claude Opus 5.5 Claude-Session: https://claude.ai/code/session_016AZWMqqHhRQ337N5EWhNC6 --- bindings/contract/cases.json | 48 ++++++++++++++++++++++++++++++++++++ 1 file changed, 48 insertions(+) diff --git a/bindings/contract/cases.json b/bindings/contract/cases.json index 0a5ea232..eab57f05 100644 --- a/bindings/contract/cases.json +++ b/bindings/contract/cases.json @@ -15,6 +15,10 @@ "x86_64-linux-gnu": { "bytes": 8349, "sha256": "ef4e226361037eac0a1763f8610774ff0658d30b58d207473c291ce5470e5c46" + }, + "x86_64-windows-msvc": { + "bytes": 8349, + "sha256": "4a7c8eef1de25dcdd693f56d016ff302a0d3e447e552c9b95a27025e1110e8df" } } }, @@ -34,6 +38,10 @@ "x86_64-linux-gnu": { "bytes": 8349, "sha256": "ef4e226361037eac0a1763f8610774ff0658d30b58d207473c291ce5470e5c46" + }, + "x86_64-windows-msvc": { + "bytes": 8349, + "sha256": "4a7c8eef1de25dcdd693f56d016ff302a0d3e447e552c9b95a27025e1110e8df" } } }, @@ -55,6 +63,10 @@ "x86_64-linux-gnu": { "bytes": 5647, "sha256": "972e2f0a364778e775e62c4c87da0669cfecdbabfd227334a92ace53f0bc8800" + }, + "x86_64-windows-msvc": { + "bytes": 5656, + "sha256": "5fef039bbcf98b3bc3c71510496f231bbd2caf5e730cab25808e39f41f35788e" } } }, @@ -73,6 +85,10 @@ "x86_64-linux-gnu": { "bytes": 8369, "sha256": "89d5100eb55c47c07c134d19ef241a4bec81a0bdfd6478a15029606388b11032" + }, + "x86_64-windows-msvc": { + "bytes": 8369, + "sha256": "802210f58565e985f7bafc615dba72fc221e22a431f0c79366e98fbdfb9b8cea" } } }, @@ -91,6 +107,10 @@ "x86_64-linux-gnu": { "bytes": 2964, "sha256": "beb7db2b1ddcd9a068a51715056f526fa31cf4bf7c1d1a621975c4c8779feec2" + }, + "x86_64-windows-msvc": { + "bytes": 2964, + "sha256": "31174376bc8deecf23dc22b4fce0b323246607b31e45161ec8217831a706ff0f" } } }, @@ -112,6 +132,10 @@ "x86_64-linux-gnu": { "bytes": 2964, "sha256": "beb7db2b1ddcd9a068a51715056f526fa31cf4bf7c1d1a621975c4c8779feec2" + }, + "x86_64-windows-msvc": { + "bytes": 2964, + "sha256": "31174376bc8deecf23dc22b4fce0b323246607b31e45161ec8217831a706ff0f" } } }, @@ -131,6 +155,10 @@ "x86_64-linux-gnu": { "bytes": 2980, "sha256": "f72515edc7f187930b6d3cc95f03657ed4270959fc32cbcb09cfafbc667d1c06" + }, + "x86_64-windows-msvc": { + "bytes": 2980, + "sha256": "848ba0fe63e56871f84d0e668925eaaada13f214d1390c4cebdad4a5a539a9ed" } } }, @@ -147,6 +175,10 @@ "x86_64-linux-gnu": { "bytes": 682, "sha256": "76fe770f9a7610175a9ecf953c02f918e8d22e75617d8b77674efcefe85d0f58" + }, + "x86_64-windows-msvc": { + "bytes": 682, + "sha256": "76fe770f9a7610175a9ecf953c02f918e8d22e75617d8b77674efcefe85d0f58" } } }, @@ -166,6 +198,10 @@ "x86_64-linux-gnu": { "bytes": 682, "sha256": "76fe770f9a7610175a9ecf953c02f918e8d22e75617d8b77674efcefe85d0f58" + }, + "x86_64-windows-msvc": { + "bytes": 682, + "sha256": "76fe770f9a7610175a9ecf953c02f918e8d22e75617d8b77674efcefe85d0f58" } } }, @@ -184,6 +220,10 @@ "x86_64-linux-gnu": { "bytes": 682, "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" + }, + "x86_64-windows-msvc": { + "bytes": 682, + "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" } } }, @@ -203,6 +243,10 @@ "x86_64-linux-gnu": { "bytes": 682, "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" + }, + "x86_64-windows-msvc": { + "bytes": 682, + "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" } } }, @@ -225,6 +269,10 @@ "x86_64-linux-gnu": { "bytes": 682, "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" + }, + "x86_64-windows-msvc": { + "bytes": 682, + "sha256": "9fee7810aadc6f32ede16fc0f5c32ceeb1b6ef63cda375d40d9e34673eec713e" } } },