[timzifer]

golang.org/x/image supplied two things: an anti-aliased rasteriser and a
bitmap font. Those are exactly the two that decide what this library's pixels
are, and a library whose whole premise is that a version always produces the
same mark cannot outsource them. A release that sharpened its anti-aliasing by
a fraction of a pixel would change every mark ever generated, and it would
arrive as a routine version bump.

go.mod now lists nothing, and TestNoDependencies keeps it that way.

The rasteriser — canvas/fill.go

Signed-area accumulation: each edge deposits, per pixel, the signed area it
sweeps in that column; a running sum along the row gives coverage. Exact for
straight edges, so filling a shape and summing the alpha returns its
geometric area — which turns "is the anti-aliasing right" into arithmetic
rather than taste. The tests assert it to within one percent for rectangles
placed off-grid, sub-pixel slivers, triangles, annular sectors and circles.

Direction folds into a sign, so the non-zero rule the canvas already relied on
falls out of the accumulation rather than needing a winding pass — reversed
inner contours still punch holes, same-wound overlaps still union.

The font — canvas/font.go

A 5×7 bitmap font over printable ASCII, written out as pictures:

'A': ".###./#...#/#...#/#####/#...#/#...#/#...#",

A wrong bit in a hex table is invisible; a wrong dot here is not. Tests check
every glyph's shape, that the set covers printable ASCII, and that no two
glyphs collide.

What the swap exposed

Only polar moved: its patch minimum fell from 24 to 13. The marks were
visually identical — a pixel diff showed at most 9 of 255, on anti-aliased
edges alone.

The cause was not the new rasteriser. It was the difference hash, which
point-sampled its 9×8 grid: shift an edge by a fraction of a pixel and the
sampled pixel crosses it, flipping bits wholesale. Thresholds calibrated
against that were partly measuring the rasteriser rather than the marks.

The grid is now box-averaged, and CheckMetricStability renders the corpus at
two resolutions and requires the mean distance to agree — so the instrument is
checked as well as the renderers:

	before the fix	after
polar mean, under a 9/255 edge wobble	−6 bits	—
any renderer's mean, across a 1.5× resolution change	—	≤ 0.6 bits

All thresholds are re-frozen against the corrected measurement. Some rose
(truchet's patch minimum 16 → 24), some fell (flowfield 28 → 20), which is what
removing a noisy component looks like. They are re-measured after fixing the
instrument, not nudged until the suite went green.

Also

• Contact-sheet labels pick the largest whole font scale that fits their cell,
  so a long prerelease string no longer runs into its neighbour.
• Gallery regenerated.
• Dependabot's x/image ignore rule is gone with the dependency; the gomod
  entry stays so that the day one is added deliberately, it is also watched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KwWthYNYf1egKNFrFS83Dm

[timzifer]

Two things worth recording, both surfaced after opening this.

Dependabot alerts on main. Pinning x/image at v0.24.0 to hold the Go
floor at 1.21 also pinned its advisories: one high and one moderate, both
excessive-resource-consumption bugs in the TIFF decoder — a package this
library never imported. Not exploitable here, but a fair illustration of the
cost of holding an old pin for an unrelated reason. This branch removes the
dependency, and the alerts go with it.

The CI matrix earned its keep again. macOS failed on a single polar pair —
0.0.3 against 0.1.3, measured at 16 bits on amd64 and 12 on arm64. Nothing
about the renderer differs: arm64 contracts float expressions that amd64
evaluates separately, which shifts an anti-aliased edge by a fraction of a
pixel, and on a borderline pair that is worth several bits.

So a bound set just under one machine's observed minimum is not a bound. Means
are stable to well under a bit — the resolution-independence check already
showed that — but the thresholds are stated in terms of minima, and those are
not. The margin moves from about fifteen percent to about thirty, blazon dist
now suggests bounds with that margin already applied, and both CONTRIBUTING and
the table say why the room is there.