format: an SVG drawing can be any size you ask for

[donislawdev]

Two new settings on svg: width and height, a whole number of pixels from 1 to 20000 each. They default to the 800 by 600 these drawings have always been, so a recipe that says nothing gets the same bytes it got before.

tfg generate --format svg --size 20kb --set width=1920 --set height=1080

Where the ceiling comes from

Not from the renderer - it never broke. Measured headless with Inkscape 1.4.4, dimensions read back from the rendered PNG with Pillow:

dimensions	megapixels	render	PNG
10000x8000	80	5.3 s	339 kB
20000x20000	400	25.6 s	1.6 MB
32000x32000	1024	43.6 s	4.1 MB
65536x65536	4295	165.3 s	17 MB

The line worth crossing belongs to a reader instead: Pillow refuses an image over PIL.Image.MAX_IMAGE_PIXELS, measured at 89478485 px, as a decompression bomb. 20000 per axis reaches 400 megapixels, four and a half times over that line, so a set can hold files on both sides of it. That is how the CSV column ceiling was chosen too.

65536 per axis was rejected with a number rather than out of caution: Inkscape manages it, but that is 165 seconds and 17 MB in somebody's CI for one line of recipe, and no measured reader has a threshold between 400 and 4295 megapixels.

There is no joint megapixel limit, unlike the picture formats, because nothing is drawn into pixels here - a picture format holds the raster in memory while it encodes, and this one writes text.

Two defects the settings would otherwise have introduced

Both invisible while the dimensions were constants, and both found by analysis before the code was written:

• the shape code assumes a margin of up to eighty units, and rand.IntN panics on a non-positive argument - measured by running it. --set width=50 would have been a panic on a value that looks entirely legal.
• a shape whose lower edge runs past the drawing paints over the label. That is the defect the strip along the bottom was added for: about one shape in ten used to land on the label, 11.0% and 10.3% measured. The strip fixed it for one canvas size, and it has to hold for every size now.

Both clamps are inert at 800 by 600 - the widest rect reaches 798 of 800, the tallest 542 of 544, two units of slack on every axis - so the guards press the sizes where they are not.

Bytes

D11 holds. Four sizes at seed 7 hash identically before and after, and stating the default out loud gives the same bytes as leaving it out.

A drawing shorter than 57 pixels has no room for the label along its bottom edge. It is still produced and still named, and the run says which files those were - the same sentence the picture formats use for a picture too small to write on.

Evidence

• four new guards in internal/guard/svgdimensions_test.go, seventeen canvases each, every case chosen for a reason
• five new mutations, all caught
• four existing SVG mutations repointed after the constants they aimed at became functions - they had gone stale silently, which is a SKIP that reads as proven
• oracle: four files rendered and measured, 1920x1080, 400x40, 50x600, 1x1, every one landing on the dimensions ordered
• full suite green, gofmt and go vet clean, 300 of 300 binary checks

🤖 Generated with Claude Code