format: one name for one setting, and one place for a picture side

[donislawdev]

What this is

Width and height were declared from scratch in ten picture formats, and the
function reading them stood in all ten as well - the same twelve lines,
differing only in the format named inside its message. This gives them one
declaration and one reader, and adds the guards that keep an eleventh format
from drifting away from both.

Two of the three clusters of shared settings were already held together by a
shared package: the container axes and the text encoding. This is the third
and largest one.

What is shared, and what deliberately is not

internal/format/imagedim owns what cannot legitimately differ - the name,
that it is a whole number, that the smallest side is one pixel, that the
number counts pixels. Each format still supplies what was measured to differ
for a reason:

• the largest side (20000 for five formats, 16384 for AVIF and JXL, 16383
  for WEBP, 256 for ICO - each is the ceiling of the thing itself)
• the sentence, because an icon is not a picture and an SVG only says how
  wide it is
• the default, which only SVG declares

The reader unifies on the best of the ten rather than the most common. Nine
built a string with fmt.Errorf. SVG returned format.PropertyValueError,
the same type the registry raises, so a reporter can ask for the parts by
name instead of parsing a sentence. The wording is SVG's, unchanged.

Guards

• TestOneSettingNameMeansOneKindOfSetting - a name written by two formats
  has to describe the same kind of setting. Asks about Kind and Unit
  only, because Max, Default and Min diverge for reasons that were
  measured first. Asked of every registered format and every shared name, so
  a cluster that does not exist yet is covered on the day it arrives.
• TestEveryDeliberateHomonymStillNamesTwoMeanings - the one deliberate
  homonym has to still be one, checked from both sides. entry_format is
  that homonym, and the guard banning it outright was written and deleted on
  2026-09-01 for reddening on it.
• TestThePictureSideBackstopRefusesWhatTheDeclarationWould plus a control
  that it did not become a wall.

The debt this repays

The range check inside those ten copies could not be reached. The registry
refuses first in engine.go and recipe/target.go, the only two production
callers - png with width=999999 comes back as exit 4 with the sentence
built from the declaration - and nothing in internal/guard reached a
generator with a value outside its range. It is kept for the reason
textenc.Parse gives about its own branches, and keeping it moved the
burden to a test, which it now has.

Measured

• tfg formats --json identical byte for byte, all twenty four declaration
  hashes unchanged
• full suite green, gofmt and vet clean, preflight all checks passed
• four new mutations, all caught. Three existing entries went stale when the
  ten helpers became one - found by staleness.py, repaired and re-run
• net 88 lines added, 300 removed across the ten format packages

No changelog entry: nothing a user can observe has changed.

🤖 Generated with Claude Code