release: 0.3.0

[donislawdev]

The owner's decision, and his alone - immutable rule 12. 0.3.0 without the
suffix, which is the whole mechanical difference from 0.3.0-rc1: the release
workflow marks a release as a prerelease when the TAG NAME carries a hyphen,
and that is what kept rc1 off the front page and out of latest.

The changelog is one section, not two

The entries that shipped as 0.3.0-rc1 were folded into 0.3.0 on the owner's
instruction, and the ## [0.3.0-rc1] heading went with them along with its link
definition - a definition left behind points at a comparison nobody can reach
from the document.

The reason is stronger than tidiness. The Breaking list is only true read
whole: the padding of a ZIP, a TAR.GZ and a WAV changed after rc1 went out,
so the two halves describe one delta from 0.2.0. Somebody upgrading needs one
list, not two.

The merged section carries 67 entries - 9 Breaking, 22 Added,
5 Security, 14 Changed, 17 Fixed - under one heading of each kind.
[Unreleased] is empty, so both of the release gate's own conditions pass.

Nine copies of the number, seven of them held

Five readers take version.Version at run time and needed nothing. The copies
were found by grepping the tree rather than from memory:

copy	what turns red
cmd/tfg/tfg.rc and its .syso	exeproperties_test.go reads the COMPILED resource, so editing the script alone changes nothing - both were recompiled with windres and each lost exactly sixteen bytes
internal/gui/icon/chickpea.rc and cmd/tfg-gui/*.syso	the same
about.png and about.xml	screenpixels_test.go - one screen of twenty five rewritten, the text 37 px narrower
web/public/**	site_test.go - thirteen pages, seventeen lines added and seventeen removed, every one carrying the version
CHANGELOG.md	changelogshape_test.go and the gate in release.yml
the mutation entry	the staleness check
README.md manifest example	nothing - it still said 0.2.0, so it had drifted through a whole release. Fixed here and recorded as O198

The run lock has a second effect, and the changelog now says so

The pre-release ritual reported a failure about the disk filling part way
through a run. It was not a regression. Measured in a container with room for
twelve names, the same command through three binaries:

binary	exit	manifest
0.2.0	5	not written
0.3.0-rc1	5	not written
this release	8	written

Bisected over the 48 commits since rc1 to the run lock, and the cause is in
engine.go: the manifest name is now taken before the first file rather than
after the last one, so the manifest has a place reserved whatever happens to the
directory. The old run left files cleanup had nothing to remove them from.

Both codes follow the frozen table. What moved is which of them the situation
produces, and somebody's CI can be reading it, so the entry says so now.

What was run

• the whole suite, 337.8 s, green
• the pre-release ritual, 179 of 179, nine phases: real files through cmd,
  PowerShell 5.1, pwsh 7, Git Bash, WSL on the Linux binary and a Mac over
  ssh, a real window opened, and a comparison against the downloaded
  v0.2.0 - eleven formats unchanged, nine changed, all nine described under
  ### Breaking
• the guard suite on Linux in Docker, amd64 and arm64, on overlayfs
• three mutations named by the guards this touches, three caught

🤖 Generated with Claude Code