guard: regenerating a screen touches the ones that changed, and a version stays linkable

[donislawdev]

Two things that went wrong while closing 0.3.0-rc1. Neither was broken code, and both were found by reading a diff rather than by anything red.

Regenerating the stored screens rewrote all twenty five

Only one had changed - About, because it draws the version. The other twenty four came back with byte identical pixels and files 12.6% to 28.9% larger.

The cause is the compiler, and it was measured rather than inferred:

	stored (1.26.7)	today (1.27.0)
generate.png	60 726 B	77 956 B (+28.4%)
preset.png	77 706 B	94 952 B (+22.2%)
recipe-refused.png	119 329 B	135 735 B (+13.7%)
twenty four files	1 866 203 B	2 278 108 B (+22.1%)

Re-encoding a committed file with today's compiler produces exactly the regenerated size. about.png, already written under 1.27, round trips byte for byte - one file under the new toolchain and twenty four under the old, in one run on one machine, so the measurement carries its own control.

So 402 kB of pure encoder churn across twenty four binary files went into the release commit, where git diff --stat says only Bin. What took it back out was somebody reading.

The writer now asks the guard's own question - does the stored reference still stand - rather than whether the bytes match. Byte equality was written here first and measured wrong on the spot: it rewrites every picture once on the first regeneration after a compiler update, which is the churn it was meant to stop. One judgement (judgePicture) now serves both the comparison and the writing, so the two cannot drift apart about what changed.

Measured after: twenty five of twenty five left alone and nothing written, while a reference replaced by a different picture is still the one and only file rewritten.

The changelog's link definitions had nothing holding them

A ## [0.3.0-rc1] heading went in without its definition, so it rendered as literal square brackets, and [Unreleased] still compared from v0.2.0 - showing the release's own changes as if they were still coming. A wrong target is worse than a missing one, because the link works.

Keep a Changelog is declared in the file's own header and in the project's git conventions, so this was a written convention with nothing behind it. Same class as the README table guard closed the same day. The guard also catches the drift running the other way, a definition left behind with no section.

Proof

Six mutations, all caught. Two of them are the two faults themselves. Full suite green, preflight --quick green on all twelve checks including staticcheck, lint and govulncheck.

🤖 Generated with Claude Code