D009 - The window matches orbistoun's, deliberately, down to the version

decided - 2026-08-26 - building pros-gui

eframe 0.29, default features off, wgpu backend, egui 0.29. The same versions and the same backend as the sibling project's shell, chosen by looking at its manifest rather than by picking what was current.

The reason is a thing worth having later: the two windows do the same kind of work over the same library, and a shared GUI crate is a sensible eventual move. It is only a move if the two are already the same underneath; otherwise it is a port, and a port is a project.

wgpu earns its place here on its own terms too. A lossless frame grab has to be displayed somewhere, and that is this repository's own future need rather than a borrowed one.

What is tested, and what could not be

A window cannot be looked at from a test, so the crate is split so that most of it is not a window:

• state.rs - the rules. One job at a time, because two racing would interleave into a window showing half of each. A failed job clears the panel it would have replaced, because a report left on screen after a refresh that failed is a claim about a target that has stopped being true - the caching mistake, made in pixels. An answer arriving when nothing was asked is dropped.
• work.rs - the asking, on its own thread. A check is five ports at a second and a half; done on the drawing thread the window stops repainting, and a window that has stopped repainting is indistinguishable from one that has crashed.
• app.rs - drawing only. The one thing tested is the wording, which must match what pros prints for the same finding. Two front ends over one library that word a verdict differently is how a shared library quietly stops being shared.

It was run, and that changed something

The claim in the last entry was that a window could not be verified from here. That was wrong: the operating system will say whether a process owns a window and whether it is answering, and the screen can be captured and looked at.

Doing so found a real defect that no test would have. The build stamp read built 1787734934 - seconds since the epoch, chosen deliberately because it needs no dependency and no locale, and entirely correct. It is also unreadable, and a stamp exists to be read at a glance. It is now computed to YYYY-MM-DD HH:MM UTC in the build script, in twenty lines of the standard era-based conversion, still with no dependency. Always UTC and it says so, because a stamp in local time is ambiguous the moment somebody elsewhere pastes it into a report.

The stamp is assembled in one place and used by both front ends, so pros --version and the window's footer cannot disagree about which binary somebody is looking at.

The rule this is the second instance of today: when something can be looked at, look at it. Reasoning about the fake's accept loop said it was fine; measuring said 487 ms. Reasoning about the stamp said it was correct; looking at it said nobody could read it.