[schreck61]

Seven defects turned up while docs/architecture.md was checked claim by claim against the 0.5.1 code. Each was confirmed in the code, and where it shows, in the app, before it was changed. Each fix is a commit of its own, with its test, so they can be reviewed one at a time. The last commit updates the architecture document.

Defect	Fix	Test
Hybrid weights files that don't load. Save weights passed NaN as the score, and train saved an infinite one when no static tuning passed the pedestrian check. JSON writes both as null, which SavedHybrid rejected. weights::load returned the first readable copy without checking that it loads, so such a file also hid the other copies.	base_score is an Option<f64>: a score that isn't finite is saved as unknown, and 0.5.1's null loads. weights::load takes the parser and skips a copy that doesn't load. The app saves the network only when it is trained or learning (has_model || learning), not an untrained one.	tests/weights_files.rs: saving with each kind of score, with and without a network, then loading; a 0.5.1 file with null; a damaged copy skipped for the next one and for the built-in one
The green wave lost the controllers. It switched every signal to fixed-time, and turning it off left them that way.	World remembers each signal's controller when the wave starts and gives it back when the wave ends, unless another was chosen meanwhile through the new World::set_controller. Rebuilds during the wave keep the remembered controller. The tooltip says what the wave does, and the bottom panel confirms each change.	New tests/green_wave.rs: a four-signal corridor with mixed controllers, one chosen during the wave, run without collisions; a single intersection left alone
The roundabout's pedestrian tooltip said pedestrians use underground passages.	Each layout without crosswalks gets its own reason.	UI text only
A Hybrid whose saved network doesn't fit said "no network beat the best static tuning".	Hybrid records that the network was left out (model_rejected), and the brain says "the saved network doesn't fit this layout".	tests/weights_files.rs: both messages
The signal timing chart never showed yellow or all-red. Each span ran from one green to the next.	StageSpan has a phase; the history records the yellow and all-red as their own spans (1,200 kept, up from 400), and the chart draws them in yellow and red.	Unit test in sim/signal.rs
The margin over the gap showed a negative value as "+0.00 s", in red.	margin_kpi keeps the sign. Rounding error just below zero still reads "+0.00 s".	Unit test in gui/mod.rs
In a corridor, the Controller list and status overlay showed the last choice, not the focused intersection's controller, after the focus moved or the green wave changed controllers.	App::sync_ctrl reads both sides' controllers from the focused intersection after a rebuild and at the end of each frame, and asks for another frame when they changed.	None: App::new needs an eframe creation context. Checked in the app

Checks

• cargo build --locked --release --all-targets
• cargo clippy --locked --release --all-targets -- -D warnings
• cargo clippy --locked --release --target wasm32-unknown-unknown --bin intersection-lab -- -D warnings
• cargo test --locked --release
• cargo deny --locked check
• The weights fuzz target for 60 s: 271,436 runs, no crashes. That run came before rebasing onto the current main, which didn't touch the weights code.

Every commit builds and passes clippy on its own.

In the app, with the scripted-screenshot settings: the green wave on and off on a corridor with mixed controllers, the timing chart at 120 s, and a three-signal corridor with different controllers as the focus moves.

Not tested in the app:

• The roundabout's tooltip: hovering can't be scripted.
• A negative margin over the gap: the cooperative tests require every margin to be at least −10⁻⁶ s, so the app doesn't produce one. The unit test covers the formatting.

Traffic behavior

None of these change how traffic runs. Only the timing chart reads the signal history, and a weights copy that loads is used as before, so bench results are unchanged.

Documentation

docs/architecture.md now says it describes the code on main: 0.5.1 and the fixes since. It covers the loading order and outcomes in §8.7, the green wave in §6.10 and the tour, the signal history and timing chart, the margin's sign, the tests, the file sizes and the repository layout. timeline.png was captured again with the Appendix A settings; only the yellow and red segments changed. The whole-window images, such as overview with its hand-drawn callouts, still show the 0.5.1 chart at small scale.

A possible follow-up

The green wave switches both sides of a comparison to fixed-time, so side B can't run an adaptive controller against a green wave on side A. Applying the wave only to signals already running fixed-time would allow that comparison. That changes how the feature works, so it isn't part of this PR.