Both states

The demo menu is the recessive icon on the right. §10 wants it reachable without looking like part of the product, and those two pull against each other — a menu nobody can find is not a demo control, and a bar of nine buttons is a second product.

The fraction is said twice on purpose, and it is the only thing here that is: Step 3 of 7 is the sentence, the seven ticks beside it are the instrument. They cannot disagree, because both are derived from one BuildStep list. Demo feed and Board simulated stay for the whole session — they are §6.1's requirement, not decoration. Agent connected is the one badge that can change, and when it does it changes glyph as well as word: absence gets its own mark rather than the connected one greyed out.

Four states, four words

The rail does not navigate. Three things already move the build between steps — the agent through navigate_build_step, the demo menu, and the topbar's progress control, whose expanded list has been a step picker since Batch 1. A fourth would be a fourth thing to keep in step with the other three, and the rail's job here is orientation.

Position carries three of the four states on its own: behind you, here, ahead. It does not carry the one that matters. A step you have already walked past can still be Issue, and keeping that visible is most of why the rail exists — so all four are written out, quietly, in the line where the estimate already lives. Writing them down is what caught the fourth: upcoming reads Not started, because the agent can move the build to step 4 with step 3 still unverified, and a row above the active one calling itself upcoming is the rail saying something untrue. The disc is the same StepMark the topbar's expanded list draws, exported rather than redrawn: two pictures of one state is how the rail and the topbar end up disagreeing about which step is blocked.

Components in this step

Derived from the connections the step owns, never listed by hand — a written list stops matching the graph the moment a wire moves. The board and the breadboard are deliberately not in it: they are the substrate, seated in step 2 and never fetched again. What the derivation cannot see is the two resistors in the LED step; they have no node in the graph because they sit in series inside a leg, and the step's own instruction names them one line above the canvas.

The workbench, live

The pressure is vertical, not horizontal. At 1280 the canvas has 1280 − 252 − 360 = 668px across and, with the dock open, 900 − 64 − 224 = 612 down — and the instruction takes another 70 of that. Which is why the canvas controls float over the well instead of sitting in a bar above it: horizontally there is room to spare, and that is where the controls take their space.

Open and close the dock while the agent is pointing at a pin. The canvas does not move. It fits itself once, when it first has a size, and never again — Batch 4's note: refitting on every resize looks tidy and throws away the focus the agent just asked for, which is a change the user did see undone by one they did not.

Everything works: inspect, show, fix, verify, run the test, open the dock, switch the view, open the inspection. Every button runs the same tool a WebMCP callback will call — there is no second path on this page.

The findings are FindingRow — the actual G-05 component, reading the actual Finding objects the panel behind the modal is reading. §7 says the detections in here use the same central state; the strongest way to keep that true is for there to be one component, so a change to how a finding reads cannot land in one place and not the other.

Two transforms, not three. The camera pane is a real viewport with its own handle, because the agent has to be able to take it to a pin. The reference beside it is a fixed viewBox — a reference view is something you compare against, not something you fly around in. It is cropped to the same focus box the agent uses, because the whole board at 380px is a picture of nothing: the correct answer has to be drawn beside the wrong one at the same magnification.

M-07 fed, not rewritten: focus trap, Escape, scroll lock and focus return are already paid for. Open it, press Show me, close it — the workbench behind is looking at what you were just shown, because the modal's Show me is the same show_correction call the panel makes.

The horn, twice

The second fault is not a wire and cannot be shown as one. This is the canvas cropped — same MicroServo, same 1:1 scene units, same mat underneath — rather than a diagram drawn for this panel, which would have been a fourth drawing scale in a product that has decided it has three. The ghost is C-18 unchanged, so the expected horn is green here and green on the workbench. The arc is the only addition, and it earns its place by carrying the quantity: you can see two horns without being able to say how far apart they are.

Press one and watch the timeline

Nine items in a flat list is a wall. They come in the three groups the demo itself has, each reading jump to it · put it back · fix it, which is the shape of a take. Reset sits above them on its own, because it is the only one that throws away what the others set up.

The two inject actions are the only thing here that is not a tool call, and they cannot be one: no tool in this product breaks a build. They go through the same act() as I fixed it, which is the right neighbourhood — in the fiction they are the same gesture in the other direction, and they land in the timeline as the person's own line.

Every entry that appears is a real tool call against the real store. §10: the demo buttons do not run a second flow. That is also why Complete project takes several seconds rather than being a state assignment — it repairs both faults through inspect_build and then verifies each remaining step, which is what completing the project is.

A sentence, not a box (rule 4). Same job as You can continue in guided demo mode: nothing is broken and nothing is blocked, so a warning-shaped container would say the opposite of the words inside it. Info rather than warning, for the same reason — a narrow window is not a fault in the build.

The canvas is not hidden — it is not rendered. §16 says the circuit is never shrunk to the point of meaninglessness, and a viewport mounted inside a 380px column would fit itself to that width once and keep the result. The steps, the findings and the agent are all still reachable, which is the other half of what §16 asks for, and the agent panel becomes the right drawer Batch 5 declined to spend on the dock — precisely so that it would still be here.