D and T: the dataset and threat-picture cycles (#45)

[bgent19]

Closes #45.

Two keys that move an axis without leaving the beat the instructor is on. Park on a Pareto detent, press T, and watch the same plan buckle under three different disruptions; or hold the same step and press D to change network. Day 2's textbook-then-theater move becomes a keypress rather than a server restart.

The beat index carries, and neither cycle assigns it

Both ladders are data-derived and their lengths genuinely move — the theater's interdiction ladder is three rungs pristine and two under strait_mined. So the index carries across and is clamped by the clamp setSpine already had. Nothing new was added to hold it in range, and neither cycleDataset nor applyThreat writes it.

The threat picture is per-dataset state

It is stashed on the way out of a dataset and restored on the way back in, checked against the target's own declarations — so the excursion to the textbook set comes home to the picture it left, rather than carrying a name the new dataset never heard of.

The dataset cycle opens on the default and moves away from it, so pressing D twice at two datasets brings you home.

T with nothing to cycle is labelled

Two of the three shipped datasets declare no threat pictures, so a silent decline is the common case — and a key that does nothing in silence is indistinguishable from one that is broken. The refusal says so on screen. The label is transient state, cleared by resolveBeat along with everything else derived.

Only D awaits, only D rebuilds

Every picture's payloads are prefetched with the dataset, because T is a keypress and anything a keypress can reach is fetched before the first keypress. T is therefore synchronous and installs spines from payloads already in hand; D is the one key that awaits, and the one structural rebuild the architecture allows.

One accepted deviation, recorded rather than overlooked: T installs new spine arrays, so repaint rebuilds the beat timeline's rows. A picture genuinely changes how long the story is, so rows have to appear or go — it is the same fact that makes setSpine's clamp necessary. What the criterion protects is the topology pass, and the test pins that directly: buildGraph has exactly one caller in the page. Grepping applyThreat's body alone could never have seen a rebuild reached some other way.

Supply figures, resolved in the panel

Removing a node zeroes its lanes and leaves its stock alone, so bravo_struck severs HUB-BRAVO and the theater still reports its full 180 of supply. A panel reading an unchanged total beside a visibly severed hub invites exactly the wrong question at the moment the instructor is making the opposite point.

The total stays the server's. A second line under it says how many hubs are severed, and the hub itself is greyed with its quantity struck through — the same absent-not-alarming vocabulary a removed lane already wears. Severance is read out of the changes block through isRemoved, so it catches a hub isolated one lane at a time as well as a removed node.

It is deliberately not a fourth NODE_STATE: those are Day 1's per-beat plan failures, and filed with them a severed hub would flicker back to healthy on the next keypress. A severed transit node is greyed and stays uncounted — nothing can move through it either, but it has no stock to strand, so counting it would inflate a figure about supply.

The note gets drawn

The instructor-written note on each disruption is displayed for the first time — it has been authored into every scenario file and drawn by nothing. It gets a slot of its own beside the headline's sentence rather than sharing it, because both are true at once on the textbook set in Cost & Risk: a flat frontier to declare and no picture to cycle.

No disruption arithmetic, and the tests enforce it

Every disrupted number is read off the server's changes block. The tests now fail if the frontend so much as names a disruption kind or reads a factor.

Review pass (second commit)

• severedNodes was building a nodeId -> [edgeId] adjacency map inside every render pass. Which lanes touch a node is topology, so it joins the crossings, the anchors and the directed lookup in frozen — solved once per dataset, handed in rather than imported.
• cycleDataset and the menu's change handler each spelled out stash-then-load. They are one switchDataset now, guarded against re-entry — which matters for exactly this key: D is designed to be pressed twice in quick succession, and two interleaved loads would build one dataset's DOM against another's payloads.
• resetView cannot finish R by itself (clearing the picture cannot re-solve the theater), so the driver reaches it through one named resetToBaseline.
• The severed class is a named export beside NODE_STATES, so a fourth state stays one edit plus a stylesheet rule.

Verification

• 390 tests pass; 12 net-new source-inspection tests in tests/test_web_assets.py (the repo verifies the frontend from Python or not at all).
• Exercised against real payloads on all three datasets: the theater's ladder shortening from three rungs to two under strait_mined, bravo_struck severing HUB-BRAVO with the 180 total intact, the labelled T refusal on textbook and greedy_trap, and a D round trip restoring the picture it left.
• Both /code-review axes run and addressed in 003f400.

🤖 Generated with Claude Code