The beat engine, keyboard controls and layer catalog (#41)

[bgent19]

Closes #41.

The keyboard becomes the primary driver. One pair of keys means the same thing all four days, and every state is one keypress from anywhere. This ticket builds the beat engine, the view scaffolding and the layer catalog that the three view tickets (#42–#44) then fill in.

The beat engine

Lives in state.js, because the beat index is app state and that module owns app state. A beat is a plain record — {unit, phase, label, apply} — and a view's ticket installs a whole spine through setSpine.

setBeat clears every derived field before applying the beat, so no beat can depend on having passed through its predecessor. That is what makes a beat render identically whether it was stepped into, stepped back to, or jumped at cold — and it is what forces the prefetch rule the spec keeps arriving at: a spine is installed complete or not at all. A beat that had to fetch could not be synchronous, and a beat that is not synchronous cannot be stepped backwards through at speaking pace.

Ends never wrap, and an out-of-range digit is a silent no-op rather than a clamp: a clamp moves the picture somewhere nobody asked for, mid-sentence.

The coarse controls address units by ordinal, not by the value in beat.unit. The instructor says "go back to the second one", and what makes it the second one is its position in the story. Reading the value directly would mis-jump on both spine shapes #42–#44 are expected to build — an interdiction ladder whose k starts at 1, and a Pareto spine with duplicate detents collapsed out.

Views are jumpable in any order and each remembers its own beat and hand toggles, so Flow & Cut stays parked on augmentation 3 while the instructor detours to answer a question.

The layer catalog

One global catalog with per-view entry states, never view-owned — the unit's strongest moments are cross-day mashups, and Day 3's cut over Day 4's interdicted network is Day 4's entire punchline.

Layers resolve from three sources in order: the view's entry table, the beat's overrides, then the instructor's own toggles. Rebuilding from all three on every beat is what keeps a beat idempotent including its layers, while leaving the hand toggles standing — R is the only thing that clears them, which is exactly where the spec spends that cost.

The old single lane label splits into caps and costrisk, because Flow & Cut wants a capacity beside its flow and Cost & Risk wants cost and risk instead; one text carrying all three put two unasked-for numbers on screen in both. All-layers-on is illegible in every scheme and is not a legal preset.

The scaffold tier

Untouched lanes dim to 30%, and this tier owns opacity — nothing else may use dimming as a channel, or two mechanisms compose into a lane at 9% nobody authored. Safe only because an empty hot set means every lane is hot, so the pristine network — the one picture whose subject is the graph itself — is never dimmed. Labels only on hot lanes falls out of that rather than being a special case.

The stroke model

The token block gains the back-row smear as one authored quantity with a derivation per pattern. Two floors apply to every dashed mark and they are orthogonal:

• pitch — is it still broken? Exactly independent of stroke weight: a dashed stroke is a union of rectangles and a Gaussian is separable, so the across-stroke term cancels out of the modulation ratio identically.
• presence — is it there? weight × duty × opacity, against 1 MAR at the 20/40 design observer.

The failure mode is the whole argument for machine-checking them: a dash that loses does not vanish, it becomes a dimmer solid line — still there, still the right colour, saying the wrong thing.

Adopted values from #25 and #27: residual arc to stroke 4 / 22 18, min cut 16 24, synthetic machinery and removed lane re-pitched to 6 22. Those last two stay under the presence floor deliberately — the machinery is scaffolding and may fade, and removal's meaning rides on the X glyph — on a recorded and closed exemption list, so the question is not reopened a third time.

tests/test_web_assets.py re-derives the 3.16 dashed-stroke floor from the smear rather than restating it, so retuning the room moves the floor instead of leaving a constant nobody can re-justify. The recorded 3.16 is compared directionally against the 3.13 the model implies: a shade conservative is a rounding, a shade generous is a mark authored under the floor with a test agreeing.

Verification

11 new tests, all under the repo's Python-or-nothing rule — no JS test runner, and the module inventory stays at seven. 311 pass.

Two judgement calls worth a reviewer's eye:

• The promoted reverse arc's 24 15 is not in the token block. Its duty of 0.615 is over this ticket's own ≤ 0.6 authoring cap, and it clears the pitch floor by 1.008× — inside the closed form's own ~1% error bar, which is to say it is a value the rule cannot actually certify. It is also geometry-specific to an arc promoted to offset 22 with a hollow head. It ships with the mark that justifies it; recorded in style.css so Flow and Cut view (Days 2-3) #42 re-opens the question rather than rediscovering it.
• The two-row anchor grew the label footprint from ±12 to ±25. Two rows are not optional — Flow & Cut shows a capacity and a flow, and the single-row layout that preceded this drew them on top of each other. It clears every distance the anchor solve enforces (74 from a node, 74 from another anchor, 44 from a crossing) and exceeds only the 14-unit lane half-band, which governs casing width rather than labels. Worth an eye against the real projector, since Authored coordinates collide at legible type sizes #17's clearances are measured work.

T and D land in #45; the beats themselves land in #42–#44.