FACE LAB

THE ONE CLOCK — the continuity proof

Everything each face draws is a function of one monotone number. If switching faces restarted anything, BEATS SINCE ARM would jump and CLOCK REBASES would climb. Switch faces mid-bar and watch neither move.

And check that the needle can move at all. A counter wired to code that cannot increment it proves nothing, so the console carries its own negative control: set BAD FACE · RE-ARMS and switch. That branch does exactly what a face with an .onAppear { engine.restart() } would do — the row turns red, the ledger says CLOCK RESTARTED, and the bar visibly snaps back to 1. Set it back to CLEAN SWITCH and the counter stops climbing. That is the proof; the zero on its own was a thermometer with no mercury.

SWITCH LEDGER

VERTICAL BUDGET — measured, not asserted

852 pt screen − 59 pt safe area − 83 pt tab bar = 710 pt. SPENT is the sum of the active face's non-spacer children and their margins, read from offsetHeight — a layout value, which a CSS transform does not affect, so this is true points at any stage scale. The spacer's own measured height is printed beside the slack as an independent cross-check; the two agree within integer rounding. Under 710 means the SE-class latent scroll (MetronomeView.swift:106-122) never actually fires.

Measured, all four states fit: CLASSIC 558 at 4/4 · RADIAL 681 · RADIAL + MAP 683 · STAGE 517. Note what that ranking says — the radial face is the tight one either way, and D1's published estimate of "~42 pt slack" does not survive adding a session pill. Dynamic Type on the gauge captions has to be capped or the dial shrinks first.

Move the METER and watch CLASSIC's number move. It did not, until 2026-08-30: this face drew every meter as one row, so it read 558 pt at 3, 4, 5, 6 and 7 alike — insensitive to the one variable that has historically broken this layout. The shipping row splits above six beats (LEDBeatRow.swift:66), so 7/4 now costs a second row — 625 pt, 85 pt of slack — and 6/8 draws the compound group divider instead of hairlines (still 558: an absolutely-positioned divider is free). A budget that cannot go up is not a budget.

ACCENT FAMILY — computed both ways

DERIVED solves each token's lightness by bisection until its WCAG ratio against #0C0C0E equals the ratio the shipping Ember hex actually measures — the targets are read off Theme.swift:123-139 at load, never typed.

Same SOLVER as accent-lab.html:1205-1227; not the same constants, and saying "same law" was too loose. Shared: bisect lightness at a fixed saturation until the ratio against the ground hits a target. Different: this page reads its targets off Ember rather than typing them, and it picks its own per-role saturations — six of nine differ from accent-lab's bright-field set (ink 8→42, dimText 30→60, ghost 12→60, borderDot 32→8, lit 72→85, deep 68→76). The one that could have mattered is burnt: accent-lab's Studio family anchors underlay to the armed colour (accent-lab.html:1214), which this page does not — but its Console family, the dark one, solves underlay against the ground exactly as here (accent-lab.html:1220, S 55 to this page's 60, target 3.00 against Ember's measured 3.03). On this field the two agree in shape and very nearly in number.

ARM-STEP is the luminance jump from resting accent to armed — the thing that makes a beat read as an onset rather than a hue. TIER SPLIT is ember-vs-cream: if it collapses, the once-per-bar law loses its second tier.

The two ΔE rows are the only ones that can fail, and that is the point. Every WCAG row above is algebraically pinned — each token is solved to its own ratio against the same ground, so ratio(lit,amber) ≡ ratio(lit,bg)/ratio(amber,bg) and the ACTIVE column reproduces the EMBER column at every hue on the wheel. What the solver does not preserve is chromatic separation: two tokens can both hit their target luminance and still slide onto each other in colour. The radial face reads by exactly that — played versus unplayed is a burnt/cold area code, and the once-per-bar law's second tier is lit/ink. Both are measured in ΔE*_ab, and their floor is 80 % of what Ember itself achieves for the same pair, so the floor is derived too. Drag the hue slider: the ΔE rows move and can go under; the seven above them will not budge.

AND THE DEFAULT HUE IS UNDER BOTH FLOORS — DELIBERATELY, BECAUSE THAT IS THE FINDING. Swept all 360 hues on this page: Ember measures 42.2 area and 45.9 tier, and 217 of 360 hues clear 80 % of that. The ones that do not are a single contiguous arc — H 161°–257°, cyan through blue into indigo — bottoming at 26.9 at H 185° (area, −36 %) and 29.9 at H 181° (tier, −35 %). That arc is where accent-lab.html's recommendation lives. The luminance law says Cyan is legal and it is right; this says a cyan family will read the played-vs-unplayed ring and the ember-vs-cream tier noticeably flatter than Ember does, and that cost is invisible to every contrast row on this table. It is not a veto — it is a number to price, and it is exactly the sort of thing an instrument with a pinned needle would never have shown.

R1 GEOMETRY — arc-minutes at the stand

θ′ = cap_pt ÷ (1.8667 × d_m), from R1 §2.2's h ≈ d·θ·0.0002909 and §1's 163 pt/in. Cap ratio 0.700 for Bebas and Overpass Mono alike, measured from the bundled TTFs (glance-lab.html:1553-1557). R1 §7.1 puts the BPM numeral in the 20–22′ optimal band at its working distance; secondary text may sit at the 14–16′ floor.

Switch to CLASSIC and read the top row. Bebas 72 is a 50.4 pt cap — 27.0′ at 1 m and 13.5′ at 2 m. At a music stand the default face is comfortably readable; at 2 m it is under R1's 14′ automotive floor, never mind the 20–22′ band its own category calls for. That single number is the entire argument for alternate faces — not that CLASSIC is wrong, but that it is a 1 m instrument, and STAGE (24.7′ at 2 m) is the same metronome answering a different room.