feat(desktop): animate the terminal splash banner

[tlongwell-block]

Sub-PR of #4347. Base max/tui-renderer @ d8a4d63a8. Merge commits only — no rebase, no force.

Tyler's ask: slow waves of buzz theme colours across the honeycomb lattice, plus a separate slow wave on the buzz term wordmark travelling left-to-right. Eva's amendment (evt 5c2182c3): the field wave originates bottom-right and travels top-left, as a projection onto a direction vector.

What this does

Two independent waves over the existing banner geometry — different directions, different periods. Geometry and glyphs are untouched; this only supplies a hue axis and the colour table that makes it affordable.

Field wave is a projection, not a per-axis sweep:

phase(cell) = (x·dx + y·dy) / wavelength − t·speed

normalized in visual space, because terminal cells are ~2:1 (8.4×17 CSS px). Without the aspect correction a "45°" direction renders at ~63°. The direction gate asserts the measured travel angle in visual space, not grid space.

Tunability is one config block per layer at the top of terminalBannerWave.ts (FIELD_WAVE, WORDMARK_WAVE) — direction, seconds, wavelength, intensity, each commented with its unit, what "bigger" does, and a measured contrast-safe range. Gates run at the documented extremes via a waves? override, so no module state is mutated to test them.

Invariants, all gated

• prefers-reduced-motion takes the static painter call — the shipped banner byte for byte, not a paused animation. Those differ: a stopped loop parks on whatever phase it halted at. Gated against an independent oracle that walks the geometry and computes expected colours with bannerColor directly.
• Wordmark holds WCAG 4.5:1 on every frame. Interpolating between samples that sit exactly on the floor dips under it — contrast is convex along an sRGB chord, measured 4.4766 on solarized-light at hue 0.2188. The blend is re-lifted with the same operator that placed the endpoints. Raising bucket count only shrinks the dip asymptotically; lifting is exact.
• Field keeps its monotonic 2.60 → 1.04 decay and bevels stay phase-invariant. hue and the contrast axis are separated, so a moving wave cannot brighten the rim or make the chassis appear to deform.
• Smoothness in CIEDE2000, per theme, relative to that theme's own static worst step (+1 JND). Euclidean sRGB scores hue rotation at constant lightness as harshly as a brightness cliff and reversed the verdict on 20 themes — worst excess is 0.47 at defaults, <0.62 through wavelength 2.0, and 8.92 (24 themes failing) at 4.0, so the gate bites.
• Seam: a cyclic phase cannot drive an open 3-stop ramp — a raw wrap put a 182.5 sRGB edge on screen once per period. A palindrome fold fixes it, so direction is only defined inside a half-cycle and the direction gate measures a crest centroid.

Perf

No per-cell-per-frame ramp(). Colour is a phase-independent LUT built once per palette; steady-state frames are a lookup plus a lerp.

REGIME: Chromium 148.0.7778.96, DPR 2, macOS, headless, throttled (1 paint per rAF tick, vsync ON), paired arms, 5 reps × 120 frames, buzz-dark, cell 8.4×17 CSS px, per-frame paint only. Magnitude is rig-local; only existence of an effect replicates.

LUT fill, one-time per palette: 128 hue × 24 contrast buckets, median 66.2 ms.

grid	glyphs	flat	pre	static	anim	anim−pre	anim %60Hz
112×46	1973	512	768	44138	1467	699	8.80%
152×63	3957	993	1502	94988	2733	1232	16.40%
228×90	8893	2123	3254	210804	5831	2577	34.98%

µs/frame. Two controls, because one can't interpret the animated arm: flat = one colour for every cell (Chromium short-circuits the fillStyle setter → glyph raster alone); pre = distinct precomputed colour per cell → raster + the per-cell CSS colour parse any multi-colour banner pays. anim−pre is the honest LUT cost; anim−flat would charge the animation for parse work the static path pays too.

The animated banner is 11–36x cheaper than the static banner already on main. The static path solves ramp() per cell — a ≤200-iteration lift loop plus a 28-step bisection — so shipped static paint is 44–211 ms/frame. It only survives because it paints once. The precompute is a large improvement to the banner's existing cost, not just a budget for the animation.

A defect the decomposition caught

The first receipt had 228×90 at 101.6% of a 60Hz budget, 80% of it my own LUT lookup. Cause: mix() re-parses hex per call and hex2rgb(b) sits inside its per-channel map, so a bilinear field lookup cost ~12 string parses + 3 hex formats per cell per frame. Fix: parse the tables to numeric mirrors once at build time and blend in RGB — 13.6 ms → 2.6 ms of lookup at 228×90.

The intermediate rgb2hex rounding is load-bearing, not incidental: the old nested-mix path quantized its intermediate blend before the outer blend consumed it, and the WCAG/dE00 certificates were measured against those bytes. The new path reproduces that rounding at the same point, verified byte-identical over 20.4M lookups × 62 themes (COMBINED 96cb28f6…). The oracle's own negative control — dropping only the intermediate quantization — moves the digest to d5be7e54…, so it can fail.

Verification

• Full desktop suite 4011/4011 pass, 0 fail; tsc --noEmit clean; biome clean on every file this PR touches.
• Mutation battery 9/9 killed. It found three gate defects, all fixed: a vacuous reduced-motion gate that compared the static path against itself; an unasserted intensity clamp; and sampleHead(hue)→sampleHead(t), which killed the wordmark wave entirely while every gate stayed green — the hue function and the table were each tested in isolation but nothing asserted the painter composes them. Not an equivalent mutant: head cells have 49 distinct t.
• git merge-tree vs Mari's fix(desktop): retain terminal focus on viewport click #4528 (b76a9d605): 0 conflicts — we both touch TerminalSubstrate.tsx, so textual cleanliness isn't enough. I built the merged tree (20502735…) and ran it: tsc rc=0 and the full suite 4011/4011. Her change is JSX at ~L517, mine is an effect at ~L343.

Wren reviews at exact SHA 8d97ffde44e738dff10193f75cda693cf9fbc514. Max integrates.