palette: give the spectral mid and the envelope trace their own roles

Decouples the keyboard strip's mid stop from accent/secondary, which the
deep-teal darkening had inverted. Adds overlay/trace (#816AA6), the first
value to clear 3:1 against the waveform. Replaces the frozen-premise test.

docs/product/visual-design-language.md

@@ -420,10 +420,26 @@ against the WCAG tests — §2.1):
 | `accent/tertiary` | pastel purple | `~194,170,232` (`#C2AAE8`) | categorical role B (tertiary zones, a distinct affordance class) |
 | `accent/hot` | brighter primary tint | `~200,236,178` (`#C8ECB2`) | hover / live / drag feedback (a lighter pastel-lime) |
 
+**Locked values, where they diverge from the starting table (2026-07-31).** `accent/secondary`
+darkened to `#38A8A0` (same hue ~176°, same saturation, lightness 0.68 → 0.44), which put its
+binding limiter on the AA 4.5:1 text-on-fill pair (4.91:1 on `bg/base`) rather than the 3:1
+indicator floor (3.94:1 on `bg/cell`). Two roles were added past the original three, since
+"exactly three accents" is no longer a constraint: `overlay/trace` `#816AA6` (a muted violet for
+marks drawn *over* an accent fill — see §the two-neighbour problem below), and the spectral
+ramp's mid stop, which keeps the original pastel teal `#84D6D0` as its own constant rather than
+aliasing `accent/secondary`.
+
+**The two-neighbour problem.** A mark drawn *over* an accent fill has to read against the fill
+*and* the surface behind it. For the envelope trace those are the lime waveform (relative
+luminance 0.648) and `bg/base` (0.024) — 9.41:1 apart, so `sqrt(9.41)` ≈ **3.07:1 is the most
+any single value can clear against both**. `overlay/trace` sits at that optimum; it has no
+headroom in either direction, and no value at all can clear 3:1 against the lime *and* on
+`bg/cell`. Such a role is confined to the band it was picked for.
+
 These are light-ish, low-saturation tints. On the **REAPER-grey surfaces** (`bg/cell` ≈
 `#3a3a3a`) they still clear **AA-large (3:1)** and the **state-indicator** floor — but the
-margin is **much smaller than it was on near-black** (roughly ~6:1–7:1 on grey vs. ~15:1
-on `#12121x`). This is a real DS-2-revision tension: **the greyer background pulls the
+margin is **much smaller than it was on near-black** (~7.6:1 primary / ~5.5:1 tertiary /
+~3.9:1 the darkened secondary on grey, vs. ~15:1 on `#12121x`). This is a real DS-2-revision tension: **the greyer background pulls the
 pastels toward the floor** at the same time the "keep it soft" rule pulls them away from
 saturation. Approach from the **soft side** (§2.1) — keep them as pastel as possible while
 still clearing the floor on **grey, not near-black** — but **if any pastel used as a state
@@ -470,9 +486,12 @@ the rest of the UI. Concretely: the ramp runs **pastel-lime (primary, low) → p
 (secondary, mid) → pastel-purple (tertiary, high)** as a three-stop gradient through the
 accent constants (or a slightly wider pastel arc that passes *through* those three anchor
 points), keeping every stop in the pastel band. This ties the spectrum to the palette:
-the same three hues that mean "live / category A / category B" elsewhere are the endpoints
-and midpoint of the spectrum here, so the strip reads as an extension of the accent system,
-not a separate neon flourish. The **active** zone still lifts to `accent/primary` +
+the same hues that mean "live / category A / category B" elsewhere anchor the spectrum here,
+so the strip reads as an extension of the accent system, not a separate neon flourish.
+**The mid stop is its own constant, not an alias of `accent/secondary`** (decoupled
+2026-07-31): the ramp is a luminance progression while the accents are categorical roles, and
+darkening secondary for a categorical reason inverted lo→mid→hi. A monotonicity test now guards
+the ordering. The **active** zone still lifts to `accent/primary` +
 its bloom, so "which zone is live" stays unambiguous over the categorical spectral bands.
 
 - **Feel:** premium, almost visualizer-grade — but soft and cohesive, a pastel spectrum