production

DESIGN_SYSTEM.md

@@ -185,18 +185,79 @@ Pinned by: `home_gutter_test` (hero and route share an edge),
 group), `activity_list_test` (a sheet is `working` or `canvas`, never the
 warm off-white).
 
+## 6b. The third pass: the glass came out, and forms became bands
+
+Two changes, made together because the second one is only possible once the
+first is done.
+
+### The glass is gone
+
+Nothing in `lib/` is translucent any more. What was removed, and why each one
+was not worth what it cost:
+
+| Was | Now | Why |
+|---|---|---|
+| `milerGlassSurface()` — `BackdropFilter` at sigma 18 on every bar | **deleted** | a full-width blur re-samples everything behind it every frame, on the mid-range Android this ships to, over a map that repaints as it pans. It was already rationed to sigma 18, which is the tell that it was tolerated rather than enjoyed. |
+| `glassCard` `#EBFFFFFF` (92% white) | `#FFFFFF` | a translucent card takes its contrast from whatever happens to be behind it — the same address is crisp over the canvas and grey over a map. |
+| `glassSheet` `#E0FDFDFE` (88%) | `#FFFFFF` | every point of transparency was contrast taken off an address read one-handed at a doorstep in sun. |
+| `glassCardLive` `#0D960019` laid *over* the card | `#FAF2F3`, the same tint flattened | with an opaque surface the overlay would cover the content rather than warm it. A live card is a card with a different fill. |
+| `glassRim` `#8CFFFFFF` — a specular highlight | `#E5E7EB`, a hairline | the highlight only read as an edge because the surface under it was see-through. |
+| `glassRed` | **deleted** | its only caller was the blur. |
+
+`ColorConstants.tint(accent, amount, {on})` replaces the
+`accent.withValues(alpha: …)` idiom wherever it was painting a *surface*.
+It resolves the composite once against a named ground and returns an ordinary
+opaque colour. Same appearance where the old code was already over white;
+predictable everywhere else. Borders and scrims keep their alpha — those are
+lines and dims, not surfaces.
+
+### Forms are bands, not cards
+
+`MilerBand` (`app_widgets.dart`). A card is right for a screen showing a *list
+of objects* — a stop, a job, a booking — because the margin is what says these
+are separate things. It is wrong for a screen that is one long thing: a form, a
+review, a receipt. There the page margin buys nothing and costs twice.
+
+- **Width.** 32–40pt off a 390pt phone, spent on two strips of empty ground
+  beside text that is mostly addresses. On the verification screen that is the
+  difference between an address on two lines and one on three.
+- **A second boundary.** The gap between two cards already separates them; the
+  outline says it again.
+
+So a band runs edge to edge, the page's ground showing between one band and the
+next is the only separator, and the margin moves *inside* as padding — air
+around words rather than air around a box. `MilerBand.pad` (20) and
+`MilerBand.gap` (10) are the two numbers, declared once, so every band's left
+edge lines up down the whole flow.
+
+A band cannot carry state in an outline it does not have, so state goes in the
+**fill**, and a finished band adds `rail` — a 3pt accent stripe down its
+leading edge. That is the one mark that survives a phone at arm's length in
+sun.
+
+**Which screens took it.** The whole first-mile flow, which is where forms live:
+`stop_verify` (collect · hand over · confirm), `shipment_capture`,
+`shipment_review`, `collect_payment`, `delivery_proof_page`. Each is now a grey
+page (`daylightSurface`) carrying full-bleed white bands — the inverse of the
+white-page/tinted-panel arrangement §6 landed on, and taken for the width
+rather than for the tone.
+
+Home, Bookings and Deliveries are **unchanged**. They are lists of objects and
+the card is still the right shape there.
+
 ## 7. Known gaps
 
 This was a foundation pass. It is not a completed product-wide redesign.
 
-- **`GlassCard` still carries three separation mechanisms** — fill, border and
-  shadow. Its rim is documented as a fix for shadows crushing on a budget LCD,
-  reported twice from a device, so it was left alone rather than changed on a
-  guess. It should be revisited *on a device* now that the ground works.
+- ~~**`GlassCard` still carries three separation mechanisms.**~~ Closed by §6b:
+  the fill is opaque, and the hairline stays — it is the documented fix for
+  shadows crushing on a budget LCD, reported twice from a device, and nothing
+  about going opaque changes that. Still worth a look **on a device**.
 - **No typography or spacing consolidation.** Arbitrary `.sp` values still exist
   across pages; `MilerType` was not rationalised into a smaller scale.
 - **Icon families are still mixed.** Lucide was adopted for Activity and the
   status vocabulary; Material rounded remains widespread elsewhere.
 - **Account/Profile was not audited or migrated.**
-- **No device or text-scale QA** of the new ground, and **no GPU profiling** of
-  `milerGlassSheet`'s blur — so no blur change was made.
+- **No device or text-scale QA** of the new ground. The blur question is moot:
+  §6b removed every `BackdropFilter` in `lib/`, so there is nothing left to
+  profile.