diff --git a/docs/DESIGN-LANGUAGE.md b/docs/DESIGN-LANGUAGE.md
index ac305e18..6e0c1df8 100644
--- a/docs/DESIGN-LANGUAGE.md
+++ b/docs/DESIGN-LANGUAGE.md
@@ -135,12 +135,19 @@ from the active terminal theme (`--bg --fg --accent --red --green --yellow
   layouts left `--seam-recessed` and joined them (DL-18.6), because two
   boundaries in one window drawn in two different greys made the quieter one
   read as an artefact rather than as something somebody chose.
-  **Amended 2026-09-11, owner:** the pane split alone left `--seam-divider`
-  for `--seam-split` — 20% of `--tone`, `--hair-strong`'s weight — and went
-  from 1px to 2px. At 12% and one pixel a grid of panes still read too faint,
-  while the strip's bottom edge did not, so the split is now the one line
-  inside the work area drawn heavier than the rest. Its hover and drag accent
-  is unchanged.
+  **Superseded for the pane split, 2026-09-18, owner: there is no split line
+  at all.** The 2026-09-11 amendment had taken it to `--seam-split` (20% of
+  `--tone`, `--hair-strong`'s weight) at 2px because a grid of panes read too
+  faint at 12% and one pixel. Panes are now rounded cards separated by the
+  stage gutter (DL-18.12), so the boundary is a background STEP — the thing the
+  split never had — and a drawn line beside a gap would read as three surfaces
+  instead of two. `--seam-split` keeps its 20% and its job — it **moved from
+  the line between two panes to the 1px edge around each one**, because
+  `--sidebar-bg` sits only a few luminance units off `--bg` and on a dark theme
+  the step alone left a card's corner legible and its straight runs invisible
+  (owner, in the dev app, same day). The divider element remains as the drag
+  target, transparent until its hover and drag accent, which is unchanged. The
+  tab strip's bottom edge keeps `--seam-divider` (DL-18.6).
   `--seam-raised` frames a surface that floats above chrome (popovers,
   dialogs).
   **The step must stay louder than the seam that marks it.** Before this rule a
@@ -1434,13 +1441,19 @@ window's identity and its actions at the same time.
   absolute.** A resize seam lights its own hairline in full `--accent` while
   hovered, and for the whole of a drag — a target that says nothing until the
   column moves is found by guessing, and the app already answered this for pane
-  splits (`.split__divider:hover`). Full accent rather than that divider's 60%
-  mix: a shell seam has a chrome surface on at least one side, where the mix
-  reads as a lighter hairline instead of a lit line; a split divider stands
-  between two terminals on `--bg` and keeps its 60%. The line is 1px and sits ON the existing hairline
-  rather than beside it, so nothing new is drawn; only the colour of a line
-  that was already there changes, which keeps DL-2.3's one-boundary-one-seam
-  reading intact. The target itself also widened from 7px to 9px, for a
+  splits (`.split__divider:hover`). **Full accent, and since 2026-09-18 the
+  split divider takes the same value:** the 60% mix was justified by a
+  difference between the two — a shell seam has chrome on one side, a split
+  stands between two terminals on `--bg` — that DL-18.12 removed, because both
+  are now a 4px chrome gutter between two `--bg` cards. One idiom, one value.
+  **What lights is the GUTTER, not a hairline (2026-09-18, owner).** Until then
+  the lit line was 1px sitting ON the shell's own hairline, so that hovering
+  changed the colour of a line already there; once the stage stopped drawing
+  that hairline (DL-18.12) the same placement read as *grabbing the sidebar's
+  own border*, which is what the owner reported. The grip now fills
+  `--stage-gutter` at `--radius-flat`, so the bar the user grabs sits between
+  the two surfaces it moves instead of on the edge of one of them. The target
+  itself also widened from 7px to 9px, for a
   reported defect rather than a preference: along a 7px target the pointer
   crossed the edge repeatedly and the cursor flickered with it. Same figures on
   the docked column's seam (DL-19.4) — one gesture, one look, both edges.
@@ -1559,6 +1572,34 @@ window's identity and its actions at the same time.
   exceptions until 2026-08-25, when DL-27.3's unread ripple became the third;
   they remain the scoped yellow exceptions recorded in DL-3.1/DL-3.2.
 
+- **DL-18.12** **A pane is a card on the stage gutter (2026-09-18, owner).**
+  `--stage-gutter` is 4px and applies to all four edges of the work area AND
+  between panes ([`06-stage-panes.css`](../src/styles/06-stage-panes.css)); the
+  corner is 6px, DL-20.1's `--radius-tab` reused rather than a sixth role
+  opened at a use site. The gutter paints `--sidebar-bg`, the chrome the
+  navigation column already owns (DL-18.7), because the pane and the stage are
+  both `--bg` and a radius between two identical surfaces is invisible;
+  [`derive-colors.test.ts`](../src/lib/derive-colors.test.ts) already
+  guarantees `sidebarBg ≠ bg` in every theme. **That step is not enough on its
+  own, and the card carries a 1px `--seam-split` border as well** — the gutter
+  proved visible and the card's own outline did not. It is a real `border`, not
+  DL-1.3's permitted inset hairline: xterm fills the padding box and would
+  paint over an inset line. This reverses the full-bleed grid of
+  2026-08-17, whose whole argument — the split line must meet the tab strip's
+  hairline — dissolves once the split is a gap rather than a line (DL-2.3). The
+  document surface takes the same gutter and radius so the stage handover does
+  not change the rectangle's shape; the browser tab is the one accepted
+  exception, since its web content is a native view the host paints over
+  [`.browser-panel__view`](../src/browser/browser-panel.tsx) and CSS cannot round
+  it. 4px and not the old 8px
+  because a terminal is measured in cells and the earlier gutter cost columns.
+  **The shell's vertical hairline went with this change.** `.window--sidebar >
+  .stage`'s `border-left` was the one structural line in the shell while it
+  separated `--sidebar-bg` from `--bg`; with the gutter painting
+  `--sidebar-bg` on the stage side, the same 1px ran between two identical
+  surfaces — a line with no step behind it, which DL-2.3 forbids. The rail and
+  the gutter are one chrome plane now, and the card's border is the boundary.
+
 ## 19. Docked side panels
 
 Added 2026-08-12 for the browser panel, and written to cover the class rather
@@ -1730,6 +1771,9 @@ direction token rebuild §9.4
   its own. The tight role joined the earlier control/surface pair on
   2026-08-16, when surface came down from 16px so the closed scale reads
   8/10/12. A value chosen by feel at a use site is not part of this scale;
+  Amended 2026-09-18 (owner): `--radius-tab`'s membership now also covers the
+  stage's panes and document surface (DL-18.12), so the 6px corner is a shared
+  role rather than the strip's private number.
   `border-radius: 50%` and the 999px capsule stay shapes rather than scale
   values.
 - **DL-20.2** One motion pair for chrome state change: `--duration` (150ms) and
diff --git a/src/styles/06-stage-panes.css b/src/styles/06-stage-panes.css
index 55ccd6e2..c46e53c2 100644
--- a/src/styles/06-stage-panes.css
+++ b/src/styles/06-stage-panes.css
@@ -1,5 +1,16 @@
 /* ── Stage (terminal area) ───────────────────────────────── */
 
+/* The gutter the pane-card geometry is derived from (owner, 2026-09-18):
+   4px on every stage edge AND between panes. On `:root` and not on `.stage`
+   because `.pane` and `.split__divider` are also mounted outside a stage — the
+   gallery's seam specimens do exactly that. The matching 6px corner is
+   DL-20.1's `--radius-tab` spelled at each use site rather than aliased here:
+   `design-language.test.ts` reads the literal role names out of the
+   stylesheet, so an alias would read as a radius picked by feel. */
+:root {
+  --stage-gutter: 4px;
+}
+
 .stage {
   position: relative;
   min-height: 0;
@@ -12,14 +23,20 @@
 .stage__tabs {
   position: absolute;
   inset: 0;
-  /* No gutter (owner, 2026-08-17). The 8px this used to carry pushed the pane
-     grid in from every edge, so the vertical split divider started 8px BELOW
-     the tab strip's bottom hairline and the two lines never met — a gap nobody
-     noticed while the divider was 3% and impossible to miss at 12%. Panes are
-     full-bleed now: the split lines run into the strip's line, and the stage
-     reads as one ruled surface instead of a card floating in the window. */
-  padding: 0;
+  /* A gutter on all four edges again (owner, 2026-09-18), replacing the
+     full-bleed grid of 2026-08-17: panes are rounded CARDS floating on the
+     gutter's chrome, so the stage no longer reads as one ruled surface and the
+     split lines no longer have to meet the strip's hairline — the whole reason
+     the 8px gutter was removed. 4px, not the old 8px, so the terminal keeps
+     its columns. */
+  padding: var(--stage-gutter);
   display: flex;
+  /* The gutter has to be a DIFFERENT surface from the pane or the radius is
+     invisible: the pane background and `.stage` are both `--bg`. `--sidebar-bg`
+     is the chrome the rail already paints (DL-18.7), and
+     `derive-colors.test.ts` guarantees it never collapses into `--bg` in any
+     theme. */
+  background: var(--sidebar-bg);
 }
 
 /* ── DL-18.6: the stage's tab strip (sidebar layout only) ─────────────────
@@ -204,15 +221,34 @@
      must still cover the document: the open board (30), Settings (35), the
      modal tier (40) and the popovers (100+). */
   z-index: 20;
-  /* Whatever gutter `.stage__tabs` gives its panes — none, since 2026-08-17 —
-     so the document lands on exactly the rectangle the terminal was occupying
-     instead of shifting by a gutter when the stage changes hands. Both values
-     move together or the handover jumps. */
-  padding: 0;
+  /* Whatever gutter `.stage__tabs` gives its panes — `--stage-gutter` since
+     2026-09-18 — so the document lands on exactly the rectangle the terminal
+     was occupying instead of shifting by a gutter when the stage changes
+     hands. Both values move together or the handover jumps, and the gutter
+     surface has to match for the same reason. */
+  padding: var(--stage-gutter);
   display: flex;
   min-width: 0;
   min-height: 0;
-  background: var(--bg);
+  background: var(--sidebar-bg);
+}
+
+/* The document is a card on the gutter like a pane is, so the handover does
+   not change the shape of the rectangle. Named children, not `> *`: the
+   markdown view-mode toggle is also a direct child and it is a CONTROL, which
+   keeps its own DL-20.1 role. Radius only, no clipping: Monaco's suggest and
+   hover widgets live inside the editor and an `overflow` here would cut them
+   off. */
+.stage__surface > .fileview,
+.stage__surface > .browser-panel,
+.stage__surface > .agent-board {
+  border-radius: var(--radius-tab);
+  /* The pane's edge, for the pane's reason — see `.pane` below. This outranks
+     `.fileview`'s own `--hair` border (`11-settings-screen.css`) only while
+     the file view IS the stage's card; inside the settings screen it stays a
+     line within a surface, which is what `--hair` is for. */
+  border: 1px solid var(--seam-split);
+  box-sizing: border-box;
 }
 
 .stage--strip .stage__surface {
@@ -223,11 +259,13 @@
   right: var(--dock-w);
 }
 
-/* The browser variant of the surface: the native view meets the stage's
-   edges — a web page brings its own margins, unlike the editor the 8px
-   gutter was measured for. */
+/* The browser variant keeps the gutter the base rule gives it (2026-09-18):
+   a page floating on the chrome like every other card is worth more than the
+   few pixels a web page's own margins duplicate. The corners of the web
+   content stay SQUARE and that is accepted — the content is a native view the
+   host positions over `.browser-panel__view`, and CSS cannot round it. */
 .stage__surface--browser {
-  padding: 0;
+  padding: var(--stage-gutter);
 }
 
 /* ── Browser surface: the browser tab's stage cover (DL-18.8) ────────────
@@ -243,6 +281,9 @@
   min-width: 0;
   min-height: 0;
   background: var(--bg);
+  /* The radius comes from `.stage__surface > *`; this is what makes the URL
+     bar's own edges follow it instead of squaring off the top corners. */
+  overflow: hidden;
 }
 
 .browser-panel__bar {
@@ -318,7 +359,9 @@
   inset: 0;
   display: flex;
   z-index: 10;
-  background: var(--bg);
+  /* The gutter's surface, not `--bg`: the zoomed pane is a card on the same
+     chrome as an unzoomed one, so its corners still read as corners. */
+  background: var(--sidebar-bg);
 }
 
 .tab-stage > * {
@@ -365,14 +408,19 @@
   transition: none;
 }
 
-/* A 2px seam between panes — no gap. `--seam-split`, not `--seam-divider`:
-   the owner asked for a heavier split on 2026-09-11 (DL-2.3), and the strip's
-   bottom edge keeps the lighter line. The grab zone is widened past the
-   visible line with a pseudo-element so a thin line stays easy to drag. */
+/* A `--stage-gutter` GAP between panes since 2026-09-18, replacing the 2px
+   `--seam-split` line of 2026-09-11: two rounded cards separated by the
+   gutter's chrome need no drawn seam, and a line plus a gap would read as
+   three surfaces. The divider element stays — it IS the gap, so it keeps
+   owning the drag — and paints nothing until hover. Its grab zone is still
+   widened past its own width with a pseudo-element. */
 .split__divider {
   position: relative;
-  flex: 0 0 2px;
-  background: var(--seam-split);
+  flex: 0 0 var(--stage-gutter);
+  background: transparent;
+  /* The sidebar grip's shape, for the same reason it has it: a 4px accent
+     reads as a bar, and a bar takes DL-20.1's 2px detail corner. */
+  border-radius: var(--radius-flat);
   transition: background var(--duration) var(--ease);
 }
 
@@ -396,9 +444,13 @@
   inset: -3px 0;
 }
 
+/* Full `--accent`, up from a 60% mix (owner, 2026-09-18): the sidebar grip and
+   this divider are now the same object — a 4px chrome gutter between two
+   `--bg` cards — and the 60% was justified by a difference between them that
+   no longer exists (DL-18.9). One idiom, one value. */
 .split__divider:hover,
 .split__divider.is-dragging {
-  background: color-mix(in srgb, var(--accent) 60%, transparent);
+  background: var(--accent);
 }
 
 /* While dragging a divider, block mouse events on terminals so text is not selected */
@@ -415,6 +467,11 @@
   position: relative;
   display: flex;
   flex: 1 1 0;
+  /* Same corner as the pane it holds (DL-18.12): `pingPane` appends
+     `.pane-ping` to the SLOT, not to the pane, so the locator's 2px line would
+     otherwise draw square ends out over the gutter where the pane's curve has
+     already receded. */
+  border-radius: var(--radius-tab);
   min-width: 0;
   min-height: 0;
   overflow: hidden;
@@ -430,6 +487,24 @@
   overflow: hidden;
   background: var(--bg);
 }
+  /* 6px (owner, 2026-09-18). The existing `overflow: hidden` is what makes it
+     visible: xterm's canvases fill the pane and would otherwise paint over the
+     rounded corners, and the same clip keeps `.is-agent-working`'s own
+     top-edge overlays inside the curve. `.pane-ping` is NOT one of them — it
+     hangs off `.pane-slot`, which carries the matching radius above. */
+  border-radius: var(--radius-tab);
+  /* The card's own edge, added the same day the gutter was: `--sidebar-bg` is
+     only a few luminance units off `--bg`, so on a dark theme the gutter alone
+     left the corner legible and the straight runs invisible — the owner saw
+     exactly that in the dev app. `--seam-split` is the token the pane boundary
+     has always owned (DL-2.3); it moved from the line BETWEEN two panes to the
+     edge AROUND each one, so it is the same 20% of `--tone` doing the same
+     job at the new geometry. A real border and not `box-shadow: inset`
+     (DL-1.3's permitted hairline): xterm's canvases fill the padding box and
+     would paint straight over an inset line, while a border sits outside it.
+     The 1px costs the terminal nothing — the grid re-fits to the padding box
+     and a cell is never that small. */
+  border: 1px solid var(--seam-split);
 
 /* The pane's two yellow signals share one visual language but not one life:
    a rail click appends `.pane-ping` for a single 1.5s locator; any pane whose
diff --git a/src/styles/19-agent-board.css b/src/styles/19-agent-board.css
index a0b5e406..95c3e255 100644
--- a/src/styles/19-agent-board.css
+++ b/src/styles/19-agent-board.css
@@ -11,11 +11,10 @@
    are kept, unmatched, as what a revert re-mounts. */
 
 /* The Board's rectangle on the stage is `.stage__surface` itself (spec §4.1),
-   and that class has carried `padding: 0` since 2026-08-17 — so unlike
-   `.stage__surface--browser`, which zeroed a gutter that no longer exists, the
-   Board's modifier needs no declaration of its own. The class stays on the
-   element as the surface's name in the DOM; a rule that only restated the base
-   would be dead the day someone tried to change it. */
+   and it takes that class's `--stage-gutter` and 6px corner unchanged
+   (DL-18.12, 2026-09-18) — the Board is a card on the gutter like a pane or a
+   document. The modifier needs no declaration of its own; the class stays on
+   the element as the surface's name in the DOM. */
 
 .agent-board {
   /* DL-34.5 / DL-4.1: the one place chrome is monospace. */