/*
	SUPERFLY — HOME SECTION 3 — THE TUNNEL (.sftn)
	==============================================
	Built 2026-11-13 to the "goal / tunnel" section on lusion.co, on the
	operator's instruction, to replace the Project AR panel (.sfvp).

	READ docs/research/2026-11-13-lusion-goal-tunnel-section.md FIRST.
	Almost every number in here is measured off their live page or read
	out of their bundle, not invented.

	FIVE ACTS, which is their arrangement. The weights live in
	home-tunnel.js and are the ONE dial that re-paces the whole band:

	    frameIn   2     the camera flies into the panel's screen
	    title     5     the headline spreads, grows and you pass through it
	    tunnel   10     rings rush past
	    frameOut  2     the camera pulls back out into a monitor
	    break   1.5     the glass shatters

	THIS IS NOT THEIR TUNNEL AND CANNOT BE. Theirs is a rigged 3D
	astronaut inside a greebled WebGL corridor with baked animation
	clips, PBR texture sets and a minute of preloading. Ours is CSS 3D:
	a perspective box with recycled rings and one video. The
	CHOREOGRAPHY is theirs; the rendering is not. Do not file the
	difference as a bug.

	=== THE THREE THINGS THAT ARE LOAD-BEARING ===

	1  THE PIN IS `position: sticky`, NEVER A ScrollTrigger PIN. This
	   repo lost a round to a mis-sized `.pin-spacer` giving the whole
	   page a horizontal scrollbar, and page 42 has had ZERO pin-spacers
	   since 2026-11-10. ScrollTrigger only supplies the 0..1 numbers.

	2  THE PANEL MOVES BY TRANSFORM, NEVER BY width/height. Measured on
	   this site in 2026-11-13's notes: driving a full-screen video's
	   width and height costs 1.66ms a frame against 0.23ms for two
	   composed transforms — 7.1x, before the operator's 142% display
	   scaling doubles the raster half again. The frame takes a
	   non-uniform scale and the video inside takes a counter-scale
	   chosen so the two COMPOSE to the uniform scale `object-fit:cover`
	   would have picked. The derivation is in home-tunnel.js.

	3  A UNIFORM RADIUS UNDER A NON-UNIFORM SCALE PAINTS AS AN ELLIPSE.
	   Each axis is divided by its own scale and written as
	   `horizontal / vertical`. The script does it; do not "simplify" it
	   back to one value.
*/

.sftn{
	/* THE RUNWAY. The reference's is 28815px — about 45 viewports —
	   because it is a studio showreel with a 3D tunnel in it. This is a
	   shop, and after the tunnel, the fly-through title, the shatter and
	   then the pull-back all came out on 2026-11-13 it is TWO beats:
	   the panel opens, and the footage holds. 220svh.

	   Raise it and every act stretches in proportion, because the acts
	   are weights in home-tunnel.js, not lengths. */
	/* 320svh, UP FROM 220 (operator, 2026-11-13: "the scroll is too
	   fast, i want slowmotion"). The runway is only half of that
	   change — the other half is the frameIn weight in
	   home-tunnel.js, which decides how much of the runway the
	   opening gets rather than how long the runway is. Both moved,
	   and together they take the open from about 310px of scroll to
	   about 600px at an 846px window. Raise this alone and the band
	   gets longer without the open getting slower; raise the weight
	   alone and the open eats the plateau. */
	/* 310svh. MEASURED OFF LUSION RATHER THAN GUESSED, 2026-11-13.
	   Stepping their astronaut section and reading their own scroll:
	   the tablet is small and inset at secTop -364 and full bleed at
	   -1147, so their opening is 783px of scroll. Ours was 602.

	   IT WENT 220 -> 320 -> 260 -> 310 ACROSS FOUR ROUNDS, and only
	   this one is a measurement. The 320 that was called too slow was
	   too slow because the page was ALSO being thrown 2.4x too far
	   per flick — the distance was never the fault. With the gain
	   fixed, 783px is about three flicks on either site, which is the
	   thing that actually matches.

	   375svh, NOT 310 — AND THE UNIT IS WHY. Their 783px was read at
	   a 640px viewport, which is 1.22 SCREENS of scroll, and this
	   runway is in svh so what has to match is the ratio, not the
	   pixel count. 310svh gave 597px at that same 640 viewport, still
	   a quarter short. The opening is (runway - 100svh) x 4/9, so
	   1.22 screens needs 375. Checked at 640: 782px against their
	   783. */
	--sftn-runway: 375svh;

	--sftn-page:   var(--wp--preset--color--base, #171717);
	--sftn-ink:    var(--wp--preset--color--contrast, #f2f2f4);
	--sftn-accent: var(--wp--preset--color--accent-1, #ec1e79);

	/* 80px at the top end, which is section 2's and section 4's gutter
	   and deliberately NOT the navbar logo's 44 — the bands either side
	   of this one sit at 80 and an 80/44/80 run reads as a mistake. */
	--sftn-gutter: clamp(20px, 4.2vw, 80px);

	/* THE HOUSE RADIUS. The navbar bar, the hero panel and section 2's
	   cards all carry 12px and were standardised onto it on purpose.
	   Change this and `--sf-navbar-bar-radius` together or neither. */
	--sftn-radius: 12px;

	/* ── THE BEZEL AND THE TILT (2026-11-13) ───────────────────
	   Operator, against two screenshots of lusion.co: "why the video
	   doesnt have border like this not border i mean the padding ..
	   and when want to zoom out the video will senget like that".
	   The shape they drew round was the DEVICE BEZEL — the dark band
	   between the tablet's edge and the footage — which is what "not
	   border i mean the padding" is distinguishing it from.

	   LUSION ITSELF IS NOT PORTABLE AND NOTHING HERE PRETENDS IT IS.
	   Their page is ONE FULL-SCREEN WebGL CANVAS, checked 2026-11-13
	   in the operator's own Chrome: zero video elements, zero DOM
	   panels, and their plane BENDS along a curve as it travels,
	   which CSS cannot do at all. What transfers is the read — a
	   flat panel seen off-axis — and that is what these are.

   THE TILT IS A BELL, NOT AN END STATE, and it is a LATE bell:
	   zero at the resting panel, climbing through the whole of the
	   growth, peaking at about 70% open, and collapsing to zero only
	   as the panel lands on full bleed. Two reasons it cannot simply
	   sit at one end, and the second is the one that settles it:

	     ·  A TILT ONLY READS WHERE THE EDGES DO. Full bleed has no
	        visible edges to be off-square, so a tilt there is not a
	        turned panel, it is a distorted picture.
	     ·  AT FULL BLEED IT WOULD ALSO LEAK. Turning a plane away
	        from the camera shrinks its far side, so a tilted panel
	        at full bleed shows the page through two corners unless
	        it is blown up to compensate — and that overscale is paid
	        in sharpness on 720p footage at the one moment it is the
	        whole screen. Peaking in the middle costs neither.

	   Mid-move is also where the operator's second screenshot is:
	   large, turned, edges plainly visible against the page.

	   THE BEZEL RUNS THE OTHER WAY — full at the resting panel,
	   gone by full bleed — for the reason the radius already has:
	   a frame is right once the panel is inset and wrong drawn round
	   the whole screen.

	   TO TURN EITHER OFF: --sftn-bezel to 0px, the three tilts to
	   0deg. The script drops the perspective out of the transform
	   entirely when the tilt is zero, so the band is byte-for-byte
	   what it was. */
	/* OFF, 2026-11-13, on the operator's instruction ("remove the
	   padding"). It is a token rather than deleted rules because the
	   whole mechanism — the ring, the per-axis pre-division, the
	   probe class that resolves this to pixels — costs nothing while
	   it reads 0: the script skips both writes, and the ::after
	   paints a zero-width border. Put a length back here and the
	   bezel returns exactly as it was. */
	--sftn-bezel:     0px;
	--sftn-bezel-col: #14141a;

	/* 1700px is about 50 degrees of field against a panel this size
	   — enough depth to read, short of the fish-eye a tighter camera
	   gives. It is applied INSIDE the device's own transform list,
	   not as a perspective property on the stage: the stage also
	   holds the 24 shatter shards, which already carry their own
	   rotateX/Y/Z and are composed flat on purpose. A perspective on
	   their parent would re-project every one of them. */
	--sftn-persp: 1200px;

	/* MEASURED OFF THEIR OWN PANEL, 2026-11-13, second pass — the
	   first was pitched by eye at 9deg and the operator was right
	   that it read as nothing ("please followww that").

	   Their PLAY REEL panel, caught mid-growth at about 1080px of a
	   1353 window, is a quad with corners near (163,172) (1168,118)
	   (1060,540) (92,450). Its LEFT edge measures about 287px and
	   its right 436 — the left side is 1.52x further from the camera.
	   For a yaw a under a camera at P, that ratio is

	       (P + (w/2) sin a) / (P - (w/2) sin a)

	   which at this panel's half width lands between about 22 and 45
	   degrees depending on how short their camera is. Theirs is a
	   real 3D camera and ours is a CSS one, so the ANGLE does not
	   port directly — the pair (angle, perspective) does. 20deg at
	   1100px reproduces their edge ratio closely, which is the thing
	   the eye actually reads.

	   THE SIGN IS THEIRS TOO: negative yaw turns the LEFT edge away,
	   which is the side their panel recedes on. Flip it and the
	   panel leans toward the copy column instead.

	   THIRD PASS, SAME DAY: THE ROLL IS THE DOMINANT CUE, NOT THE
	   YAW. The operator pointed at the astronaut-tablet section
	   rather than PLAY REEL and said to go and watch it. Driven
	   through its growth in their own Chrome by dispatching wheel
	   events (their scroll is virtual — the document is 640px tall
	   and nothing moves with window.scrollTo), the panel reads:
	   small, flat and framed on the way in; plainly turned at about
	   three quarters open; square again at full bleed. So the LATE
	   bell is right.

	   What was wrong was the mix. Measuring the tablet's right edge
	   in the operator's own screenshot of the peak — it runs from
	   about (1368, 20) to (1262, 900) of a 1901x906 frame — gives
	   atan(106/880) = 6.9 degrees of IN-PLANE ROLL. That single
	   straight leaning edge is what the eye reads as senget; the
	   yaw only supplies the depth behind it. Roll went 1.5 -> 6 and
	   the yaw came down 20 -> 12 to make room for it.

	   NOTE THE CAMERA, 1700 -> 1200. A long camera flattens
	   a yaw into something close to a skew; the depth is most of what
	   makes the move read as a panel turning rather than a picture
	   being squashed. Do not raise the angles without shortening the
	   camera, or the other way round. */
	/* ALL THREE ZERO, 2026-11-13 (operator: "make it match").
	   MEASURED, NOT ASSUMED: stepping lusion.co's own section and
	   shooting it at the small panel, at about 45% open and at full
	   bleed, their tablet is SQUARE in every settled frame. Ours was
	   the only one leaning.

	   The tilt in the screenshot that started this is almost
	   certainly driven by scroll VELOCITY on their side — it leans
	   while you are moving and springs back when you stop — which is
	   why it never appears in a settled shot. Ours leaned by scroll
	   POSITION, so it sat tilted while the page was still, which is
	   a different thing and is what did not match.

	   NOTHING IS DELETED. The script drops the perspective out of the
	   transform entirely while these read 0, so this costs nothing;
	   put degrees back and the lean returns exactly as it was. */
	/* BACK ON, 2026-11-13 (operator: "why in my section video project
	   AR if scroll to full screen the video not senget like before").

	   They were zeroed earlier the same day on "make it match", after
	   I showed that lusion.co's tablet is square in every settled
	   frame. That finding still stands and the note below it is still
	   the reason the reference looks the way it does — but the lean is
	   wanted here, so this is a deliberate departure from the
	   reference rather than an oversight. Do not zero them again
	   without asking.

	   The roll is the one that reads — 6deg of it, measured off the
	   operator's own screenshot of the peak. The yaw and pitch only
	   supply the depth behind it. */
	--sftn-tilt-y: -12deg;
	--sftn-tilt-x: 5deg;
	--sftn-tilt-z: 6deg;

	/* Written by the script, every one of them. Never set by hand. */
	--sftn-f:    0;   /* 0 rest panel  ->  1 full screen              */
	--sftn-g:    0;   /* 0 full screen ->  1 the closing monitor      */
	--sftn-br:   0;   /* the shatter's own 0..1                       */
	--sftn-nav:  90px;/* measured off the real navbar at load         */

	/* THE RESTING PANEL IS SIZED FROM ITS HEIGHT, NOT ITS WIDTH, and
	   that is what keeps the heading and the panel both on one screen.
	   Driven from the width (`min(46vw, 880px)`) the panel grew with
	   the viewport while the screen's HEIGHT did not, so on a wide
	   short window it ate the room the three-line heading needs and the
	   fitted type collapsed. Height first, width derived from the
	   footage's own 16:9 — measured: 361x642 at 1440x860, 378x672 at
	   1900x900, 323x574 at 1024x768, which is near enough one
	   proportion everywhere. */
	/* AND IT IS CAPPED BY WHAT THE HEADING LEAVES, which is new since
	   the heading went back to section 2's ladder. Fitted, the type
	   shrank to suit the panel; laddered, it does not — so at 1900 the
	   112px step stands three lines 336px tall and the panel ran 217px
	   PAST THE BOTTOM of a 900px window. The panel is the thing that
	   gives now, because the type is fixed by the shared set.

	   `--sftn-head-space` is measured by the script. There is no loop
	   here: the heading's height does not depend on the panel. */
	/* The panel's own bottom inset. It does NOT have to equal the side
	   gutter, and at 1900 the gutter is 80px — 80px of the height the
	   112px heading step has already made scarce. */
	--sftn-panel-bottom: clamp(16px, 1.6vw, 28px);

	/* THE PANEL TAKES THE LEFT HALF, WIDTH FIRST.

	   46vw IS THE REFERENCE'S OWN PROPORTION — its tablet measures
	   about 880 of 1908, running from the gutter to just past the
	   middle of the page, with the copy in the right half. This band
	   shipped sized from its HEIGHT instead, to guarantee it cleared
	   the heading, and on a 1886x900 window that produced a panel 37%
	   of the width against the reference's 46 — which is what read as
	   "not 50/50".

	   SO THE HEIGHT IS A CAP ON THE WIDTH, NOT THE OTHER WAY ROUND.
	   Where the window is too short for 46vw at 16:9, the panel gives
	   up WIDTH to keep the ratio rather than being driven from a height
	   that ignores the page's proportion entirely. On a tall window the
	   cap never binds and the panel is exactly the reference's.

	   The ratio is the footage's own 16:9, so `cover` has nothing to
	   crop and the counter-scale in home-tunnel.js has no aspect
	   mismatch to correct. */
	/* THE HEADING IS OUT OF THE STAGE, so the height the panel has to
	   fit into is the whole screen less its own inset top and bottom,
	   not the screen less a heading. --sftn-head-space is still
	   written by the script and is still what sizes the lead block
	   below; it is simply no longer a term here. */
	--sftn-panel-room: calc(
		100svh - 2 * var(--sftn-panel-bottom)
	);
	/* ── THE RESTING PANEL IS 4:3, AND THAT IS WHAT FIXES BOTH
	      COMPLAINTS (2026-11-13) ──────────────────────────────────
	   First "under the titleeee its too many space", then — once the
	   row was lifted to close that — "the space on the bottom is too
	   much". Both are the same fact: a 488-tall panel in an 888 stage
	   leaves 400px of air, and moving it only ever decides which end
	   gets it.

	   THE PANEL WAS 16:9, THE FOOTAGE'S OWN RATIO. On a 16:9-ish
	   screen that is the worst possible resting shape, because panel
	   aspect = screen aspect means it can only be small in BOTH
	   dimensions at once: filling the height would also fill the
	   width and leave no zoom at all. Measured at 1440x900, a 16:9
	   panel tall enough to close the gaps is 84% of the width, which
	   opens 1.19x — not a move.

	   A 4:3 PANEL BREAKS THAT. It fills the height at a much smaller
	   width, so the air goes and the zoom stays: at 1886x888 it rests
	   900x675 with 106px above and below and still opens 2.1x.

	   THE FOOTAGE IS STILL 16:9 AND IS NOW CROPPED AT REST, which is
	   correct and is already handled — object-fit: cover re-crops as
	   the box reshapes from 4:3 to the screen's ratio, and the
	   counter-scale in home-tunnel.js is derived as a max() of both
	   axes precisely so it holds whichever one dominates. It is also
	   what makes the resting panel read as a device rather than as a
	   letterbox, which is what the reference's tablet is.

	   HEIGHT FIRST, WIDTH DERIVED. The height is what the air depends
	   on, so it is the term that is capped; the width follows from
	   the ratio and only has to be stopped from running away on a
	   very wide monitor. */
	/* 76svh / 60vw -> 70svh / 52vw, 2026-11-13 (operator: "the video
	   size also reduce a bit"). BOTH TERMS HAD TO MOVE, because which
	   one binds depends on the window's shape: at 1440x900 the width
	   term is the smaller and at 1900x900 the height term is, so
	   lowering either alone leaves one common desktop unchanged. */
	--sftn-panel-h: max(180px, min(70svh, calc(52vw * 3 / 4)));
	--sftn-panel-w: calc(var(--sftn-panel-h) * 4 / 3);

	/* ── THE COPY COLUMN IS A WIDTH NOW, AND THE ROW IS CENTRED AS A
	      PAIR (2026-11-13) ────────────────────────────────────────
	   Operator: "reduce a width , cause if the width is too high , the
	   button also will follow the width ... but make it that section
	   still center".

	   IT WAS `right: 0` AGAINST THE PAGE GUTTER, so the copy took
	   whatever the panel left and the band always spanned the full
	   page. Shrinking either side would have opened a hole at the
	   right rather than closing in — which is what "still center"
	   is about.

	   So the two are sized first and placed second: the row is panel +
	   gap + copy, and `--sftn-row-x` centres that whole thing, falling
	   back to the gutter on a window too narrow to centre it in.

	   IT IS ALL IN vw, DELIBERATELY. A percentage would resolve
	   against a different containing block in each of the two places
	   that read it — the stage for the device, the intro's padding box
	   for the copy — and the two would drift apart by the gutter. The
	   cost is the scrollbar: the ROW's centring is out by about half
	   of it, roughly 7px, which nobody can see. What must not drift is
	   the gap BETWEEN the panel and the copy, and that cannot, because
	   both are placed off this one token. */
	--sftn-copy-w: clamp(280px, 26vw, 420px);
	--sftn-row-w:  calc(var(--sftn-panel-w) + var(--sftn-colgap) + var(--sftn-copy-w));
	--sftn-row-x:  max(var(--sftn-gutter), calc((100vw - var(--sftn-row-w)) / 2));

	/* ── WHERE THE ROW SITS (2026-11-13) ───────────────────────
	   Operator: "i want the content is centerrrr the screen so the
	   title will be on the top".

	   THE PANEL USED TO CENTRE IN WHAT THE HEADING LEFT, which is
	   not the same thing as centring on the screen and on a 1440x900
	   window was 146px lower: the band under the heading runs 316 to
	   880, so its middle is 596 against the stage's 450.

	   IT IS CENTRED ON THE STAGE NOW, CLAMPED SO IT NEVER RIDES UNDER
	   THE HEADING. The max() is the whole rule, and both halves earn
	   their place:

	     ·  the second term is the centred position;
	     ·  the first is the heading's own floor. The type is fixed by
	        the shared set — the note above says the panel is the thing
	        that gives — so on a window too short to centre a panel
	        this size, it sits as high as the heading allows rather
	        than shrinking.

	   Measured at 1440x900: the floor binds, and the panel still goes
	   from y 410 to y 316 — 94px up, centre 502 against the stage's
	   450. On a taller window the centred term wins outright.

	   THE COPY READS THE SAME TOKEN, in the min-width: 1024px block at
	   the foot of this file, so the paragraphs stay centred on the
	   panel rather than on whatever the grid row happens to be. That
	   is the reference's own arrangement and it is why the two cannot
	   be allowed to drift apart. */
	/* THE HEADING IS NOT IN THIS VIEWPORT ANY MORE (operator,
	   2026-11-13: "i dont want the title in one viewport ... only the
	   video and description , the title is on the top"), so there is
	   no heading floor left to clamp against and the row is simply
	   centred on the stage.

	   That is what the markup change bought. The title used to live
	   inside the sticky stage, which meant it and the panel were
	   always on screen together and the panel could never reach the
	   middle — the heading took 316px of a 900px screen and the max()
	   that used to be here spent the difference. The title is a
	   normal-flow block ABOVE the stage now: you read it, scroll past
	   it, and the pinned screen holds the video and the copy alone.
	   Which is the reference's own arrangement, and the reason its
	   tablet is dead centre. */
	/* ── THE LIFT, AND THE TENSION BEHIND IT (2026-11-13) ─────
	   Operator: "under the titleeee its too many space". Measured at
	   1886x888 with the lead padding already cut to a quarter, the
	   gap from the title's ink to the panel's top edge was 210px —
	   0.236 of a screen, against lusion.co's 0.125.

	   ALL OF THAT REMAINDER IS THE CENTRING. The panel is 488 tall in
	   an 888 stage, so centring it leaves exactly 200px above, and no
	   amount of padding can take that back.

	   AND THE TWO CANNOT BOTH BE HAD. Air above a centred panel is
	   (screen - panel) / 2, so closing the gap means a BIGGER resting
	   panel — and a bigger resting panel is a smaller zoom. At the
	   size that would match their gap outright the panel opens 1.2x
	   instead of 2.2x, which is barely a move at all. The panel is
	   the whole point of the band, so it keeps its size and the row
	   is lifted off centre instead.

	   11vh lands the gap on their 0.125 at this height. What it costs
	   is that the panel no longer sits dead centre once the title has
	   gone — about a tenth of a screen high, with the extra air
	   below. Set this to 0 and the centring is exact again. */
	/* BACK TO 0 — the panel fills the stage now, so there is no
	   surplus to push to one end and the row is centred again, which
	   is what was asked for in the first place. The lift is kept as a
	   token because it is the lever if a heading ever grows. */
	--sftn-row-lift: 0px;
	--sftn-row-top: max(0px, calc(50svh - var(--sftn-panel-h) / 2 - var(--sftn-row-lift)));

	/* THE GAP BETWEEN THE PANEL AND THE COPY. The reference measures
	   150px of 1875 — 8% — and this band shipped with 17%, which is
	   what read as "too much space". It is a real column gap now, not
	   whatever was left over after pushing the text to the right
	   gutter. */
	/* 7vw -> 2.8vw, 2026-11-13, TO BUY THE COPY COLUMN WIDTH. The
	   reference's own 8% is right for its two short paragraphs; this
	   band now carries three built from the product's description, and
	   below about 1500 the GRID TRACK is what sets the measure, not
	   `.sftn__body`'s cap — so the gap is the only term that was
	   actually in play there. At 1440 it is 40px rather than 101, all
	   of which goes to the text. Above 1500 the cap binds instead and
	   this changes nothing. 7vw is the reference's value, kept here. */
	--sftn-colgap: clamp(20px, 2.8vw, 72px);

	/* The house social-rail clearance. `.sfsr` is fixed, 46px wide (40
	   below 782px) and vertically centred on EVERY page; 58 is the
	   number `.sfidx`, `.sfcg` and `.sfbw` all use. It unwinds on its
	   own once the gutter passes 58, so there is no media query. */
	--sftn-rail: max(0px, calc(58px - var(--sftn-gutter)));

	position: relative;
	/* THE LEAD IS ADDED TO THE RUNWAY, NOT TAKEN OUT OF IT. The
	   heading block is in flow at the top of the section, so without
	   this the pinned travel would be shorter by exactly its height
	   and the open would speed up every time the headline gained a
	   line. The script writes the measurement. */
	height: calc(var(--sftn-runway) + var(--sftn-lead, 0px));
	color: var(--sftn-ink);
	isolation: isolate;

	/* ONE SURFACE FOR THE HOME PAGE. Every band on page 42 carries the
	   carbon weave, and a new one copies BOTH halves or it shows a
	   seam: the light diagonal is the sheen on the tow, the dark one
	   the shadow under it, and with one alone it reads as corduroy.

	   THIS BAND SHIPPED WITH THE COLOUR AND NOT THE WEAVE, which is
	   worse than it sounds. A flat `background-color` paints OVER the
	   body's own weave, so it did not read as a missing texture — it
	   read as a rectangle of plain carbon between two woven bands.
	   Measured on the live page: sections 1, 2, 5 and 6 and the body
	   all computed two `repeating-linear-gradient`s, and this one
	   computed `none`.

	   The local harness could never show it: the weave is on the
	   harness's `body`, and a band that paints over it looks identical
	   to one that does not when the band fills the screen. */
	background-color: var(--sftn-page);
	background-image:
		repeating-linear-gradient( 45deg, rgba(255,255,255,.025) 0 2px, transparent 2px 4px),
		repeating-linear-gradient(-45deg, rgba(0,0,0,.35)        0 2px, transparent 2px 4px);
}

/* ---------------------------------------------------------------- *
	THE STAGE
 * ---------------------------------------------------------------- */

/* ── THE LEAD ────────────────────────────────────────────────────
   The heading, in normal flow, ABOVE the sticky stage. Scrolling past
   it is the first beat of the band.

   IT IS NOT IN THE STAGE AND MUST NOT GO BACK IN. The stage is
   sticky and holds the video and the copy; anything in it is on
   screen for the whole pin, which is exactly what the operator asked
   to end. The section's own height adds this block's height on top of
   the runway (see --sftn-lead), so moving the title out here costs
   the pinned travel nothing.

   The script measures this and writes --sftn-lead, because the
   heading is three lines on a stepped ladder and its height is not
   knowable from CSS. The ScrollTrigger start is offset by the same
   number, so the panel does not begin opening until the stage has
   actually pinned. */
.sftn__lead{
	position: relative;
	z-index: 2;
	/* THE BOTTOM PADDING IS THE SPACE UNDER THE TITLE and it was
	   clamp(28px, 4vw, 72px) — 72px on a wide screen, on top of the
	   air the centred panel already leaves below it. Cut to a quarter
	   of that. The rest of the gap is the panel being centred on the
	   stage, which is a separate decision and the lever for it is
	   --sftn-row-top. */
	padding: var(--sftn-nav) var(--sftn-gutter) clamp(6px, .9vw, 18px);
	padding-left: calc(var(--sftn-gutter) + var(--sftn-rail));
}

.sftn__stage{
	position: sticky;
	top: 0;
	height: 100svh;
	overflow: hidden;
	/* `svh`, not `vh` or `dvh` — the small viewport height, the one
	   with the phone's toolbars showing, so the panel fits at the
	   worst moment. Same reasoning as `.sfop`. */
}

.sftn__sr{
	position: absolute;
	width: 1px; height: 1px;
	margin: -1px; padding: 0; border: 0;
	overflow: hidden; clip: rect(0 0 0 0); clip-path: inset(50%);
	white-space: nowrap;
}

/* ---------------------------------------------------------------- *
	ACT 1 — THE RESTING LAYOUT

	Their arrangement: heading across the top, the device left, two
	paragraphs right. Ours is the same on a carbon page.
 * ---------------------------------------------------------------- */

.sftn__intro{
	position: absolute;
	inset: 0;
	padding: var(--sftn-nav) var(--sftn-gutter) var(--sftn-gutter);
	display: grid;
	/* TWO COLUMNS, THE FIRST THE PANEL'S OWN WIDTH PLUS THE GAP. The
	   copy then starts a fixed distance after the panel's right edge
	   rather than being shoved against the right gutter, which is what
	   the reference does and what closes the hole. */
	grid-template-columns: calc(var(--sftn-panel-w) + var(--sftn-colgap)) 1fr;
	grid-template-rows: auto 1fr;
	gap: clamp(16px, 2.4vw, 32px);
	box-sizing: border-box;
	pointer-events: none;
	z-index: 2;
}

/* The heading spans both columns; only the copy sits in the right one. */
.sftn__intro > .sftn__introText:first-child{
	grid-column: 1 / -1;
	grid-row: 1;

	/* THE HAIRLINE UNDER THE HEADING. The reference has none, but it
	   has a pale page and a drawn ribbon doing that work; on carbon the
	   title floated free of the row beneath it. A hairline is the
	   house device — `.sfidx` rules its rows the same way — and it
	   gives the band a top edge without touching the column gap, which
	   is the reference's own breathing room and should stay empty.

	   It is inside `.sftn__introText`, so it fades out with the rest of
	   the copy on `--sftn-f` rather than hanging there over the video. */
	padding-bottom: clamp(16px, 1.8vw, 30px);
	border-bottom: 1px solid rgba(242, 242, 244, .12);
}

/* The text is what leaves; the device is a separate layer that grows,
   so only this block fades. `--sftn-f` is the frame-in, so the copy is
   gone by the time the screen is full. */
/* ONE WORD OF THE HEADING. inline-block, because an inline box
   cannot be transformed and the whole reveal is a transform.

   THE SPACES ARE REAL TEXT NODES BETWEEN THESE, not padding and not
   margins: an inline-block holding only a space collapses to zero
   width, which is the trap navbar.css and the house button both
   record. Leaving the spaces outside the spans means the line still
   breaks exactly where it would have without the split. */
.sftn__w{
	display: inline-block;
	will-change: transform, opacity;
}

.sftn__introText{
	opacity: calc(1 - var(--sftn-f) * 1.6);
	transform: translate3d(0, calc(var(--sftn-f) * -40px), 0);
	will-change: opacity, transform;
}

.sftn__eyebrow{
	margin: 0 0 10px;
	padding-left: var(--sftn-rail);
	font-family: 'Chakra Petch', 'Inter', system-ui, sans-serif;
	font-weight: 500;
	font-size: 11px;
	letter-spacing: .16em;
	text-transform: uppercase;
	color: var(--sftn-accent);
}

/* SECTION 1 AND 2's TYPE, VALUE FOR VALUE — Inter 700, the stepped
   ladder, leading 1, tracking -.075em, uppercase. SIX other
   stylesheets carry these as literals and CLAUDE.md says to change all
   of them or none: home-opening, category-grid, video-panel,
   bike-ribbon, dealer-wall, singapore. This file is the seventh.

   They are literals rather than one shared token on purpose: a custom
   property substitutes at computed-value time on the element where it
   is DECLARED, so a `:root` token declared in one stylesheet and read
   from another resolves by load order and fails silently. */
.sftn__title{
	margin: 0;
	padding-left: var(--sftn-rail);
	font-family: 'Inter', system-ui, -apple-system, sans-serif;
	font-weight: 700;
	line-height: 1;
	letter-spacing: -.075em;
	text-transform: uppercase;
	color: var(--sftn-ink);

	/* SECTION 2's TYPE, ALL OF IT, SIZE INCLUDED (operator, 2026-11-13:
	   "the heading i want my section 2"). Family, weight, SIZE, leading,
	   tracking and case are `.sfcg__title`'s exactly, so this file joins
	   the set CLAUDE.md says to change together or not at all —
	   home-opening, category-grid, video-panel, bike-ribbon,
	   dealer-wall, singapore, and now this one.

	   The ladder is STEPPED, not a clamp, because section 2's is: only
	   the size moves, the leading and tracking being ratios of it.

	   IT WAS FITTED TO THE MEASURE FOR AN HOUR and that is reversed
	   here. What the fit bought was a heading that FILLED its width in
	   three lines; what it cost was being the only band on the home
	   page off the shared ladder. The ladder won.

	   SO THE SENTENCE IS WHAT FILLS THE WIDTH NOW, not the type size.
	   There is NO max-width and the lines are NOT broken by hand — one
	   long sentence wrapping at the measure, which is exactly how
	   `.sfcg__title` works and why it runs edge to edge in two lines.
	   Three hand-broken lines sat at about half the measure at 1440 and
	   read as a small heading in a big space.

	   REWRITE IT AND RE-CHECK THE LINE COUNT at 1280, 1440 and 1900.
	   The 112px step at 1900 is a third bigger than the 72px one, so a
	   sentence that is two lines at 1440 can be three there. 77
	   characters lands on two at both. */
	font-size: 30px;
}
@media (min-width: 768px){  .sftn__title{ font-size: 48px } }
@media (min-width: 1024px){ .sftn__title{ font-size: 60px } }
@media (min-width: 1280px){ .sftn__title{ font-size: 72px } }
@media (min-width: 1900px){ .sftn__title{ font-size: 112px } }

/* ONE LINE PER SPAN, BROKEN BY HAND. Three lines is the arrangement —
   a measure cannot be trusted to produce exactly three at every width,
   and the fit is computed from the longest line, so the break has to be
   known. NEVER a `<br>`: `.sfv` already recorded what a `<br>` does to
   a section that rebuilds its text nodes. */
/* NOTHING PRODUCES `.sftn__tline` ANY MORE. The heading is one
   wrapping sentence and its reveal plays on the block, because
   splitting it put the sheen word on a line of its own —
   home-tunnel.js carries the full reasoning. Kept as a note rather
   than a rule so the next person does not go looking for the
   stylesheet that lost it. */

/* THE SHEEN WORD, PORTED FROM `.sfcg__word` AS-IS. One span, a
   gradient clipped to the glyphs, looping for ever — ink to magenta to
   ink, so it reads as a light passing through the word rather than as
   a second colour. It is NOT a scroll effect and must not be made one.

   TWO THINGS IN IT ARE LOAD-BEARING and both are recorded elsewhere in
   this project from other sections:

   1  `background-clip: text` MUST BE WRITTEN AS BOTH LONGHANDS. With
      only the standard one, Blink paints a solid magenta band with
      invisible type inside it — `reviews.css` lost a round to exactly
      this, and `category-grid.css` records it again.
   2  THE NEGATIVE-MARGIN / POSITIVE-PADDING PAIR gives the last glyph
      the room the background box would otherwise clip. Remove the
      padding and the final letter loses its right edge; remove the
      margin and the word pushes its neighbour along. They move
      together or not at all. */
.sftn__sheen{
	background-image: linear-gradient(
		100deg,
		var(--sftn-ink)    0%,
		var(--sftn-ink)   30%,
		var(--sftn-accent) 50%,
		var(--sftn-ink)   70%,
		var(--sftn-ink)  100%);
	background-size: 300% 100%;
	-webkit-background-clip: text;
	background-clip: text;
	color: transparent;
	padding-right: .08em;
	margin-right: -.08em;
	animation: sftn-sheen 4.5s linear infinite;
}

@keyframes sftn-sheen{
	from{ background-position: 150% 0 }
	to  { background-position: -150% 0 }
}

@media (prefers-reduced-motion: reduce){
	.sftn__sheen{ animation: none; background-position: 50% 0 }
}

.sftn__body{
	/* CENTRED ON THE ROW, NOT BOTTOM-ALIGNED. The reference sets its
	   two paragraphs against the middle of the device, not against the
	   foot of the screen — bottom-aligned they read as a caption to the
	   section rather than as the copy beside the panel. */
	grid-column: 2;
	grid-row: 2;
	align-self: center;
	/* THE COPY COLUMN IS WIDER THAN THE REFERENCE'S NOW (operator,
	   2026-11-13: "i want you added the width , cause the width is too
	   small"), and the two numbers below are the whole of it.

	   IT WENT 46ch -> 52ch -> 66ch. The first two steps were chasing
	   the reference's own 28.3% of the page, which is right for two
	   short paragraphs and wrong for what this band says now: the copy
	   was rewritten from product 983's own description the same day
	   and went from two paragraphs to three. At 52ch that set eleven
	   lines in a column narrower than the video beside it, which reads
	   as a newspaper column rather than as the other half of a layout.

	   THE RIGHT-HAND PADDING IS THE OTHER HALF OF THE WIDTH, and it is
	   easy to raise the cap and wonder why nothing moved: the column
	   is a 1fr track, so above about 1500 the cap is what binds and
	   below it the padding is. Both had to come down together — 140px
	   was the reference's ~12% of the page, and 80px is still enough
	   that the text does not run at the right gutter.

	   52ch and 140px are the reference's values and are kept here so
	   going back is two numbers. */
	/* A WIDTH, NOT A MEASURE IN ch, SINCE 2026-11-13. The column is
	   placed rather than filled now (see --sftn-copy-w), so the cap and
	   the layout are the same number and cannot disagree. `right: 0` in
	   the desktop block still backs it off the gutter on a narrow
	   window, where this cap is not what binds. */
	max-width: var(--sftn-copy-w);
	padding-right: 0;
}

.sftn__p{
	margin: 0 0 16px;
	font-family: 'Inter', system-ui, sans-serif;
	font-weight: 500;
	font-size: clamp(14px, 1.05vw, 17px);
	line-height: 1.55;
	color: rgba(242,242,244,.82);
}
.sftn__p:last-of-type{ margin-bottom: 0 }

/* THE BUTTON UNDER THE COPY, and it is not only a route to the
   product. The right column was 170px tall against the panel's 330 at
   1353x640, so the copy floated with 65px of nothing above it and 95
   below — which is what read as the section being unfinished. The
   button takes the column to about 250 and the two sides balance. */
.sftn__body .sftn__btn{
	margin-top: clamp(20px, 2.4vw, 34px);
	/* THE BUTTON SHRINKS TO ITS LABEL (operator, 2026-11-13: "if the
	   width is too high , the button also will follow the width").

	   `.sftn__btn` is `display: inline-flex`, which shrink-wraps in a
	   block — but `.sftn__body` is a FLEX COLUMN, and a flex item
	   stretches across the cross axis by default whatever its own
	   display is. So the button grew with the copy column, which is
	   the one thing in this band that should not. */
	align-self: flex-start;
}

/* Below 1024 the two columns stack and the paragraphs go under the
   panel, which is where the reference puts them on a phone too. */
@media (max-width: 1023px){
	.sftn__intro{ grid-template-columns: 1fr }
	.sftn__body{
		grid-column: 1;
		align-self: end;
		max-width: none;
		padding-right: 0;
	}
	.sftn__p{ font-size: 14px }
	.sftn__p:nth-child(n+2){ display: none }
}

/* ---------------------------------------------------------------- *
	THE DEVICE

	Its LAYOUT box is the resting panel and never changes — that is
	what makes the whole move free. Everything else is transform.
 * ---------------------------------------------------------------- */

.sftn__device{
	position: absolute;
	/* THE ROW IS CENTRED AS A PAIR — see --sftn-row-x at the top of
	   this file. It falls back to the gutter on a window too narrow to
	   centre the row in, which is what this used to be. */
	left: var(--sftn-row-x);
	/* IT CENTRES IN WHAT IS LEFT UNDER THE HEADING, not in the stage.
	   Centred in the whole stage it rode up behind the headline; the
	   reference puts the device below the heading and the paragraphs
	   beside it. `top`/`bottom` + `margin: auto` is the absolute-
	   centring trick, and the band it centres in is the one the intro
	   leaves. */
	/* THE TOP OF THE PANEL'S BAND IS MEASURED, because the heading's
	   height is now fitted and so is not knowable from CSS. The script
	   writes `--sftn-head-space` (the eyebrow and the three lines, plus
	   their gap) after every fit. Hardcoded, the panel rode up over the
	   heading the moment the type got bigger — which is exactly what
	   happened on the first pass of this change. */
	/* CENTRED ON THE STAGE, FLOORED BY THE HEADING — the token is
	   argued at the top of this file. It was top/bottom + margin:auto,
	   which centres in the band the heading leaves rather than in the
	   screen; --sftn-panel-bottom is now only the height cap's own
	   term and no longer positions anything. */
	top: var(--sftn-row-top);
	bottom: auto;
	margin: 0;
	/* 46vw IS THE REFERENCE'S OWN PROPORTION — their tablet measures
	   about 880 of 1900. The cap is what stops it running away on a
	   very wide monitor. */
	/* 16:9 IS THE FOOTAGE'S OWN RATIO, so `cover` has nothing to crop
	   and the counter-scale in home-tunnel.js has no aspect mismatch to
	   correct. Both terms come off one token so the copy column beside
	   it can be placed from the same number. */
	height: var(--sftn-panel-h);
	width: var(--sftn-panel-w);
	z-index: 3;
	/* NO TRANSFORM HERE, AND THAT IS DELIBERATE. The script reads this
	   element's `offsetLeft`/`offsetTop`/`offsetWidth`/`offsetHeight`
	   to learn the resting rect, and those are LAYOUT values that
	   ignore a transform — but a transform written in CSS would still
	   have to be composed back into every frame the script writes, and
	   the first thing to go wrong would be a panel half a height out.
	   Centre it with `margin: auto`, not with `translate`.

	   The script replaces nothing until it runs, so with no JS at all
	   the band is a video beside some text. That is a perfectly good
	   fallback rather than a blank screen. */
}

@media (max-width: 1023px){
	.sftn__device{
		left: var(--sftn-gutter);
		right: var(--sftn-gutter);
		width: auto;
		height: auto;
		aspect-ratio: 16 / 9;
	}
}

.sftn__frame{
	position: absolute;
	inset: 0;
	overflow: hidden;
	border-radius: var(--sftn-radius);
	background: #000;
	transform-origin: 50% 50%;
	will-change: transform, border-radius;
	/* THE BEZEL AND ITS LIGHT. Theirs is a signed-distance rounded box
	   drawn in a fragment shader with a real light position; this is
	   the two-shadow emboss `.sfdw` and `.sfcg` already share, which is
	   the closest thing this site has to it and costs nothing. */
	box-shadow:
		-4px -4px  8px rgba(255,255,255,.10),
		 5px  5px 12px rgba(0,0,0,.70);
}

/* ── THE BEZEL ───────────────────────────────────────────────────
   A RING, NOT PADDING, and that is why it can exist in this band at
   all. Padding, a border or an inset box on the frame are all
   LAYOUT, and this whole section is built on the device's layout box
   never changing — the note on .sftn__device says so in as many
   words. This is one absolutely positioned pseudo-element with a
   border, painting over the outer band of footage. Nothing lays out.

   IT COSTS A BAND OF PICTURE rather than shrinking the picture,
   which is what a real bezel does: the screen is a window cut in the
   device, not the device drawn around a screen.

   THE WIDTHS ARE PRE-DIVIDED PER AXIS, exactly as the radius in
   home-tunnel.js is and for the same reason — these are LOCAL widths
   on a box that is about to be scaled non-uniformly, so one value
   would paint a thicker band on one axis than the other. The script
   writes b/sx and b/sy, which paint as one thickness; under the tilt
   it then foreshortens across the panel, which is what a real bezel
   does.

   border-radius: inherit copies the frame's own two-axis radii, so
   the outer corner matches and the INNER corner falls out of the
   border algorithm — outer less the width, per axis, which is the
   device corner for free.

   IT RIDES THE FRAME'S OWN OPACITY, so it dissolves with the frame
   when the glass goes and never outlives the panel it frames.

   A PSEUDO-ELEMENT CANNOT TAKE AN INLINE STYLE, which is the whole
   reason both widths arrive as custom properties written on the
   frame and read here. .sfbr__tag records the same constraint. */
.sftn__frame::after{
	content: "";
	position: absolute;
	inset: 0;
	z-index: 2;
	pointer-events: none;
	border-style: solid;
	border-color: var(--sftn-bezel-col);
	border-width: var(--sftn-by, 0px) var(--sftn-bx, 0px);
	border-radius: inherit;
}

/* MEASURING ONLY — never on screen. --sftn-bezel is a clamp() and a
   custom property is NOT computed to pixels, so getPropertyValue
   hands the script the token text and parseFloat on it returns the
   FLOOR, 8, silently and at every width. home-tunnel.js adds this
   class for one frame inside measure() and reads the ::after's
   borderTopWidth, which IS a used length. It is the same trick the
   file already uses to read the gutter off .sftn__intro's padding
   rather than off the token. */
.sftn__frame.is-sftn-measuring::after{
	border-width: var(--sftn-bezel);
}

.sftn__video{
	position: absolute;
	inset: 0;
	width: 100%;
	height: 100%;
	object-fit: cover;
	transform-origin: 50% 50%;
	will-change: transform;
	display: block;
}

/* ---------------------------------------------------------------- *
	ACT 5 — THE SHATTER

	Theirs is a `broken_glass_animation.buf` — a real simulated break.
	Ours is 24 clip-path polygons over the closing panel, flying out on
	a seeded random spread so every visitor sees the same break and the
	band can be screenshotted twice and compared.
 * ---------------------------------------------------------------- */

.sftn__shards{
	position: absolute;
	inset: 0;
	z-index: 7;
	pointer-events: none;
	transform-style: preserve-3d;
	opacity: var(--sftn-br, 0);
}

.sftn__shard{
	position: absolute;
	background:
		linear-gradient(135deg,
			rgba(236,30,121,.55) 0%,
			rgba(255,255,255,.75) 42%,
			rgba(120,190,255,.55) 100%);
	will-change: transform, opacity;
	backface-visibility: hidden;
}

/* ---------------------------------------------------------------- *
	THE CLOSING BLOCK

	Theirs ends on "CONTINUE TO SCROLL". This band is replacing the
	Project AR panel, so it ends where that one did — the statement and
	the pill to the product.
 * ---------------------------------------------------------------- */

.sftn__say{
	position: absolute;
	inset: auto 0 0;
	padding: 0 var(--sftn-gutter) clamp(40px, 7vh, 96px);
	display: grid;
	justify-items: center;
	gap: clamp(18px, 2.6vh, 34px);
	text-align: center;
	z-index: 8;
	opacity: var(--sftn-out, 0);
	transform: translate3d(0, calc((1 - var(--sftn-out, 0)) * 28px), 0);
	pointer-events: none;
}
.sftn__say.is-live{ pointer-events: auto }

.sftn__sayText{
	margin: 0;
	max-width: 22ch;
	padding-left: var(--sftn-rail);
	font-family: 'Inter', system-ui, sans-serif;
	font-weight: 700;
	font-size: clamp(22px, 3.1vw, 54px);
	line-height: 1.04;
	letter-spacing: -.035em;
	color: var(--sftn-ink);
}

/* The accent words. Instrument Serif italic is section 3's own face
   and arrives on this band's own handle — `superfly-hero-fonts` is
   registered 43 times and the first registration wins, so adding a
   family to it silently does nothing. Georgia is the fallback and is
   about 15% narrower, so a missing enqueue reads as type that is
   merely mis-set rather than as an error. */
.sftn__sayText em{
	font-family: 'Instrument Serif', Georgia, serif;
	font-style: italic;
	font-weight: 500;
	letter-spacing: -.01em;
	color: var(--sftn-accent);
}

/* ── THE HOUSE BUTTON (chosen 2026-11-13) ────────────────────────
   Design 01 of the five proposed at docs/home-concepts/
   button-designs.html, at a 12px radius — both the operator's call
   ("i want 12 border radius and i will used option 1").

   12px IS THE HOUSE RADIUS and that is why it was offered: the navbar
   bar, section 3's video panel, the category cards and the dealer
   cards all carry it. The button was a 999px pill, which is the one
   shape on this page that matched nothing else on it.

   THE LABEL ROLLS, LETTER BY LETTER. Every character is two stacked
   copies in its own clipped column; on hover the pair slides up one
   line, staggered 22ms apart, so the word retypes itself left to
   right. The arrow does the same thing sideways in its white disc —
   one leaves right as its twin arrives from the left.

   IT IS THE NAVBAR'S OWN DEVICE, not a new one. navbar.css has rolled
   its menu labels this way since 2026-11-10, which is the whole
   argument for this one over the other four: the site already does
   it, so the button reads as part of the same object rather than as
   an effect bolted to it.

   ── TWO THINGS THAT ARE LOAD-BEARING, both already recorded in
      navbar.css from its own roll ──

   1  white-space: pre ON EVERY CHARACTER. Each one is an
      inline-block, and an inline-block holding only a space collapses
      to zero width — "Shop Project AR" renders as "ShopProjectAR"
      without it.
   2  THE CLIP IS overflow: hidden ON A BOX ONE LINE TALL, and the
      line-height and that height must agree. At 1.15em both the
      resting copy and its twin sit exactly one line apart, so the
      roll lands on the baseline it left. Change one and change the
      other.

   THE MARKUP IS BUILT BY SCRIPT, so there is nothing to write to the
   database for it — enhanceButtons() in home-tunnel.js splits the
   label and injects the disc on load. With no JS at all the anchor is
   still a solid magenta button with its label, which is the right
   failure.

   ── WHEN THIS GOES SITE-WIDE ──
   It is scoped to .sftn__btn for now. Standardising it means moving
   these rules to a shared stylesheet under one class and pointing
   .sfvp__btn, .sfdw__cta and the rest at it — a pass of its own,
   because each of those carries its own size and placement. */
.sftn__btn{
	position: relative;
	display: inline-flex;
	align-items: center;
	justify-content: center;
	gap: 14px;
	height: 58px;
	padding: 0 32px;
	border: 0;
	border-radius: 12px;
	background: var(--sftn-accent);
	color: #fff;
	text-decoration: none;
	font-family: 'Inter', system-ui, sans-serif;
	font-weight: 700;
	font-size: clamp(11px, 1vw, 14px);
	letter-spacing: .1em;
	text-transform: uppercase;
	white-space: nowrap;
	overflow: hidden;
	transition: background-color .3s ease;
}

/* The clipped line the columns roll inside. */
.sftn__btn-roll{
	display: inline-flex;
	overflow: hidden;
	height: 1.15em;
	line-height: 1.15em;
}
.sftn__btn-col{ display: inline-block; position: relative }
.sftn__btn-col > span{
	display: block;
	white-space: pre;
	transition: transform .5s cubic-bezier(.76,0,.24,1);
	transition-delay: calc(var(--i, 0) * 22ms);
}
.sftn__btn-col > span + span{ position: absolute; left: 0; top: 100% }

/* The disc. Both arrows share one grid cell, so the outgoing and the
   incoming one are on the same pixel and the swap cannot drift. */
.sftn__btn-pip{
	position: relative;
	flex: 0 0 auto;
	width: 26px; height: 26px;
	border-radius: 50%;
	background: #fff;
	display: grid;
	place-items: center;
	overflow: hidden;
}
.sftn__btn-pip svg{
	grid-area: 1 / 1;
	width: 12px; height: 12px;
	fill: none;
	stroke: var(--sftn-accent);
	stroke-width: 2.2;
	transition: transform .5s cubic-bezier(.76,0,.24,1);
}
.sftn__btn-pip svg + svg{ transform: translateX(-26px) }

@media (hover: hover) and (pointer: fine){
	.sftn__btn:hover{ background: #f5318a }
	.sftn__btn:hover .sftn__btn-col > span{ transform: translateY(-100%) }
	.sftn__btn:hover .sftn__btn-pip svg{ transform: translateX(26px) }
	.sftn__btn:hover .sftn__btn-pip svg + svg{ transform: translateX(0) }
}

@media (prefers-reduced-motion: reduce){
	.sftn__btn, .sftn__btn *{ transition: none !important }
}

/* ---------------------------------------------------------------- *
	REDUCED MOTION

	The whole band is one scrubbed move, so there is nothing to tone
	down — it collapses to its resting state: the panel, the heading,
	the paragraphs and the button. One screen, no runway.
 * ---------------------------------------------------------------- */

/* THE COPY IS TIED TO THE PANEL'S BAND, not to the grid row.

   The grid gives row 2 everything under the heading, so a copy block
   centred in it sits wherever that leftover happens to be — at
   1440x900 that was a centre of 580 against the panel's 596, and once
   the panel moved up to centre on the stage the two would have been
   78px apart. The reference sets its paragraphs against the middle of
   the tablet, so here they are absolutely positioned over exactly the
   panel's band and centred in it.

   DESKTOP ONLY. Below 1024px the panel is full width and the copy
   stacks under it, which is the max-width: 1023px block above; this
   must not reach that layout. Both rules are plain classes, so this
   one wins there by source order alone — keep it below that block. */
@media (min-width: 1024px){
	.sftn__body{
		/* AN ABSOLUTELY POSITIONED GRID ITEM IS LAID OUT AGAINST ITS
		   GRID AREA, NOT THE GRID CONTAINER, and that cost a deploy:
		   with the body still in column 2 / row 2, a left of
		   gutter + panel + gap was measured from the column's own
		   left edge and the copy landed at x 1679 on a 1440 screen,
		   101px wide and off the page.

		   Spanning the whole grid makes the containing block the
		   grid's CONTENT box — which is the intro's padding box, so
		   it already starts at the gutter and under the navbar. Every
		   offset below is therefore relative to those two, which is
		   why the gutter is not added again and why the top subtracts
		   the navbar. */
		grid-column: 1 / -1;
		grid-row: 1 / -1;
		position: absolute;
		/* The gutter comes off because this box's containing block is
		   the intro's PADDING box, which already starts there, while
		   --sftn-row-x is measured from the screen's edge. */
		left: calc(var(--sftn-row-x) - var(--sftn-gutter)
		           + var(--sftn-panel-w) + var(--sftn-colgap));
		right: 0;
		top: calc(var(--sftn-row-top) - var(--sftn-nav));
		height: var(--sftn-panel-h);
		display: flex;
		flex-direction: column;
		justify-content: center;
	}
}

@media (prefers-reduced-motion: reduce){
	.sftn{ --sftn-runway: 100svh }

	/* THE TILT GOES, THE BEZEL STAYS. A panel swinging through nine
	   degrees of yaw is motion in the sense this setting is asking
	   about; a dark band round the footage is a look. Zeroed here,
	   the script drops the perspective out of the transform list
	   altogether, so there is not even a 3D context left behind. */
	.sftn{ --sftn-tilt-y: 0deg; --sftn-tilt-x: 0deg; --sftn-tilt-z: 0deg }
	.sftn__shards{ display: none }
	.sftn__say{
		position: static;
		opacity: 1;
		transform: none;
		pointer-events: auto;
		justify-items: start;
		text-align: left;
		padding-top: 24px;
	}
}
