Thread a box in pixels down the draw, one multiply from its parent's

A box in pixels was composed back up the move chain, on a grid fine enough
that the walk rounded once, while a widget's offer was threaded down through
its ancestors' offers. Two routes to one length, which is what
`Holds::through` allowed for -- and the offer's route broke at a region node.
`offered_region` fell back to `UiRegion::FULL` there, and `redraw` resolved
that against the node's slot entry, which holds the box its parent *placed*
the node in. Under a `Scroll` that is as long as the content rather than the
viewport, so everything below was re-asked at a width its own answer had
produced and the old answer confirmed itself: shrinker seed 220 on `reorder`
left a widget 290px out.

`ActiveData` now keeps a widget's box as lengths of its parent's box --
`given_len`, and `offer_len` for the box it was first asked about -- and
`DrawInfo` carries the pixel lengths, threaded down one `Len::to_px` at a
time: the box its parent gave it, then the part of that box its own answer
placed the drawing in, which `placed_lens` states once for both `placed_box`
and the walk. `Painter::px_size` and `px_len` read that value, and
`UiRenderState::asked_px` takes the same steps back up the parent chain where
a local redraw starts part-way down the tree. Neither chain has a coordinate
frame in it, so neither can break at a region node, and warm and cold reach
every length by the same expression.

Three things follow. `Holds::through` is the exact preimage of
`px + floor(rel * box)` -- two divisions, no allowance, the whole of a box
mapping back to itself. A local redraw asks in the box its parent gave it and
only where that box is as long as the offer, which retires `redraw`'s third
ask and the region-node exception beside it; `draw_inner` places the answer
inside that box itself. And symbolic regions are left to the GPU, hit testing
and remaps, where `Moves::resolve` is the only walk: `wide.rs`,
`Moves::compose`, `Moves::size_of`, `px_of`, `px_region`, `offered_region`
and `slot_wide` are gone, 252 lines of `core/` net.

`px` is deliberately not stored beside those lengths. A resize every widget's
`Holds` admits redraws nothing, so a stored pixel length would be stale on
every widget in the tree with nothing on it to say so, and refreshing it costs
a walk down every reused subtree on the resize path.

Instructions:u, medians of 21 runs, seed 1 at depth 8:

| phase | before | after | |
| --- | ---: | ---: | ---: |
| `cold`, 200 frames | 313.1M | 312.9M | -0.04% |
| `resize` | 408.1M | 405.6M | -0.61% |
| `many` | 1,924M | 1,756M | -8.75% |
| `scroll` | 357.3M | 323.4M | -9.49% |
| `repaint` | 363.3M | 315.4M | -13.18% |

`cold` and `resize` have all twenty-five work counters identical, so those
two rows say the draw path costs the same threaded as composed. The other
three do less work: `repaint` goes from 23 draw requests and 13 widget draws
a frame to 1 and 1, `scroll` from 20 and 11 to 8 and 2, `many` from 273 and
186 to 207 and 157. Primitive writes are unmoved in every phase.

Verified: `view`, `minimal`, `random`, `tabs` and `text` render
byte-identical at 1920x1200 against `5b78002`, as does the `tabs` touch
replay before and after the gesture, and a live resize of `random` to
1280x800 is identical both to the old head's and to a cold render at that
size. The oracle passes 100 seeds in release and 120 in debug -- the debug
run is the one that exercises the `Holds` assertion -- and the fifteen
shrinker cases pass at 400 seeds of depth 5 and 1000 of depth 6. Seed 220 is
`unsettled::a_widget_under_a_region_node_is_asked_in_the_box_that_node_was_offered`,
which needs both halves of this to fail: the old chain with the old allowance
passes it, and the old chain with the exact preimage does not.

`AGREE_STEPS` stays 2. One step passes the 100-seed oracle and fails the
400-seed shrinker on `resize-size` by 0.002 px, so what is left there is the
resize path re-expressing a part as a fraction of a box that changed length,
not a length reached two ways.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/cases/unsettled.rs

@@ -3,13 +3,15 @@
 //! frame that had not settled: a wrapping text shaped at a width it was
 //! measured in rather than the one it was given. The rest are a widget
 //! measured again in a box its own answer had decided, where the old answer
-//! is a fixed point whatever the content now says. The last two are neither:
-//! one box length, composed two ways, landing either side of the boundary
-//! that decided whether a child was drawn at all, and one box as long as the
-//! box a widget was offered but somewhere else.
+//! is a fixed point whatever the content now says. The last three are
+//! neither: one box length, composed two ways, landing either side of the
+//! boundary that decided whether a child was drawn at all, and two boxes
+//! reached through a region node's own entry rather than through the offer
+//! that node was given.
 
 use iris::harness::Harness;
 use iris::prelude::*;
+use iris::random::Branch;
 
 /// Six widgets, shrunk from a 402-widget tree the fuzzer found. Nothing about
 /// the tree changes -- every widget is marked for redraw and the frame is
@@ -428,12 +430,12 @@ fn plant_nested_scrolls(h: &mut Harness) -> Vec<WidgetId> {
     vec![text.id(), inner.id(), filler.id(), span.id(), root.id()]
 }
 
-/// A local redraw asks a dirty widget in the box its parent asked it in, and
-/// then again in the box its parent chose from that answer. Skipping the
-/// second ask because the two boxes are the same *length* left this inner
-/// scroll, which owns a region node, drawn at its offer. The offer is the
-/// outer scroll's whole viewport and the final box is 24px above it -- the
-/// height of the sized child the outer scroll snaps to the end of -- so the
+/// A local redraw asks a dirty widget in the box its parent gave it, and only
+/// where that box is as long as the one it was offered; anything else is a
+/// question its parent has to ask. This inner scroll's offer is the outer
+/// scroll's whole viewport and the box it was given is 24px shorter -- the
+/// height of the sized child the outer scroll snaps to the end of -- so what
+/// it must not do is settle itself. It was drawn at its offer once, and the
 /// inner scroll and its text stayed 24px too low.
 #[test]
 fn redrawing_one_widget_does_not_move_what_scrolls_around_it() {
@@ -454,3 +456,101 @@ fn redrawing_one_widget_does_not_move_what_scrolls_around_it() {
     }
     assert!(wrong.is_empty(), "{}", wrong.join("\n"));
 }
+
+/// Ten widgets, of the shape `tests/shrink.rs` reduces the oracle's seed 220
+/// to. The pad owns a movable region and is the scroll's content, so the box
+/// the scroll places it in is as long as that content while the box it was
+/// offered is the viewport -- and with no padding to tell those two apart,
+/// the span inside it looked like it was still at its offer. So everything
+/// under the pad was asked again in the *placed* box, the offer resolving
+/// against the node's own entry, which holds that box: the texts kept the
+/// widths they had, the content stayed the length those widths make, and the
+/// old answer confirmed itself. What the branch adds is a tree that differs
+/// rather than a box that moved, since a probe measured at the wrong width
+/// takes the other side.
+fn plant_under_a_node(h: &mut Harness, swapped: bool) -> (Vec<WidgetId>, [WeakWidget<Span>; 2]) {
+    let probe = rect(Color::RED).add(&mut h.rsc);
+    let wide = rect(Color::GREEN).add(&mut h.rsc);
+    let narrow = rect(Color::BLUE).add(&mut h.rsc);
+    let branch = Branch {
+        probe: probe.add_strong(&mut h.rsc),
+        wide: wide.add_strong(&mut h.rsc),
+        narrow: narrow.add_strong(&mut h.rsc),
+        threshold: 213.0,
+    }
+    .add(&mut h.rsc);
+    let wrapped = wtext(
+        "Wrapping shapes one source into as many lines as the box \
+         leaves room for, so a paragraph's height is an answer and not a setting.",
+    )
+    .size(16)
+    .wrap(true)
+    .add(&mut h.rsc);
+    let plain = wtext("one line, overflowing whatever it is given")
+        .size(16)
+        .wrap(false)
+        .add(&mut h.rsc);
+    let row = |h: &mut Harness, mut children: Vec<StrongWidget>| {
+        if swapped {
+            children.rotate_left(1);
+        }
+        Span {
+            children,
+            dir: Dir::RIGHT,
+            gap: Px::ZERO,
+        }
+        .add(&mut h.rsc)
+    };
+    let texts: Vec<StrongWidget> =
+        vec![wrapped.add_strong(&mut h.rsc), plain.add_strong(&mut h.rsc)];
+    let inner = row(h, texts);
+    let pair: Vec<StrongWidget> = vec![branch.add_strong(&mut h.rsc), inner.add_strong(&mut h.rsc)];
+    let outer = row(h, pair);
+    let pad = Pad {
+        padding: Padding::ZERO,
+        inner: outer.add_strong(&mut h.rsc),
+    }
+    .add(&mut h.rsc);
+    h.rsc.widgets_mut().set_region_node(pad.id(), true);
+    let root = Scroll::new(pad.add_strong(&mut h.rsc), Axis::X).add(&mut h.rsc);
+    h.set_root(root);
+    (
+        vec![
+            probe.id(),
+            wide.id(),
+            narrow.id(),
+            branch.id(),
+            wrapped.id(),
+            plain.id(),
+            inner.id(),
+            outer.id(),
+            pad.id(),
+            root.id(),
+        ],
+        [outer, inner],
+    )
+}
+
+#[test]
+fn a_widget_under_a_region_node_is_asked_in_the_box_that_node_was_offered() {
+    let mut warm = Harness::new((900, 1200));
+    let (ids, spans) = plant_under_a_node(&mut warm, false);
+    warm.frame();
+    for span in spans {
+        warm.rsc[span].children.rotate_left(1);
+    }
+    warm.frame();
+
+    let mut cold = Harness::new((900, 1200));
+    let (cold_ids, _) = plant_under_a_node(&mut cold, true);
+    cold.frame();
+
+    let mut wrong = Vec::new();
+    for (i, (&w, &c)) in ids.iter().zip(&cold_ids).enumerate() {
+        let (got, want) = (warm.region(&w), cold.region(&c));
+        if got != want {
+            wrong.push(format!("widget {i}: warm {got:?} cold {want:?}"));
+        }
+    }
+    assert!(wrong.is_empty(), "{}", wrong.join("\n"));
+}

tests/scenario/mod.rs

@@ -42,11 +42,19 @@ const OUTER: (f32, f32) = (1920.0, 1200.0);
 const INNER: (f32, f32) = (640.0, 900.0);
 const STILL: (f32, f32) = (900.0, 1200.0);
 
-/// The same box, to a step of the grid per level of nesting between the two
-/// ways of reaching it. A move, a repaint and a row of shares land on the
-/// same number; what is left is a box centred in a fraction of its parent
-/// against the same box centred in its own pixels. A step is a thousandth of
-/// a pixel, where this was a twentieth of one before any of it was on a grid.
+/// The same box, to two steps of the grid between the two ways of reaching
+/// it. A move, a repaint, a row of shares and every length in pixels land on
+/// the same number. What needs the slack is a position: a box centred in a
+/// fraction of its parent against the same box centred in its own pixels,
+/// and a box re-expressed as a fraction of a parent that changed length.
+/// A step is a thousandth of a pixel, where this was a twentieth of one
+/// before any of it was on a grid.
+///
+/// **One step is not enough**, tried 2026-09-17 once a length in pixels
+/// stopped being composed: it passes the 100-seed oracle and fails the
+/// 400-seed shrinker on `resize-size`, seeds 384 and 162, by 0.002 px. So
+/// what is left here is the resize path's own rounding rather than a length
+/// reached two ways.
 const AGREE_STEPS: i32 = 2;
 
 /// A way of changing what a span holds. Each is a shape worth its own case: