ai-app/docs/HANDOFF.md

Handoff

Where the work in flight stands for a session picking it up cold. Keep current invariants, measurements, and failed hypotheses here; this is not a decisions log.

Where things stand

Canonical Iris main is ca2b4b2 (#17, the headless rig). #18 split/18-position-chain is open in /home/bob/repos/iris-pr18; its local head is d21a215, eighty commits, pushed. Built-in alignment is complete there; see "Built-in alignment" below for the retained-layout details. No PR reviews were present when checked on 2026-09-15.

The current head completes LAYOUT.md §2's position chain and the requested leftover behavior. A child whose length is only leftover is not drawn when nothing is left. A child that also asks for pixels or a relative fraction keeps that part and overflows as before.

29c7881 replaces the parallel resize rules with one retained-layout contract, Holds: the interval of box lengths for which a widget's drawing and reported size stay valid. Reading Painter::px_len or px_size narrows the interval to the length read; Painter::holds lets the widget widen it. Parent validity is the intersection of the ranges its children induce. This contract is trusted. A widget that declares an incorrect range is a defective widget; Iris does not add defensive work to recover optimizations from a false declaration.

9d8415d deletes OrthoSize, which f437495 had restored as an explicit Span::ortho(OrthoSize::{Children, Full}). It was the rule beside a widget written a second time: a span across itself is as long as its longest child unless a rule already says how long it is, and Painter::ruled(axis) is how it asks which case it is in. Under a rule it does not read its children across that axis at all -- the answer is not wanted, and reading one is what would make its size depend on theirs. Not under one, the largest fixed orthogonal length is reported and any relative or leftover child makes it report leftover, which is the old conservative behavior. Either way the children's Holds ranges propagate through final-box drawing, so a resize repositions them without redrawing when their own contracts permit it.

a8898aa gives a length with no share in it its own type again, which is the separation canonical main had and this branch had collapsed: Len is pixels plus a fraction of a box -- a position being the length from the box's start, which is why a span is two of them -- and LayoutLen is a Len plus a claim only a container dividing its room can answer. What was UiScalar is Len, what was Len is LayoutLen, and From<Len> for LayoutLen is the one-way step. 490918b renames Painter::ruled to has_exact_size, over SizeRule::exact: "is there a rule here" and "is my report moot" are the same question only while Exact is the only rule, and Min/Max/Clamp are queued.

71c9c39 replaces the public Painter::place distinction with an opt-in widget property. .region_node() gives a widget one independently movable retained region; Widgets::set_region_node can change that choice at runtime and causes one structural redraw. Widgets without the property remain at the default shallow chain depth: Iris recursively remaps their retained primitive and mask regions when they move. .scrollable() enables a region node on its content once as its convenient default; raw Scroll::new respects the caller's choice, and the property can be disabled later without breaking scrolling. Span and Align do not add nodes to their children.

Do not make a span choose its orthogonal size by comparing children in pixels at its current width. A fixed child and a relative child can create multiple self-sizing fixed points; generated seed 13 settled differently warm and cold under that attempted implementation. A Holds interval says where an already chosen answer stays valid, but cannot make that circular choice unique. The same circularity is what a leftover cap would put into SizeRule::Max -- see the clamp item under "Next".

The implementation also fixes three counterexamples found while finishing the rewrite:

• An asked-but-undrawn size dependency must name the widget that asked as its parent. Using the asker's parent skipped a reader and made generated seed 10 settle differently warm and cold.
• Scroll must return the answer from the first box it asked about, whether that answer came from a retained length or a fresh measurement. Returning the final placed answer only on the retained path advanced one fixed-point iteration and broke seed 86.
• A widget retains the layer it was entered on, not the last child layer its painter visited. The old value drifted on local redraw and put a redrawn tab background above its retained text.
• Span's leftover/no-leftover split is a strict layout decision, not a rounding tolerance. Its Holds range must use the same exact divided boundary as drawing; a tolerant endpoint retained zero-height children at the boundary in generated seed 16. 5ed9e87 moved that boundary rather than softening it, and 39e4ca2 removed the move: on the grid the box and the sum are the same count. See "Fixed point" below.

core/src/ui/holds.rs, the retained tests, and the generated cold-layout oracle pin those rules. Seeds 10 and 86 are now in the ordinary generated set.

Warm layout agreed with cold, and the rigs that got it there

Fixed in aea878d, and the bisect was a red herring. git bisect named 95fb4f9, the commit that made Masked report Size::LEFTOVER; the failing tree has no Masked in it and that commit changes nothing else but a debug_assert, so the attribution was wrong and the write-up built on it ("Scroll clips through a Masked") was too. Read the tree rather than the bisect next time a commit that cannot be the cause is named.

The tree, five widgets, now running in tests/cases/unsettled.rs as redrawing_one_widget_does_not_move_what_scrolls_around_it:

Scroll(Y) { Span(DOWN) { Scroll(X){ text } as a region node, rect 87x24 } }

UiRenderState::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, and it skipped that second ask whenever the two were the same length. They are not the same box: the outer scroll offers its whole viewport and places the span 24px above it, snapping to the end of a content 24px longer than its box, so the offer was as long as the final box and 24px below it. A region node writes the box it drew in into its own move entry, so the inner scroll stayed where the offer put it. d3b0ebf had already compared whole boxes for parent_must_place and left this one a length comparison -- one decision with its two halves disagreeing.

The rigs. 98d4e98 splits iris::random::grow into plan(seed, depth, &edits) and build(rsc, &plan). A seed cannot be made smaller, which is why a failure the oracle found could never be handed to the shrinker; a Plan can, and tests/scenario/ holds the fifteen cases both rigs now run over the same trees. So any failing seed reduces directly:

SHRINK_SEED=18 SHRINK_DEPTH=6 SHRINK_CASE=repaint-some \
  cargo test --release --test shrink -- --ignored --nocapture

That took seed 18's 277 widgets to 5. The oracle prints the command to reduce whatever it failed on. Seeds still mean the trees they meant: the 1000-seed depth-6 run gives the same three failures with the same boxes before and after the split, which is the check to repeat if the generator is touched again.

Nothing is left, as of d8ae9c3. All fifteen shrinker cases pass at 400 seeds of depth 5 and the oracle passes 1000 seeds of depth 6, both for the first time. What closed the last two was one line: redraw's second ask, in the box its parent chose from the widget's answer, handed that box over as if it were an offer, so draw_inner ran placed_box on an already-placed box and applied the widget's own alignment to it twice. That only showed where the alignment was the widget's own to apply, which is why it stayed hidden while Stack, Pad and Scroll overrode every child.

Closed by it: shrinker seed 288 on region-node (11 widgets from 42, a Text inset 8.8px at each end) and oracle seed 326 at depth 6 (43 from 205, 88px out, around two Branches). Neither reduced small enough to hand-write, so generated seed 20 at depth 4 is the record -- it is in SEEDS, so cargo test fails without the fix rather than only the ignored long run. Closed earlier by aea878d: shrinker seeds 174, 175 and 2, oracle seeds 18 and 190.

Depth is what found these. Nothing failed at 100 seeds of depth 4, which is all the oracle had ever routinely run. Before fixed point, 5ed9e87 fails the 1000-seed depth-6 run on seed 40 after a resize; that one is older than all of the rest and is also gone.

Run the long two before believing a rounding change. They are 14 s, 57 s and 145 s, which is nothing against how long the two above took to find:

cargo test --release --test generated -- --ignored a_long_run_of_seeds_agrees
SHRINK_CASE=all SHRINK_SEEDS=400 SHRINK_DEPTH=5 \
  cargo test --release --test shrink -- --ignored --nocapture
IRIS_GENERATED_SEEDS=1000 IRIS_GENERATED_DEPTH=6 \
  cargo test --release --test generated -- --ignored a_long_run_of_seeds_agrees

And run the ordinary suite in debug: the Holds assertion in draw_at is what says a validity range stopped containing the box a drawing was made in, and it is compiled out of the release runs above.

Verification

At e166e00, the current head:

• cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo test --workspace: sixteen targets green, 83 suite tests, 17 core unit tests, 11 generated cases. What stays ignored is the long runs and the profiling rigs; no known defect is ignored any more.
• The release oracle at 100 seeds in 14.3 s, all fifteen shrinker cases at 400 seeds of depth 5 in 57 s, and 1000 seeds of depth 6 in 145 s. The last two pass for the first time.
• view, minimal and random render byte-identical at 1920x1200 against 98d4e98. tabs and text changed twice on purpose: 08c9d5a moved an antialiased edge by less than a pixel (see "Performance") and 2bc6bdf gave stack and pad children their own alignment back, which is what tabs and text were asking for and what canonical main draws. Take the oracle rather than a render as the reference from here; view, minimal and random are the three still worth comparing.
• Twenty-five rig work counters identical between 60367d8 and 08c9d5a on the many fixture, which is what makes their timings comparable, and unmoved again by 38eba54.

At 394d514, kept because several of these have not been re-run since:

4febabf on top of it -- Fixed wrapping rather than saturating -- passed the workspace tests with all features (105), the release 100-seed oracle in 9.5 s, and rendered random, tabs and text byte-identical at 1920x1200 against 394d514. The measurements are under "Performance".

• cargo fmt --all --check
• cargo build --workspace --all-features
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-features: 105 passed, 10 ignored
• The release generated cold-layout oracle passed 100 seeds in 9.4 s, and a shrinker case at 300 seeds in 3.5 s: both run a thread per core but one.
• The release shrinker passed all five cases it had then -- resize, repaint, resize-repaint, reorder, size-change -- at 300 seeds of depth 5. That claim does not carry to 98d4e98, which runs fifteen cases over the oracle's trees rather than five over its own, and fails at 400 seeds of depth 5. Neither does the 1000-seed depth-6 line that used to sit here: re-run, 5ed9e87 fails it too, on seed 40 after a resize.
• tabs, view, minimal, text and random render byte-identical at 1920x1200 across the whole fixed-point sequence, and across this session's renames and arithmetic changes on top of it.
• tests/cases/drift.rs passed its 20,000-move exactness check in release mode.
• The seeded random example, live-resized from 1920x1200 to 1280x800, is byte-identical to a cold 1280x800 render. Both PNGs hash to 1d397c57b9914a2e596fa907029bc6aab4629e4d74deb0715e81325c703bcdb3. The run used the Venus adapter backed by the host RX 7900 XT.
• minimal, text and view render byte-identical at 1920x1200 across 8220a78. tabs differed only in the widget count it printed about itself, which is two wrapper types smaller -- so it is no longer a byte-identical reference and the generated oracle is the check that matters. That last claim was then left to cover d3b0ebf as well, and it does not: tabs changed twice more there, and nobody looked (2026-09-16). See "Alignment as a property changed two examples" below. A claim about a render holds for the commit it was checked at and no further.
• At preceding head 29c7881, reference renders against /home/bob/repos/iris-main-cmp at ca2b4b2 covered: tabs, view, minimal, and text at 1920x1200; tabs cold at 900x1200; live resize from 1920x1200 to 900x1200; and the tab interaction before and after replay. Every comparison had zero differing pixels. The live-resize image is also byte-identical to the cold 900x1200 image.

The built-in-alignment pre-submit review was run in four passes. It caught three distinctions the smaller tests had missed: an answer's validity must include the drawing made in its final placed box; a region-node box is resolved through its parent move rather than through its own move twice; and boxes with equal dimensions but different positions still need the parent that placed them. The 100-seed oracle and both 300-seed shrink cases pass after those fixes. The earlier region-node review caught an index-reuse hazard when removing a node; its move entry now remains alive until every descendant has migrated. The generated oracle then exposed the exact Span threshold described above. The earlier retained-layout review caught the layer defect above, corrected validity-range inversion for negative relative extents, and removed an impossible-state unwrap from Scroll. The orthogonal-sizing review caught both the circular longest-child choice and the incompatible visual effect of making Full the default.

Performance

The 3x this section used to report was a fixture artifact, withdrawn 2026-09-16. random.rs's Branch picks which of two subtrees to draw by comparing a measured pixel length with a threshold, so the fixture's shape moves with the thing being measured. Seed 1 at depth 8 draws 88 widgets and writes 2,298 primitives a frame at 5ed9e87, and 115 and 8,209 at bd6de71 -- three and a half times the work behind a number read as three and a half times the cost. 394d514 gives the rig Edits::fixed_branches; the oracle keeps measured branches, which is the whole point of them.

Check the work counters before comparing two commits' times. The rig prints drawn widgets, widget draws and primitive writes. An undrawn leftover child still moves them and no flag removes that, so the trees are near rather than equal -- and 9d8415d changed the generator itself, since an X span's OrthoSize::Full became a size rule.

Measured on the fixed-shape fixture, seed 1, depth 8, 500 frames of many:

	drawn widgets	primitive writes	instructions	cycles	IPC
5ed9e87, before fixed point	100	4,272	1,761M	688M	2.561
4cbb242	97	3,951	2,093M	~819M
4febabf, wrapping	97	3,951	1,912M	771M	2.48
60367d8, rounding	97	3,951	1,915M	777M	2.465
08c9d5a, truncating	97	3,951	1,800M	715M	2.516

The three ends were measured as medians of 25 runs of binaries built for the reading; 4cbb242's row is still the single reading it was taken from, and 4febabf's is medians of nine from 2026-09-16. 5ed9e87's numbers reproduced to within 0.1% and 0.7% across those two sessions, which is what says the rig is sound.

60367d8 and 08c9d5a are the one pair here that can be compared directly: all twenty-five of the rig's work counters are identical between them, so -6.0% instructions and -8.0% cycles is the same work at a different speed. Against the float head the head is +2.2% and +3.9%, down from +8.7% and +12.9% -- but take neither for the grid's cost. Those two do different work: the float head draws 100 widgets to the head's 97, writes 4,272 primitives to 3,951, remaps 38 subtrees a frame and redraws 43 where the head remaps 53 and redraws 27, and renders 49 texts to 42. Dividing by primitive writes does not correct for a different mix of remapping and redrawing.

The fused multiply-add this section used to name does not exist, withdrawn 2026-09-16. 5ed9e87's binary holds no FMA instruction at all, and none appears when it is rebuilt with -C target-feature=+fma,+avx2 either. Rust does not contract a + b * c, so LerpUtil::lerp's from + (to - from) * self is subss, mulss, addss on every target there is. The float side never had the single instruction the grid's multiply was being compared against, here or anywhere.

Nor is the multiply the cost. Each row below is 4febabf with one thing taken out, in one worktree so the rows compare to each other, 500 frames of many at seed 1 depth 8, medians of nine to eleven runs. Every row leaves the work counters where they were, so unlike the cross-commit numbers above these compare one tree against itself: all twenty-seven counters diffed identical for the all-three row that the conclusion rests on, and six of them checked for the others.

	instructions	cycles
4febabf	1,921M	772M
shift_round out of Fixed::mul, truncating instead	1,878M	774M
is_full short-circuits out of UiSpan::within	1,910M	765M
zero short-circuit out of Fixed::scaled	1,914M	768M
all three	1,803M	715M
branchless shift_round, both short-circuits out	1,964M	773M

The "all three" row is the head as of 08c9d5a; it re-measures there at 1,800M and 715M, which is the 0.2% a rebuild moves.

Taking the rounding out alone removes 43M instructions and moves cycles by +2M, which is inside the noise: the grid's arithmetic is instructions the machine has spare issue width for, not time.

IPC has reversed since this section recorded it, and that is the same finding from the other side. It was 2.58 before fixed point and 2.59 after, which is what "the cost is instruction count at unchanged throughput" was read from. On the re-measured medians it is 2.54 before and 2.48 after: the head both executes more instructions and retires them more slowly, so there is something to find beyond counting them. The all-three build recovers half of it against its own worktree's baseline, 2.49 to 2.52.

What moves cycles is whether UiSpan::within inlines, and the three cuts are superadditive because of it. Separately they are worth 61M instructions and 8M cycles between them; together, 118M and 57M -- four fifths of the instruction gap and two thirds to three quarters of the cycle gap, which is the looser number. The mechanism is visible in nm: <UiSpan>::within is a symbol in 4febabf and in neither 5ed9e87 nor the all-three build. Shortening its body past the inliner's threshold is what pays, and only the whole set of cuts reaches it. That is where to push: a cheaper Holds or a smaller Len::within is worth trying for this reason rather than for its own instruction count.

Cycle noise is larger than this section used to claim. Across ten sets of nine to eleven runs, each of one unchanged binary, cycles spread 1-3% usually and 6.7% in the worst set -- not the 0.23% recorded from three runs, which was a lucky set read as the machine's precision. Worse, instructions:u in this VM intermittently returns a garbage value -- 11.0e9, 25.8e9 and 28.4e9 against a true 1.9e9, in roughly a quarter of readings. cycles:u does it too, seen 2026-09-16: 7.1e9 and 11.3e9 against a true 715M, and in one set of three runs all three cycle readings were garbage while all three instruction readings were good -- so a filter that drops anything off by a factor has to report how many readings it kept, or an empty sample reads as a zero. Take medians of nine or more and do not trust a single perf stat. ex_div_busy was the one counter that held to 0.1% across runs.

Instruction counts hold to 0.02% within a binary and move 0.5% across a rebuild of the same source, which is why the two tables above disagree about 4febabf by 9M: their baselines were built in different worktrees. The difference against 5ed9e87 was 149M in both builds. So compare against a baseline built beside the thing being measured, quote a delta rather than an absolute, and re-measure the baseline whenever anything is rebuilt.

"The division is 0.00% of cycles" was read off the wrong function. It is true of AxisRemap::apply_scalar, where most moves are translations and the multiply is 1.5%. It is not true of the run: Holds::through divides twice on every call and accounts for essentially all of the 2.23M i64 divisions a 500-frame many executes, which hold the integer divider busy 21.30M cycles -- the steadiest counter on this machine, 0.1% across runs, and 2.8% of the run. A controlled extra div_toward in through, against a de-const control so the inlining does not move, costs 6.3 cycles each, so the two real ones are about 14M. 5ed9e87 divides twice there too, in floats, so this is a real cost but a smaller difference than the source-line profile makes it look, where div_toward's num / den is the largest single grid line in the head at 106M of 5,361M over 4,000 frames, against shift_round's 71M.

Recovered this session, each measured before and after on one commit with byte-identical renders proving the layout unchanged -- on the old fixture, 689.6M cycles and 1,886M instructions down to 638.9M and 1,657M:

• cb1bba4 works a move out once per subtree rather than per scalar: -3.8% instructions, -0.8% cycles. The gap between those two is the finding.
• 1940e85 translates a whole region at once, since that is what a translation is: -4% instructions, -3.4% cycles.
• d75a1e2 and 4cbb242 skip multiplies that cannot change anything -- composing through a full box, and lerp where the ends are equal: together -2.7% instructions, -1.8% cycles when they landed. That no longer reproduces at 4febabf: removing either one now saves a little, and removing both is part of the 118M/57M above. Wrapping changed what a multiply costs, and a guard priced against a saturating multiply is priced against something that is gone. Re-price a short-circuit before keeping it.

Tried and rejected, with numbers, so they are not tried again:

• A float reciprocal for the division: +6% cycles. The conversions cost more than the idiv did there. That was apply_scalar's division; the one in Holds::through is a different question and has not been tried.

• Branchless shift_round: +6.7% cycles when tried alone, and re-tested 2026-09-16 with both short-circuits also removed, where the branchy form reaches 1,803M/715M: 1,964M instructions and 773M cycles, worse than the head on both. It is the size of the rounding that keeps within out of line, not its branch, so making it branchless makes it bigger and loses twice.

• Removing the per-child hash lookup in remap_subtree with mem::take: 0.0%. FxHashMap on a hot line is not the cost.

• Short-circuiting apply_scalar where the fraction is nought or one: +17%. (From the previous session.)

• #[inline] on UiSpan::within, whose prologue and epilogue were a sixth of its own cycles: -0.2% instructions, +1.5% cycles. That is not in tension with the inlining finding above: forcing the body it has now into every caller loses, and shrinking the body until the inliner takes it of its own accord wins. Shrink it; do not annotate it.

4febabf makes Fixed wrap rather than saturate (Bryan, 2026-09-16: a coordinate past the range is not going to draw reasonably anyway, so wrap and break clearly): -8.6% instructions, -6.6% cycles, more on cycles than the 2.7% the instruction share predicted. from_f32 still clamps and Holds keeps its saturating narrow, since a range of box lengths past i32 really is unbounded.

08c9d5a makes Fixed::mul drop to the step below rather than round (Bryan, 2026-09-16: truncation is preferable at this point). The three cuts this section priced together landed as one commit, since separately they are worth 61M instructions and 8M cycles and together 115M and 62M -- the whole of the superadditivity above. Fixed::scaled was its own zero short-circuit and is gone; UiSpan::within lost both is_full tests and is inlined again, which nm confirms.

What it costs, all of it measured rather than asserted:

• A share lands a thousandth of a pixel short of its row rather than on it. an_uneven_nesting_still_gives_every_share_the_same_length now says what is still exact -- each share starts where the last ended, the row ends at its own edge, each edge is on the even division or one step below.
• A value and its negation are no longer the same distance from where they came, since the drop is toward negative infinity on both sides of zero, so a flipped span can sit a step from its mirror image. The alternative that keeps that symmetry is the sign branch, which is part of what was bought.
• tabs is no longer a byte-identical render: 4,664 of 2,304,000 pixels differ, in single-pixel-wide runs along 80 columns of one band of rounded rects, which is an antialiased edge moved less than a pixel. view, minimal, text, random and the tab replay are unchanged.
• The warm-against-cold oracle is not worse: 100 seeds pass, the fifteen shrinker cases at 400 seeds of depth 5 fail only on seed 288 as before, and depth-6 seeds 18 and 190 pass with 326 failing as before.
• Holds::through had to be re-derived, and the derivation for one truncation either side is measurably too narrow. See the item under "Retained-layout invariants".

Earlier, against #18's own history: the retained rewrite took many from 25.17M instructions a frame at 691e3eb to 6.14M at 29c7881, and resize from 16.08M to 8.39M. That fixture has since changed twice; do not compare across it.

Retained-layout invariants

• A retained drawing belongs to the layer it was made on. Asked for again on another layer it is redrawn, since nothing about its geometry says it is in a list that paints at a different moment. A container that measures a child by drawing it therefore measures on the layer that child will draw on -- Painter::child_layer_at addresses one -- or it pays two draws a frame forever and keeps whichever the second ask left.

• A widget that clips its contents to its box reports its box: Scroll and Masked both report LEFTOVER, and a debug_assert holds any widget that set a mask this draw to it. Overflowing is otherwise ordinary and a text too tall for its box says so -- which is why the assertion is narrowed to mask-setters rather than to every widget. Masked is the only one that sets a mask; Scroll reports its box because a drawing placed in a longer box is scaled by recomposition, not because it masks. So a LazySpan is unaffected: it clips by not drawing, and may report its whole content length, which is what the Scroll above it measures. One that set its own mask would have to report its box and hand the content length up another way.

• A span is as long across itself as its longest child, unless a rule beside it says how long it is -- and then it does not read its children there at all, since the answer is not wanted and reading one is what makes its size depend on it. OrthoSize was that second case written twice and is gone (9d8415d); Painter::ruled is how a container asks which it is in, and the only thing a widget may learn about a rule over it.

• A region node holds a whole UiRegion in its parent node's coordinates. UiRegion::FULL is the identity. Widgets opt in with .region_node() or Widgets::set_region_node; ordinary widgets share the nearest ancestor node. Region nodes therefore add chain depth only where moving a whole subtree through one entry is useful.

• A widget's ActiveData::region is its box in its parent node. A region-node widget draws in FULL; its box lives in its node. Moving an ordinary retained subtree instead remaps its primitive, mask, and active regions. Remapping stops at a descendant region node after rewriting that one entry.

• Changing region_node redraws the subtree once to rebuild the coordinate boundary. The property belongs to widget identity, which is safe because a widget has one parent. .scrollable() sets it once; raw Scroll::new does not, and Scroll never reasserts it while drawing.

• The first box a parent asks about is the offer. A later box chosen from the child's answer is the final box, not another independent answer. Dirty widgets are re-asked at the offer and only then drawn in the final box. An offer composes through its ancestors' offers, not through their current placed boxes.

• An answer is reusable only where both its measurement and the drawing made in its final placed box remain valid. The final drawing's Holds interval is translated back into lengths of the offered box and intersected with the answer's interval.

• A container's near-edge override is owed only where the box really is the child's own answer. Stack gives every child the box box_of derives from its sizing child, so that one child has no room in it and would be placed twice; every other child is handed a box that owes nothing to it and keeps its own alignment. Pad reports its inner plus the padding, so where its box is that answer the inset box is exactly the inner, and where the box is bigger the slack is the inner's to sit in -- it overrides nothing (2bc6bdf). Scroll still overrides, on both axes and deliberately: the box it hands its content along the scrolling axis is the content's own length, and where shorter content sits across the viewport is Scroll's own alignment by design.

• A box a parent has already placed is asked about at the near edge. The second ask in redraw is in the final box, not at an offer, so applying the widget's alignment to it again places its content twice (d8ae9c3).

• Equal box lengths do not imply equal placement. An ordinary widget whose offered and current boxes differ in position must involve its parent again; a region node can settle itself only when its own alignment, rather than a container override, determines the final box. Both halves of that decision compare whole boxes -- whether the parent must place it, and whether the offer already was the final box. Comparing lengths for the second left a region node drawn at its offer (aea878d).

• A retained drawing can be reused only when its Holds interval contains the new pixel box on both axes, its parent node is unchanged, its region-node choice matches the retained structure, and the widget is clean. A valid ordinary subtree may move without redrawing because its regions are recursively remapped.

• Painter records size-dependency edges only when a parent reads a child's size or hint. An undrawn measured child remains recorded so a later change reaches the parent that decided whether to draw it.

• Dirty widgets settle deepest-first. dirty_size_under prevents a reader from taking a retained answer while something below that answer is still dirty; the walk is an optimization against laying out twice, not a second validity mechanism.

• Declared non-leftover lengths are resolved by the widget's parent where the widget is drawn. A declared-length change therefore redraws the parent.

• A pixel comparison is equality: lengths are whole counts of 1/1024 px, so a change too small to reach the next step is not a change and one that reaches it is, however little of a pixel it is worth.

• A length given in pixels is that many pixels wherever it ends up, and structurally rather than by luck: Len::within adds a part's own pixels rather than scaling them, and both ends of a gap carry the same fraction, so the multiply that rounds is the same on each and cancels. Pinned by a_length_in_pixels_is_that_many_pixels_however_it_is_nested, over a row buried under three containers that are each a fraction of their parent; swept over 2,100 box widths and exact at every one.

• A length given as a share is not, and cannot be: children asking for the same fraction come out one or two steps apart, 0.001 to 0.002 px. A position is the quantity that gets rounded, so the row fills exactly and no two children leave a seam, and the spread between lengths is what that costs. Exact composition would shrink it, not remove it -- five equal lengths cannot fill a row whose step count is not a multiple of five. Both halves of this are in equal_shares_differ_by_at_most_two_steps_and_fill_the_row.

• Text shaping is retained separately from line breaking. A greedy line break remains valid from its longest produced line through the width at which it was made, and TextView reports that interval through Painter::holds.

• Span's decision to distribute leftover is a pixel question. The room to divide is len * fixed - total.px; pure leftover children are undrawn where there is none. The box a parent hands back and the sum of what the children asked for are counts of the same step, so the boundary needs no margin and the validity interval is split exactly at it.

• Scroll reports Size::LEFTOVER; what it takes from the first box it asked about is its content's length, which it clamps and holds by rather than reports. Its drawing can survive container-length changes only over the interval in which clamping and its current offset do not change.

• A move that keeps a box's length is a translation, and an offset is exact on the grid. A box that also changed length has to re-express each part as a fraction of the new one, and that division and multiplication round: 39e4ca2 translates where from.len() == to.len() and scales only where it must, which is what made the shrinker's resize case agree exactly. This inverts the float-era rule, and the measurements behind that rule are why tests/cases/drift.rs exists: in floats, offsetting both ends of a span shortened that fixture's row by 0.071 px over 20,000 moves and 0.712 over 200,000, while re-expressing fractions stayed exact. 20,000 moves is five minutes of scrolling at 60Hz. On the grid the drift is gone either way, and tests/cases/drift.rs pins that it stays gone.

Built-in alignment

Committed as d3b0ebf in /home/bob/repos/iris-pr18.

What is in it: align is a widget property beside region_node and the size rule, Aligned is deleted, .align()/.center() set the property, and the fuzzer covers size rules, alignment and region nodes and changes all three at runtime. Region nodes had no generated coverage at all before that.

Alignment is two fractions, not four directions

Decided 2026-09-15 (Bryan). A widget's alignment is one f32 per axis, so a quarter of the way along an axis is expressible. AxisAlign's three familiar positions are named constants over that number, which is what every expression already uses: the layout math only ever reads AxisAlign::rel(), so nothing downstream changes shape.

The default is the middle on both axes, because the two edges are the ones that assume a direction -- which edge is the near one depends on the writing system and on which way a container runs. "Near edge" throughout this document means the start of the box in the box's own orientation, which for a reversed span is its visually far end, not "top left".

Placement cannot be applied after the fact

Do not re-attempt "draw the widget, then move its drawing to where its alignment says". Three attempts failed, and the reason is structural: the move is a change of coordinate frame, and no consistent split of the stored state carries it.

• Move ActiveData::region with the drawing, and a later local redraw asks a differently rounded question. Measured: (0,0)..(0,305.936) re-expressed as (0.5,-81.5)..(0.5,224.436) reads its length back as 305.93604, which crosses Span's leftover/no-leftover boundary -- the one that must be exact -- and draws a child a cold layout leaves undrawn.
• Leave region alone, and placed accumulates without bound, because try_reuse returns a clean subtree's size without walking into it: only the top widget's placed is recomputed while an ancestor's shift carries the whole subtree. Measured 7,048,813 where a cold layout says 456.

The replacement applies alignment in the two places that already exist and are exact. Where the size is known before drawing, from a rule, declared_box hands the child its aligned box directly -- one draw, no move. Where the size is only known after drawing, the widget is re-asked in its placed box, with its alignment forced to the near edge on the second ask so it terminates; that ask goes through try_reuse, which moves by recomposing, which tests/cases/drift.rs pins as exact. That deletes ActiveData::placed and shift_subtree. The cost is a second ask for a measured widget that is not near-aligned, which is exactly what Aligned cost before this work.

Placement compounds, which is what the align override is for

A container that reports a child's size while handing that child a bigger box gets its content placed twice: once by the child, once by the box around it. Found three times before the class was fixed rather than the instances -- Stack::size(Child(i)), Scroll's orthogonal axis, and Pad.

Painter::widget_aligned(child, region, align) is the override: an Option<RegionAlign> carried in DrawInfo and resolved once in draw_inner, so the root resolves like anything else. Pad, Stack and Scroll pass the near edge for children whose size they report. ActiveData keeps both the resolved alignment, so a local redraw asks the question its parent asked, and the widget's own, which is what a change is compared against -- an override means the answer is the parent's to give again.

Alignment as a property changed two examples, unnoticed

Found 2026-09-16 by looking at tabs beside canonical main rather than at its pixel count. d3b0ebf deleted Aligned, and 8220a78 deleted SetSize, so .align(), .center(), .sized(), .width() and .height() all set properties on one widget where each used to wrap it in another. Chaining them therefore means something different now, and two of it showed:

• A stack or pad child could not be aligned at all, because those containers overrode every child with the near edge. tabs's counters asked for Align::RIGHT and sat at the left; text's narrow panel asked to sit at the top of its row and filled it. Fixed in 2bc6bdf; both match main again.
• .width() overwrites what .sized() set on the same axis, so tabs's rrect.sized((100, 100)).center().width(leftover(2)) was a 100-tall bar across two shares where main draws a 100x100 square centred in them. Answered by d21a215 (Bryan, 2026-09-16): one widget keeps one length per axis, and the second length needs a second widget that does as little as possible. WidgetPtr already was that widget and is now Wrapper, with .wrapper() to make one -- Wrapper and not Wrap, and .wrapper() and not .wrapped(), so neither reads as the text setting.

text still differs from main in its aligned panel, and that one is a decision rather than a defect: 9d8415d deleted OrthoSize, so a Y span across itself is as long as its longest child instead of full width. The example needs a width, not the framework a change.

A widget occupies its box, and must not report more than it draws

Scroll reports Size::LEFTOVER on both axes: it clips its content to its box, so it can neither take less of one nor honestly ask for more. The content's length is what it scrolls through, not what it is. Reporting the content length instead made the framework place a 400-long drawing in a 200-long box, and placement by recomposition scales in that case, because a part stored at fraction 2 of its box stays at fraction 2 of a box twice as long. Reporting the content's cross length had the box around it place content already placed.

Where content shorter than the viewport sits is now Scroll's own alignment. Its Holds widening -- "content of a fixed length that fits sits at the start of any box it fits in" -- is therefore gated on near alignment: anchored anywhere else it is a part of the room left over, so it moves with every length the box takes and the drawing holds for that length alone.

Stack gives every child the box its sizing child defines, through the new Painter::box_of, for the same reason.

95fb4f9 adds the debug_assert this asked for, and narrows it to what is actually true: a widget that set a mask this draw must report inside the box it drew in. As "a reported size does not exceed its box" it fires on ordinary overflow instead -- measured, a hundred fuzzer trees produce thousands, every one a text too tall for the box it was offered, which is what a text is meant to say. The same commit fixed Masked, which was passing its inner's size up; tests/cases/scroll.rs has a clipping widget that reports its content, so the assertion is itself covered.

Fixed point

Layout decides on a grid rather than in floats, in four commits: 7548139 the number, 4e28f10 positions, bd6de71 lengths, 39e4ca2 the last of the pixels and Holds. Decided with Bryan on 2026-09-15.

• Fixed<SHIFT> is an i32 counting 1 / 2^SHIFT. Adding and subtracting are exact; a multiply drops to the step below and a conversion between grids takes the nearest step. Two routes to one place that land on one number are the same place, so everything downstream compares for equality. The multiply truncates as of 08c9d5a -- see the item in "Performance" for what that bought and what it costs. shift_round is still there for to_scale, which has no caller outside its own test.

• Px is 1/1024 px, Rel is 1/2^24 of a box, Weight is 1/65536 of a share. PX_SHIFT and REL_SHIFT are the only statement of the first two, and the shader's copy is prepended from them by render::module_source rather than written again in WGSL.

• A weight is not a fraction: a list divides its room by the total of its weights, so Weight trades precision for the range to hold a whole list, and Rel::ratio turns two weights into a share on the finer grid.

• Arithmetic wraps, as of 4febabf. It saturated first, so that a clamped coordinate kept the ordering a wrapped one inverts, and that cost a twelfth of layout's instructions to keep order two million pixels out where nothing draws. MIN and MAX still stand in for an unbounded end, which is safe only because every use compares against them and none adds to them; from_f32 is the one operation that clamps.

• Px was 1/64 first. The residue of a length reached two ways is one rounding, so it scales with the step: at 1/64 that was 0.016 px, enough to move a box, and at 1/1024 it is a thousandth of a pixel. Range is +/-2.1M px and conversion to f32 is exact to 16,384 px.

• What the fuzzers ask for is a step per level of nesting, which is two for these trees. Truncating made each of those steps a whole one rather than half of one; it did not add a level. Traced on 2026-09-16 to the same box reached two ways, each rounding where the other does not -- not accumulation, and not one place. Two of them are fixed in bdab558:

  • Scroll wrote a box it had been given back out as its own length in pixels. Centring a part in rel 1 lands a step from centring it in px 900, because a(x - y) and ax - ay do not round alike. Content that fills the viewport unscrolled is handed back as it came, and the repaint and resize-repaint cases became exact.
  • Span placed each child a step from where the last ended, carrying every share's rounding along the row. A position is now the fixed parts before it, exact, plus one rounded share. Two hundred equal shares of a 1000 px row ended at 999.999 and now end at 1000. What is left is a box centred in a fraction of its parent against the same box centred in its own pixels, one step per level between them. Closing it means alignment resolved in pixels everywhere -- which costs the retained resize path, since it is the fractional form that re-centres a subtree without redrawing it. Not worth it at a thousandth of a pixel.

• Holds::through inverts px + rel * box, which drops a step, so the answer is an interval even for a single length: inverting the length alone gives a point that need not contain the box the part was drawn in. What it allows for has two parts and they are not the same shape.

  • The two routes to a length -- composed down the chain against measured against the window -- get three half steps either side. That is the floor: two fires the Holds assertion on eleven generated cases. Four was tried and was overshoot, and tightening it moved none of the rig's twenty-five work counters, so slack here buys no reuse (38eba54).
  • The multiply on the way in gets one more step at the top of the range and nothing at the bottom, since truncation only ever drops. The whole of a box has no multiply in it, however many pixels are added to it, and allowing for one there compounded a step per level down a chain of widgets each taking the whole of its parent (60367d8).

  It cannot go to zero while a length reached two ways is two numbers. The range has to contain the box a drawing was made in, which the Holds assertion in draw_at checks, and it must not contain a box the drawing does not hold for, which the warm-against-cold oracle checks. Those two only coincide where composition is exact -- the i64 item under "Next" -- and shrinking this allowance is the test of whether that worked. Too wide is the unsound side of it. Run the generated cases in debug, since the assertion is compiled out of every release run.

• A pointer, a wheel notch, a shaped glyph advance and a window size arrive as floats and are put on the grid where they arrive. Vec2 stays what the GPU and the platform speak; PxVec2 is what layout decides in.

The leftover boundary

Fixed in 5ed9e87, which closed the shrinker's reorder case. Neither of its two red seeds was about reordering.

A span's box in pixels, compared with what its fixed and relative children fill, is the same number whenever the parent sized that box from the span's own answer -- and the box comes back through the chain a few bits off. So 0.00003 px decided whether a leftover-only child existed: warm rounded under and left it undrawn, cold rounded over and drew it at zero length. Both layouts are stable and the pixels are identical either way, which is why only the warm-against-cold oracle could see it.

A structural decision may not be taken where boxes structurally land. The fix was not a tolerant comparison -- that is what generated seed 16 punished -- but a boundary moved by HOLDS_EPSILON_PX of room, with the validity range split exactly at the moved boundary. The move and the constant are gone since 39e4ca2: on the grid the box a parent hands back and the sum of what the children asked for are whole counts of the same step, and both routes land on the same count, so the boundary needs no margin. What the section still records is why a decision may not be taken on a hair's breadth, and the regression that pins it. tests/cases/unsettled.rs's a_box_that_only_rounds_past_its_fixed_children_leaves_nothing_over is the six-widget regression, shrunk from 266; it needs the span above the one that divides, because without a box composed through it both trees round the same way.

Scroll's content_len <= container_len sits on the same coincidence but is continuous there -- it chooses between two Holds ranges that both contain the current length, so a rounding difference costs a redraw rather than a different layout. Checked while fixing this; nothing else reads a box in pixels to decide something structural.

Rigs and reproduction

Ordinary framework verification:

cd /home/bob/repos/iris-pr18
cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace

cb955f1 put the ordinary tests in tests/cases/, as modules of one tests/suite.rs target -- eleven links became one, and with profile.test's debug = "line-tables-only" a rebuild of iris's test targets went from 14.3 s to 7.7 s and target/ from 45 GB to 13 GB. Pick a module out with cargo test --test suite layout::. The fuzzers and the *_cost measurements are still their own targets.

The fuzzers take a thread per core but one (9d8415d), since a seed grows, lays out and drops its tree alone: the oracle's hundred seeds went from 68 s to 9.4 s and a shrinker case at 300 seeds from 18 s to 3.5 s. A failing seed still shrinks and panics on its own thread.

Run the long generated oracle only after ordinary tests pass:

cargo test --release --test generated -- --ignored a_long_run_of_seeds_agrees

tests/generated.rs compares a warm incremental tree with a cold tree of the same state. IRIS_GENERATED_SEED, IRIS_GENERATED_SEEDS, and IRIS_GENERATED_DEPTH select what it covers. tests/shrink.rs reduces a failing tree, over the same fifteen cases and the same trees since 98d4e98, so a seed the oracle fails on goes straight to it -- the failure prints the command. SHRINK_SEED, SHRINK_SEEDS, SHRINK_DEPTH, and SHRINK_CASE (a case name, or all) select what it tries. Turn what it finds into a test of its own rather than leaving a seed as the record.

The fifteen cases live in tests/scenario/mod.rs, which both targets include by #[path]. Adding one there gives it to both rigs at once; a case that only one of them knows is how the two drifted apart in the first place.

The float head is checked out at /home/bob/repos/iris-float-cmp, at 5ed9e87 with 394d514's Edits::fixed_branches cherry-applied to src/random.rs and tests/layout_diagnostics.rs but not committed, since the flag is the only way to measure the two on one tree. Build it and iris-pr18 with --test layout_diagnostics and run the binaries directly rather than through cargo. A comparison is only worth reading when the --features layout-diagnostics counters match on both sides; check primitive writes, widget draws, reuse remapped and placed by redrawing before any timing. Against the float head they do not match and will not, so that comparison is a bound rather than a measurement -- the pair worth measuring is two fixed-point commits, where all twenty-five can be diffed. Dump them with

IRIS_SEED=1 IRIS_DEPTH=8 IRIS_FRAMES=500 IRIS_PHASE=many \
  <instrumented binary> --ignored --nocapture \
  | grep -E '^ +[a-z].*[0-9.]+$' | grep -v ' ms$' | sort

and diff the two, which is the check that says a change is free.

The headless reference set must be run one process at a time because the rig reuses one compositor. Comparison worktrees need separate target directories. Useful commands:

./scripts/run-headless.sh tabs --mode 1920x1200@60Hz --shot /tmp/tabs.png
./scripts/run-headless.sh tabs --mode 900x1200@60Hz --shot /tmp/cold.png
./scripts/run-headless.sh tabs --mode 1920x1200@60Hz \
  --resize 900x1200@60Hz --shot /tmp/resized.png
./scripts/run-headless.sh tabs --mode 1920x1200@60Hz \
  --replay /tmp/tabs.touch --shot /tmp/replay.png

The replay used for the final check was:

0 down 1728 24
80 up 1728 24
400 down 1836 1116
480 up 1836 1116
800 down 1836 1116
880 up 1836 1116

tests/layout_diagnostics.rs is the retained CPU rig. Select cold, many, repaint, size, scroll, or resize with IRIS_PHASE; use the feature for explanatory counters and an uninstrumented release binary under perf for instruction totals.

Next

The next small LAYOUT.md §2 item is LazySpan. Region nodes now cover the independently movable-subtree use case; do not restore a separate child-placement API. 2d86058 brought docs/LAYOUT.md §2 and §3 to what shipped; §4, §5 and the density section are still stale -- they name Painter::place, SetSize, desired_width, apply_rest, Len::dp, Aligned and MaxSize, none of which exist.

Built-in alignment and size goes on #18 rather than after it (Bryan, 2026-09-15: #18 is unreviewed and already large enough that most lines get read anyway). The size half landed as 8220a78, the alignment half as d3b0ebf.

Do not restore OnResize::Translate; retained translation is now expressed by the same Holds contract and box chain.

Queued from this work, in order:

• Scroll should take a direction rather than one axis: vertical, horizontal, or both. Reporting LEFTOVER on both axes is already the right shape for it.

• Restore max_width/max_height as SizeRule::{Min, Max, Clamp}. 8220a78 deleted MaxSize and its builders, so that is public API owed back. A clamp resolved where the box is decided also fixes MaxSize's reading of px_size, which pinned its interval to one exact box on both axes and redrew its whole subtree on any resize. The clamp boundary is a hard layout decision: its Holds range must be exact and split at the crossover, the way generated seed 16 taught for Span. On the grid the crossover no longer needs moving off where boxes land -- see "Fixed point" -- but it does need both sides of the comparison to be Px.

  A cap may not contain leftover (found 2026-09-16, answering whether the parent could resolve one). Exact works with a share in it because it makes the widget's report moot -- ruled() is true, nothing reads the report, and the parent divides a weight that does not depend on what it divides. A cap must read the report, so rule and report are in one equation, and a share puts the division into it too: a child capped at one share contributes its drawn pixels to the row's total while it fits and a weight once it does not, which moves the room, which moves the share. The two assignments are each self-consistent, which is the multiple-fixed-point failure that generated seed 13 punished for orthogonal sizing -- and resolving it the way flexbox does, by freezing violated children and dividing again, costs an ask per round where the retained contract allows one offer per child. min(report, cap) is also not a Len: Len is a sum of the three parts, and the smaller of two of them is not.

  So a cap takes pixels and a fraction and no share, which is what Len is since a8898aa. Still awaiting Bryan: whether a Max narrows the box the child draws in, or only what the parent reports for it.

• Compose in i64 and narrow only when storing (Bryan, 2026-09-16). Chosen for the share spread; what it buys is not that, found on 2026-09-16 and worth reading before starting.

  The range inside a single operation is already protected -- Fixed::mul widens to i64 for the product, Holds::through does all of its arithmetic in i64 and narrows once at the end, and div_toward takes i64. What is not protected is precision across a chain: UiSpan::within composes a box through its parent with four multiplies and every one comes back to the i32 grid before the next starts, which is the "step per level of nesting" in "Fixed point" and what forces Holds::through's two-route allowance. A Wide<SHIFT> held through a composition and narrowed when it is written to ActiveData or a primitive would round once instead of once a level, at no cost in stored size.

  But it cannot make equal shares the same number of screen pixels, because the CPU's grid is not what puts them on the screen. prelude.wgsl decodes the raw rel/px counts into f32, walks the move chain in floats -- with the comment saying so, "what has to hold is that this agrees with itself frame to frame, not that it matches the CPU to the last bit" -- and then lands each edge with snap_floor(rel * dim) + snap_floor(px), which is a whole pixel. Measured: every box edge in the tabs render is a hard edge, no partially covered column anywhere along the band, and the two pad(10) gaps are exactly ten pixels on both sides of the truncation change. So the screen invariant is already as good as integers allow: a length in pixels is that many pixels, and equal shares differ by at most one whole pixel because three equal integers cannot sum to 1000. Layout's own one-or-two step spread is below what the shader can express, and shows only where it pushes a value across the floor -- which is what moved tabs's corner arcs by a pixel at 08c9d5a.

  What is left to buy is layout's own decisions, which is not nothing: the Span leftover boundary and the clamp crossover are structural decisions taken on a pixel comparison, Holds validity is a pixel interval, and warm-against-cold agreement is the thing every rig here measures. The success test is Holds::through's two-route allowance shrinking from three half steps toward zero, in debug where the assertion is live.

  Two things to decide before starting: how many fractional bits the intermediate keeps, since chained multiplies accumulate them and an i64 runs out too; and whether it keeps UiSpan::within inside the inliner, which is what actually moves cycles there -- a Wide that makes the body bigger loses on the axis the truncating multiply won on.

Other queued work, in dependency order:

• UiRenderState behind Rc<RefCell<_>>.
• Split Len/layout length and add density-independent pixels.
• Input restructuring: pointer capture, drag slop and axis, cancellation, mask-aware hit testing, and timestamps.
• Retained paints, selection, overlays, and shared runtime state.
• Generic desktop/Android hosts and reusable example/APK tooling.
• Application-owned fonts and replaceable glyph-atlas buckets.
• Positioned text overflow and cluster-safe ellipsis.

The archive is a reference, not a patch: it predates returned Size, the current box chain, and the current length types. Recreate changes on current types and keep app/session concepts out of Iris.