iris/ai-app
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Land the layout repair's findings and retire the working log

Browse Source 
`docs/LAYOUT_LOG.md` existed until the one-ask protocol landed. It has, as
Iris PR #19 replacing #18, so what has to outlive the log moves into
`docs/LAYOUT.md`: the hint rule and the one ask a span makes, the reported
shares that are left and the two-answers design that would remove them, what
the fuzzers reach and what they cannot, the hypotheses not to repeat, and the
counter and dump commands with the numbers they print now. The handoff becomes
what it is for -- where the work stands, which is waiting on a review.
This commit is contained in:
iris-ai committed 2026-09-19 02:57:17 -04:00
1 parent e4fada298d
commit 3ecd676bcf
3 files changed
+191 -1183
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

docs/HANDOFF.md
+67 -504
View File
@@ -1,516 +1,79 @@
# Handoff
Where the Iris retained-layout work stands for a worker picking it up cold.
This file contains current decisions, the implementation plan and its checks.
The durable layout design and measurement method are in `docs/LAYOUT.md`.
The temporary investigation record is in `docs/LAYOUT_LOG.md`; delete that
log when the one-ask protocol lands, after moving any fact that must survive.
Where the work in flight stands. The durable layout design, the measurement
method and the findings that outlived the working log are in
`docs/LAYOUT.md`.
## Where things stand
## The Iris layout repair is submitted
Canonical upstream Iris `main` is **`ca2b4b2`** (#17, the headless rig). PR
#18's pushed branch is `split/18-position-chain` at **`e44dea3`**. Its
detached comparison checkout is `/home/bob/repos/iris-layout-baseline`. It is
the reviewed baseline this work must preserve or improve.
**Iris PR #19** (`layout/one-ask` at `cadfba0`, 123 commits over upstream
`main` at `ca2b4b2`) is open and waiting on Bryan. It **replaces #18**, which
is closed and whose branch `split/18-position-chain` stays as the reviewed
baseline the newer work was measured against. Bryan decided on 2026-09-19
that the experiment replaces it, because it is more performant where it
matters.
The continuation is `/home/bob/repos/iris-layout-experiment`, now on branch
**`wip/hint-first`** at **`f6242aa`**, one commit over `wip/one-ask` at
`a888717` (nine commits over `4328eac`, the head of `wip/transparent-frames`,
which is unchanged). Both branches are pushed to `origin`. `wip/one-ask`
replaced the old step 3 plan with the one-ask protocol below, `1512d84` and
`23523ee` made the frame a length of the window, and `e8a5792` is what the
step 1 review found. `f6242aa` is the step 5 result: a span takes a child's
length from its hint or rule and asks it once, in its slot. It passes every
check:
The checkout is `/home/bob/repos/iris-layout-experiment`, on branch
`wip/hint-first`, which is the same commit as `layout/one-ask`. Both are
pushed to `origin` (`iris-ai/iris`). What the branch does and what it leaves
undone is in the PR description; the settled protocol is in `docs/LAYOUT.md`
under "Frames, decided boxes and padding".
| check at `f6242aa` | result |
Verification at `cadfba0`, all run in this checkout:
| check | result |
| --- | --- |
| `cargo fmt --all --check`, clippy `-D warnings`, with and without `layout-diagnostics` | clean |
| `cargo test --workspace` (debug) | 123 suite, 20 core, 11 generated, all green |
| `cargo test --workspace` | 123 suite, 20 core, 11 generated, green |
| `cargo test --release --test generated` | 11/11 |
| shrinker, 400 seeds, depth 5, all sixteen cases | agree, 63 s (34,488 widgets) |
| 1000 seeds at depth 6 | agree, 277 s |
| 2000-seed depth-4 scan, all sixteen cases | agree, 283 s (82,203 widgets) |
| cold layout of 400 trees at depth 5, dumped and diffed against `a888717` | byte-identical (34,488 widgets) |
| renders, the `tabs` replay and the `random` resize | **not re-run since `a888717`**; step 3 below |
The cold-layout dump is `tests/layout_dump.rs`, new at `f6242aa`: it prints
every widget's box for many grown trees so two commits can be diffed on cold
layout, which the warm/cold oracle cannot see a change to.
What implementing it corrected in the plan is in `docs/LAYOUT_LOG.md`; the
four that would have shipped as wrong layout are a share inside padding
losing the padding twice, a root resolving its own rule twice, a rule changed
over two pads relocating the column under them instead of dividing it again,
and a resize leaving a short scroll's window-tall content where it was. Each
is pinned as a named test.
The renders and the replay are done and recorded in `docs/LAYOUT_LOG.md`:
against #18, `minimal` and `tabs` are byte-identical (before and after the
reference gesture), `random` differs in two pixels of glyph antialiasing,
`text` moves one padded block one pixel, and a live resize of `random`
matches a cold render at that size byte for byte. Re-run at `a888717`, all
five are byte-identical to the same renders at `a30971e`, so the scroll fix
below changed none of them.
Not done: the pre-submit review of `f6242aa`, the render set at it, the doc
pruning and the landing (steps 1, 3 and 6 below). Step 5's measurements and
the two ideas it tested are in `docs/LAYOUT_LOG.md` under "What the step 5
measurements found".
The worker's older step 3/4 experiment is preserved as branch
`wip/step3-experiment` (one commit over `4328eac`) and as
`~/repos/iris-step3-experiment.patch`. It is evidence, not the protocol.
The app's Iris pin is unchanged.
## The two rules to protect
These outrank the accumulated machinery:
1. A changed tree lays out exactly as if it had been drawn that way from the
start. The warm/cold oracle and shrinker test this.
2. Lengths are predictable. `px` is that many pixels; `rel(0.5)` is half of
the frame decided for the widget, wherever it sits; `leftover` is a share
of the room left after every sibling's `px` and `rel` lengths are resolved.
Do not fix a failure with a tolerance, another measurement flag, a special
case in `Span`, or another layout method.
## What the previous plan got wrong
The full account is in `docs/LAYOUT_LOG.md`. The short version, because it
is the third plan for this repair and the next one should not repeat it:
- **Every plan kept the second draw.** The old protocol drew a widget in the
box it was asked in, then drew it *again* in the box its own answer placed
it in whenever the first drawing's `Holds` did not cover that box. All the
offer machinery -- `offer_place`, `offer_part`, `at_offer`, `measured()`,
the local-redraw deferral -- existed to remember which of the two draws was
the question. The plans tried to define that bit better; the defect was
that there were two draws at all.
- **The step 3 plan then over-corrected.** It said "every drawing must hold
for the answer box it supplies", and the worker implemented exactly that as
an assertion in `place`. A wrapped `Text` asked at 45 px whose longest word
is 89.5 px cannot satisfy it, and neither can any widget that reads its box
and reports something other than it. The answer box is not a question, so
no contract about it can be demanded of the widget.
- **It also let a caller narrow a frame by position.** A frame narrowed to a
region (the worker's share frames) does not move when the part it sits in
moves; only a frame narrowed to a *length*, put back into the part on
every placement, does.
## The protocol now in the experiment
**A widget draws once, in the box it is asked in. Its answer is placed inside
that box by re-expressing the drawing. Nothing is drawn again in a box an
answer chose.** `Holds` is a contract about the ask box alone, consulted only
to decide whether a re-ask can be skipped. This is `draw_inner` at `1512d84`:
```rust
let reused = (!stale)
.then(|| self.retained_answer(id, part, info))
.flatten()
.and_then(|answer| {
let extent = placed_extent(part, answer.0, declared, info.fill(), align);
self.try_reuse(id, part, extent, info, rsc).map(|()| answer)
});
let answer = reused.unwrap_or_else(|| {
if old.is_none() {
old = self.remove(id, false, rsc);
}
let answer = self.draw_at(id, part, info, old.take(), rsc);
let extent = placed_extent(part, answer.0, declared, info.fill(), align);
if extent != part {
self.relocate(id, extent, info, rsc);
}
answer
});
```
`try_reuse` checks the drawing against `part` and relocates it to `extent`;
the old `place` (redraw in the answer box) is gone, and with it every offer
field's purpose. `ActiveData` keeps `part` as the ask box, `asked` as where
it was asked and `placed` as where it was put (renamed in step 4).
A container that puts an answer somewhere other than where it asked says so
with a new call that never runs the body:
```rust
/// Puts a child asked about in this draw somewhere else in this
/// widget's box: its answer, placed in this part instead. The drawing
/// is re-expressed there rather than made again -- what a row does once
/// it knows every slot, having measured each child from its cursor.
pub fn place_at<W: ?Sized>(&mut self, id: &StrongWidget<W>, place: [Place; 2])
```
A frame is narrowed by a *length of the window*, never a region and never a
fraction of the parent's frame -- a row's slot cannot be written as a
fraction of the row. The box stays whatever `place` names; only a declaration
places the box inside it, by the child's alignment, on every placement:
```rust
pub fn widget_at<'s, W: ?Sized>(
&'s mut self,
id: &'s StrongWidget<W>,
narrow: [Option<Len>; 2],
place: [Place; 2],
) -> DrawResult<'s, 'a, W>
```
`Span` takes each child's length along itself from its hint where one says
-- a rule, or a widget that always reports the whole of its box -- and
otherwise asks the child once from its cursor (`Within(From(cursor..far))`).
Then it moves a drawn fixed child to its slot, and asks every other child
once in its decided slot, a share with the frame narrowed to it. This is
`Span::draw` at `f6242aa`:
```rust
let size = match painter.size_hint(child, axis) {
Some(len) => {
measured.push(None);
len
}
None => {
let room = Place::Within(Part::From(along(cursor, far)));
let size = painter
.widget_at(child, [None; 2], axis.pair(room, across))
.size();
measured.push(Some(size));
size.axis(axis)
}
};
...
let slot = along(from, start);
let place = axis.pair(Place::Fill(Part::From(slot)), across);
let mut narrow = [None; 2];
if len.leftover > Weight::ZERO && shares {
narrow[axis as usize] = Some(slot.len());
}
let used = match (measured, narrow[axis as usize]) {
(Some(size), None) => {
painter.place_at(child, place);
size.axis(!axis)
}
_ => painter.widget_at(child, narrow, place).len(!axis),
};
```
A hint is the length the child's draw would report: `Painter::size_hint`
resolves a fraction in it against the asking widget's frame, which is the
frame a child asked with nothing narrowed gets, and pins that frame where it
did. `Scroll`, `Masked`, a `Stack` without a sizing child and the fixture's
`Branch` hint `LEFTOVER`, since each always reports it; `Rect` and `()`
already did. A widget that reports a share its children gave it -- a span
with a `Rect` in it -- has no hint and is still asked twice, which is the
whole of what step 5 left.
`Stack` asks non-sizing children with `Part::Sized(len)` of what its sizing
child decided, on every axis that is not a share -- a box of that length
where their own alignment puts it, and that length as their frame -- and
`Scroll` asks its content once in the viewport and `place_at`s it to the
scrolled offset.
A local redraw asks the retained question again -- the same place of the box
the parent was *asked* in -- and, if the answer stands, puts the fresh drawing
back at the retained place of the box the parent's answer *chose*. Both halves
are needed: seed 2 at depth 4 (a stack sized by its text) fails without the
second.
A widget its parent asked more than once in one draw -- a share child, asked
in the room and then in its slot -- has two questions and one record, so it
cannot settle locally: `redraw` defers it to the parent the way it defers a
widget whose declared length changed (`ActiveData::re_asked`, set by
`widget_at` when the child is already in `children`). Seed 946 at depth 6
found the case: a fixed-height column that is a share while its rect fits
and a fixed width once it does not, so emptying it changes the room answer
and not the slot answer.
A symbolic length a child pinned now composes through `Part::Of` where the
part is the whole box less pixels, and pins the parent's own length otherwise
(`in_parent`). Dropping it let a zero `Pad` reuse a drawing across a narrowed
frame of the same pixel length; the shrinker found six such seeds at depth 5.
## Decisions
Decided with Bryan on 2026-09-17 and 2026-09-18, kept where still true.
### One draw method, in a box decided from above
`Widget::draw` remains the only layout method. A container's body runs only
in a box its parent offered or decided, never in a box derived from the
container's own answer. **The experiment extends this to every widget:** a
leaf is not drawn in its answer box either. Its drawing is re-expressed
there, which for a text means the block it shaped at the asked width is
positioned inside the box its reported size chose, and its lines do not
change. `examples/text.rs` and `random` have not been rendered since; do that
before landing and inspect any change.
### Frames are narrowed by every length decided from above
A declared `px` or `rel`, a resolved share, and the box a stack's sizing
child decided (Bryan, 2026-09-18: the sizing child, if any, determines how
the rest are laid out) all narrow the frame.
`declared_lens` still excludes `leftover`, which is right: a share has no
length until the span divides its room, and it narrows the frame at the
placing ask instead.
### Padding is an inset, and the frame is a length while the box is a region
Bryan, 2026-09-18: padding is an inset. It subtracts from both the child's
frame and its box and adds itself to the reported size, so `rel(1.0)` inside
padding fills the parent without overflowing. A span's frame never subtracts
siblings; only the padding subtracts from it. No outset kind and no mixed
kind for now; the name stays `Pad`. Worked example, 900 px row:
```rust
let row = (rect(Color::RED).width(24), wtext(PARAGRAPH).wrap(true).pad(16)).span(Dir::RIGHT);
```
The text is asked in 900 − 24 − 32 = **844** px and wraps there; a
`rel(1.0)` inside the same padding is 900 − 32 = **868** px and overflows the
row by exactly the icon's width. With the pad in a share instead, both are
the share less 32. An icon *after* the padded text overflows; a user who
wanted otherwise meant `leftover`.
Implemented at `1512d84`. `Pad` reads its own frame (`Painter::frame_len`,
which pins it), takes the padding off, and hands that down as the child's
frame, while the box it gives is the inset part of its own box. The two are
different lengths whenever the box is narrower than the frame -- which is
exactly the wrapping case above.
### A share never adds room beyond the deciding box
`Scroll` resolves its content length from the fixed part of the answer and
makes it at least the viewport (`4328eac`). Unchanged.
### Existing fixed-point and box-chain design stays
Unchanged; see `docs/LAYOUT.md`.
## Implementation plan
Work in `/home/bob/repos/iris-layout-experiment` on `wip/hint-first` from
`f6242aa`. Make each step a warning-clean commit, push it, and run its named
checks before the next. If a step exposes a different mechanism, stop and
update this handoff rather than papering over it.
The order is 1, 3, 6: review the step 5 commit, render at it, then land.
Steps 2, 4 and 5 are done and kept below for what they decided.
Anything a fuzzer finds is shrunk first (`SHRINK_SEED=<seed>
SHRINK_DEPTH=<depth> SHRINK_CASE=<case>`, which now prints each level's
frame, ask, box and size warm against cold), pinned as a named test in
`tests/cases/unsettled.rs`, then fixed under the two rules above. Do not add
a second draw back.
### 1. Review the step 5 commit
The six commits `4328eac..a888717` were reviewed as one diff (done at
`a888717`; it found one wrong layout, fixed at `e8a5792` and pinned by
`scroll::content_that_fits_is_placed_in_the_viewport_and_not_in_the_window`,
and three stale comments). **Still to do: the pre-submit review of
`a888717..f6242aa`**, which the planner wrote and measured but did not
review as a separate pass. It touches `Span::draw`, `Painter::size_hint`,
four `size_hint` implementations and the dump rig. Things to look at:
- `Painter::size_hint`'s pin: it sets `frame_own_len` where the hint has a
fraction. A child asked in its slot afterwards pins the same frame through
`in_parent` anyway, so the pin only matters for a caller that reads a hint
and never asks the child. Confirm that reading is right, or delete the pin
and say why.
- A hinted share child with no room is never asked and `Span` calls
`painter.undraw` on it, which is what takes its last frame's drawing down.
A drawn-in-the-room child got the same call before; check nothing else
relied on the room draw having happened first (`size_deps` records the
hint read, and `draw_at`'s `asked` handles a dep that is not a child).
- The two comments in `Span::draw` above the passes, and the `Stack` hint's
comment.
Check: the ordinary suite, then `cargo test --release --test generated`.
Anything the review changes in `Span` or `Painter` re-runs the three long
fuzzers and the cold dump (commands under step 6).
### 2. Make the frame a length and the box a region
**Done at `1512d84` and `23523ee`.** The settled rule is in `docs/LAYOUT.md`
under "Frames, decided boxes and padding"; what implementing it corrected in
the plan -- four of them wrong layout that would have shipped -- is in
`docs/LAYOUT_LOG.md`.
### 3. Render and replay
Done at `adbedaf` and re-run at `a888717`, against `e44dea3` (#18); what
each render showed is in `docs/LAYOUT_LOG.md`. **Not yet run at `f6242aa`.**
Render the five examples, the `tabs` replay and the `random` resize at the
reviewed head and compare with the same renders at `a888717` (a
`git worktree add ../iris-a888717 a888717` with its own target directory,
one process at a time). Expected: byte-identical, every one -- the cold
dump already says so for 400 random trees, but the fuzzer grows no `rel`
rules and the examples do, so `text` is the render that can still differ.
Any difference is a finding: read the records before touching anything.
`view` has no counterpart in #18 and is compared with `a888717` only. The
commands are under step 6.
### 4. Rename and delete
**Done at `a30971e`.** `ActiveData` and `DrawInfo` now say `part` for the box
a widget was asked in, `asked` for the place it was asked at and `placed` for
where its drawing was put; `LayoutHolds::frame` is `window`, since those
ranges are window pixels and the frame's own entry is the `frame_len` pin
beside them, and `Painter::frame_own` is `window_own`. `answers_at` had one
caller and is inlined there. `ActiveData::measured` is kept: `place_in` reads
it, and what it says -- the answer rather than the last drawing's report --
is worth a name.
### 5. Restore the expected retained cost
**Done at `f6242aa`**, by one change: a span takes a child's length from
its hint or rule and asks it once, in its slot, instead of drawing it in the
measuring room first. Work counters, seeds 1 and 13, depth 8, widget draws /
distinct widgets, beside `e44dea3` (#18) and `a888717` (before it):
| seed 1 | e44dea3 | a888717 | f6242aa |
| --- | --- | --- | --- |
| cold | 369/261 | 342/288 | 264/232 |
| many | 157/95 | 118/95 | 41/41 |
| size | 16/12 | 3/3 | 3/3 |
| scroll | 2 | 1 | 1 |
| resize | 13/13 | 44/13 | 36/13 |
| seed 13 | e44dea3 | a888717 | f6242aa |
| --- | --- | --- | --- |
| cold | 1330/707 | 1278/982 | 758/627 |
| many | 524/159 | 429/366 | 16/16 |
| resize | nothing | nothing | nothing |
The command is
```sh
IRIS_SEED=1 IRIS_DEPTH=8 cargo test --release --features layout-diagnostics \
--test layout_diagnostics -- --ignored --nocapture layout_cost
```
reading `widget draws` and `distinct widgets` from each phase's block, and
`hottest widget draws` with `IRIS_PHASE=resize IRIS_FRAMES=2` for who is
drawn how often.
`many`, `size` and `scroll` are at their floor: every draw is a marked
widget, or the parent a marked widget deferred to. `cold` is under #18 at
both seeds. What is left is `resize` at seed 1, 36 draws of 13 widgets where
#18 drew 13, and the `cold` draws over the distinct count (32 at seed 1,
131 at seed 13). Both are one mechanism, understood and not yet worth a
change: a **reported share** -- a span whose children report `leftover`, a
stack sized by a child, a wrapper round either -- has no hint, so its parent
still asks it in the room and again in its slot, and each of those asks
draws, since the room drawing divided the room and does not hold for the
slot. Under the root's resize redraw that multiplies down the tree: the
hottest widget at seed 1 is a `Span` drawn 8 times in one frame, two
levels of reported shares under it drawn 4. The design that would remove it
is in `docs/LAYOUT_LOG.md` ("two answers per record"); do it only once an
app screen shows the cost, and measure that screen first.
Two ideas from the previous version of this step were tested and are **not
to be done**; the measurements are in `docs/LAYOUT_LOG.md`:
- Not reading the span's own length where no slot depends on it
(`wip/inset` at `5b181bc`, pushed): identical counters at every phase.
The symbolic pin survives a resize because every ask box is symbolically
stable across one; the earlier "44 to 20" came from share spans reusing
their room drawing in their slot, which is wrong layout rather than saved
work.
- Resolving a rule that is a fraction of the frame from the record instead
of redrawing: `reuse outside: a frame length` is 3 draws of 264 cold at
seed 1 and 1 of 758 at seed 13. Nothing to recover.
Report every phase at both seeds, work counters first, medians only when the
work agrees.
### 6. Full verification and landing
Run, in the experiment checkout:
```sh
cargo fmt --all --check
cargo clippy --workspace --all-targets -- -D warnings
cargo clippy --workspace --all-targets --features layout-diagnostics -- -D warnings
cargo test --workspace
cargo test --release --test generated
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
SHRINK_CASE=all SHRINK_SEEDS=2000 SHRINK_DEPTH=4 \
cargo test --release --test shrink -- --ignored --nocapture
```
The last line is the 2000-seed depth-4 scan over all sixteen cases; the
shrinker runs the same cases as the scan and reduces anything it finds, so
no temporary test body is needed any more. `Rng::new` uses `seed | 1`, so
adjacent even/odd seed pairs describe the same tree. All of these passed at
`f6242aa`; re-run whatever a later commit could affect.
A change that could move cold layout also gets the dump diff, which the
oracle cannot replace:
```sh
IRIS_DUMP_SEEDS=400 IRIS_DUMP_DEPTH=5 cargo test --release --test layout_dump \
-- --ignored --nocapture | grep -E '^[0-9]+ [0-9]+ ' > /tmp/after.txt
```
once at the commit before and once after (check the earlier commit out in
the same checkout so the build is incremental; the rig file is untracked
there, copy it in with `git show wip/hint-first:tests/layout_dump.rs`), then
`diff` the two. Zero differing lines is the expectation for a cost change.
Render `view`, `minimal`, `random`, `tabs` and `text` at 1920x1200 and inspect
every intentional change. Also replay `tabs` and compare a live resize of
`random` with a cold render at the same size. Read the installed graphics
skill before rendering and confirm the renderer; an llvmpipe fallback can
produce a plausible PNG. The headless rig reuses one compositor, so run one
process at a time and give comparison worktrees separate target directories:
```sh
./scripts/run-headless.sh tabs --mode 1920x1200@60Hz --shot /tmp/tabs.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 reference replay is:
```text
0 down 1728 24
80 up 1728 24
400 down 1836 1116
480 up 1836 1116
800 down 1836 1116
880 up 1836 1116
```
Before submitting, run the pre-submit review. Then land: move any surviving
fact from `docs/LAYOUT_LOG.md` into `docs/LAYOUT.md` (the settled protocol,
the measurement method including the dump rig, and the "not to be done"
results under step 5), delete the log, update this handoff to the next
actual task, and push every coherent commit. How the branch reaches
upstream is Bryan's call and has not been asked: `wip/hint-first` is some
sixty commits over #18's `split/18-position-chain`, which is still open, so
either #18 lands first and this follows as one PR, or this replaces #18.
Ask before opening anything. Update the app's Iris pin only when the Iris
change is ready.
## Follow-on work, not part of this repair
- CPU round-to-nearest and shader nearest-pixel snapping are approved as one
separately verified change. Neither has landed. Re-derive `Holds::through`
for the new rounding and run both long fuzzers plus the render set.
- Smaller layout items remain in `docs/LAYOUT_LOG.md`: an undrawn share's
gap, nested share weights, inconsistent zero-divisor fallbacks, and the
stale `f32` identity comment.
| shrinker, 400 seeds, depth 5, sixteen cases | agree, 60 s |
| 1000 seeds at depth 6 | agree, 177 s |
| 2000-seed depth-4 scan, sixteen cases | agree, 295 s (82,203 widgets) |
| cold dump of 400 trees at depth 5, diffed against `f6242aa` | identical, 34,488 widget boxes |
| counters, seeds 1 and 13 at depth 8 | unchanged from `f6242aa` (the table is in `docs/LAYOUT.md`) |
| five renders, the `tabs` replay, a live resize against a cold render | all byte-identical to `a888717` |
The renderer was Vulkan on the host's RX 7900 XT: `src/default/render.rs`
asks for `Backends::PRIMARY` and `.expect()`s the adapter, so a PNG at all
means it was not the software fallback.
## What is next, in order
1. **Bryan's review of #19.** Fixes to it are unreviewed code: repeat the
`pre-submit-review` passes over whatever each round changes, and re-run
the three fuzzers and the dump diff for anything that touches `Span`,
`Painter` or `render_state`.
2. **The app's Iris pin**, once #19 lands. It is still on the old submodule
commit, and the branch changes API the app uses: `SizeRule` beside a
widget rather than a wrapper, alignment as a widget property, `rest`
renamed to `leftover`, `OrthoSize` gone.
3. **Round-to-nearest**, CPU and shader together as one verified change.
Bryan approved it on 2026-09-17 and neither half has landed; the
derivation, the form to use and what to re-check are in `docs/LAYOUT.md`
under "Rendering the grid (pending)".
## Smaller layout items, none urgent
- An undrawn `leftover` child still contributes its gap, so a vanished child
leaves a double gap.
- Nested spans pass `leftover` weight up, so three leftover children in one
inner span beside one in another get three quarters to one quarter. No
other layout system does that; confirm it is wanted.
- A span can overflow itself without bound, so boxes of negative length reach
children and nothing states what a widget may assume about one.
- `Fixed::div` by zero answers `MIN`/`MAX` while `ratio` answers `ZERO`; both
are caller bugs under `debug_assert`, but the fallbacks differ.
- The comment on the `local == UiRegion::FULL` shortcut in `widget_at` says
composing through `FULL` "is not quite the identity in f32"; on the grid it
is exact and the shortcut is performance only.
- `docs/LAYOUT.md` §4, §5 and the density section name `Painter::place`,
`SetSize`, `desired_width`, `apply_rest`, `Len::dp`, `Aligned` and
`MaxSize`, none of which exist. Do not restore `OnResize::Translate` or
`OrthoSize`.
- `LazySpan`, then `SizeRule::{Min, Max, Clamp}`. A cap may not contain
`leftover`; whether `Max` narrows the child's drawing box is still a real
product decision.
`leftover`; whether `Max` narrows the child's drawing box is a product
decision.
- `Scroll` taking a direction rather than one axis.
Other product work remains in `docs/PLAN.md` and the focused documents it
links. Do not mix it into the Iris layout branch.
Other product work is in `docs/PLAN.md` and the focused documents it links.
Do not mix it into the Iris layout branch.
docs/LAYOUT.md
+124 -8
View File
@@ -4,12 +4,13 @@ A widget draws once and records its size on the `Painter`. Reading a child
`DrawResult::size()` records a retained size dependency; drawing the child
without reading that result does not make the parent's size depend on it.
§1, §2 and §3 have landed in Iris (#16 and #18). §4 to §6 and the density
section retain the rationale of the design but still name types that have
since been replaced; they are not an API reference. `docs/HANDOFF.md` is the
current transparent-frames work and its checks; `docs/LAYOUT_LOG.md` is what
the sessions doing that work found, kept until it lands. The sections at the
end of this file are durable design moved out of the handoff on 2026-09-18.
§1 landed in Iris as #16. §2 and §3 were #18, which is closed: Iris PR #19
carries those commits whole plus the retained-layout repair built on them. §4 to §6 and the density section
retain the rationale of the design but still name types that have since been
replaced; they are not an API reference. `docs/HANDOFF.md` is where the work
in flight stands. The sections from "Frames, decided boxes and padding"
onwards are the settled design, the findings that outlived the working log,
and the measurement method.
## Design
@@ -333,7 +334,38 @@ that box by re-expressing the drawing, and nothing is drawn again in a box an
answer chose. `Holds` is a contract about the ask box alone, read only to
decide whether a re-ask can be skipped. A container that puts an answer
somewhere other than where it asked says so with `Painter::place_at`, which
never runs the body. This is `wip/one-ask` in the experiment checkout.
never runs the body.
The answer box is never a question, so no contract can be demanded of a
widget there: a `Text` asked at 45 px whose longest word is 89.5 px cannot
promise its drawing holds for the box its own answer chose. An earlier plan
asserted exactly that and was unsatisfiable.
A container that can learn a child's length without drawing it does.
`Painter::size_hint` answers from the child's rule, or from
`Widget::size_hint` where it implements one, resolved against the asking
widget's frame; `Scroll`, `Masked` and a `Stack` without a sizing child hint
`LEFTOVER`, since each always reports the whole of its box. `Span` takes each
child's length from its hint where there is one and asks that child exactly
once, in the slot it decided; a child with no hint is drawn in the room left
from the cursor, because a text has to wrap at the width actually there, and
its drawing is moved to its slot with `place_at`. Reading a hint records a
size dependency and pins the frame where the hint declared a fraction, the
same pin a rule that is a fraction takes.
**What that leaves, and the design that would remove it.** A widget that
reports a share its children gave it -- a span whose children report
`leftover`, a stack sized by such a child, a wrapper round either -- can have
no hint, so its parent asks it in the room and again in its slot, and both
asks draw, since the room drawing divided the room. Under a resize that
multiplies down the tree: at seed 1, depth 8, the hottest widget in the
resize frame is a `Span` drawn 8 times. Removing it means two answers on one
record -- a room answer from an explicit measuring ask (`Painter::measure`,
so which ask is the measurement is stated rather than inferred) and a slot
answer from the placing ask, with the drawing belonging to the placing one --
plus a local redraw that re-asks both questions and marks the parent if
either answer moved, which also retires `re_asked`'s deferral. Do it only
once an app screen shows the cost, and measure that screen first.
**There is one coordinate unit, the window** (`1512d84`, `23523ee`). Every box in the
tree is a region in window units; a widget's frame is a *length* in the same
@@ -489,14 +521,68 @@ of its parent's frame and a local redraw walked back up the parent chain in
- Alignment is one value per axis and defaults to the middle because neither
edge is neutral without a direction. One widget has one length per axis; a
second length requires a second widget through `.wrapper()`.
- No measurement is a different state from a measured zero, so
`ActiveData::answer` is an `Option` rather than a zero size.
### What the fuzzers tolerate
### What the fuzzers tolerate, and what they reach
Warm and cold pixel regions must compare exactly; there is no step
allowance. When a row's grid-step count is not divisible by the number of
children, individual share widths differ, but every rerun of that layout
must still agree exactly.
- The routine runs are ten seeds (fast oracle), 400 at depth 5 through the
shrinker, and 1000 at depth 6. The 2000-seed depth-4 scan over every case
is what found seeds that had failed on every commit for a week, so run it
after any layout change. `Rng::new` is `seed | 1`, so an even seed and the
odd one above it are one tree.
- The shrinker panics per thread at the first failing seed, so a run reports
at most one seed per chunk of about 58. Check a single seed against the
unpatched code before calling it new.
- Depth finds things and so does breadth; widen one axis at a time and
record which.
- If a seed fails once and never again, suspect the rig's determinism before
believing an edit fixed it: seed 30 at depth 5 did exactly that on
2026-09-19 and reverting each part of the change did not bring it back.
- The cases are what a change is *then* compared at. Until the `size-resize`
case was added, every one of them compared at the window the change was
made in, and a length kept as a fraction of the wrong box agrees there and
parts from cold at every other window. There is still no case that changes
a tree twice with a resize between, and none that resizes twice.
- The fuzzer grows no `rel` rules and never re-parents a widget, so neither
a fraction resolved against a frame nor a subtree changing hands is
covered by any generated tree. Both have cost a defect that only a
hand-written expectation or a render caught.
### Failed hypotheses worth not repeating
- **"Which draw is the measurement" cannot be defined.** Four bookkeeping
rules and two experiments each fixed some seeds and broke others. The bit
existed only because a widget was drawn twice; one draw leaves it nothing
to name. A plan that proposes bookkeeping for a distinction should first
ask whether the distinction has to exist.
- **A frame narrowed by a region does not move when its part moves.** A
narrowed frame must be a *length*, put back into the part on every
placement, exactly as a declared rule already is.
- **Choosing between a fixed and a relative child in pixels at the span's
current width** admits multiple self-sizing fixed points; seed 13 settles
differently warm and cold under it. The same circularity is what a cap
containing `leftover` would put into `SizeRule::Max`.
- **A tolerant endpoint on the span's leftover split** retains zero-height
children (seed 16). The boundary is derived through the inverse of the
expression that places children, and is pinned by
`unsettled::a_box_that_only_rounds_past_its_fixed_children_leaves_nothing_over`.
- **Not reading a span's own length where no slot depends on it** saves
nothing (`wip/inset`, measured 2026-09-19): counters identical at every
phase. A symbolic pin survives a resize because every ask box is
symbolically stable across one, so the pin only fails where an answer
changed, which is a real relayout.
- **Resolving a rule that is a fraction of the frame from the record instead
of redrawing** recovers nothing: `reuse outside: a frame length` is 3 of
264 cold draws at seed 1 and 1 of 758 at seed 13.
- **A `git bisect`** once named a commit that could not be the cause; read
the tree rather than the bisect when that happens.
### Rendering the grid (pending)
`snap_floor` in `prelude.wgsl` adds half a layout step before flooring,
@@ -515,6 +601,36 @@ inlines.
## Measuring layout cost on this machine
- **Counters, at two seeds, before anything else.**
IRIS_SEED=1 IRIS_DEPTH=8 cargo test --release --features layout-diagnostics \
--test layout_diagnostics -- --ignored --nocapture layout_cost
reports each phase's `widget draws` and `distinct widgets`, and
`IRIS_PHASE=resize IRIS_FRAMES=2` names who is drawn how often. Where the
one-ask protocol stands, draws / distinct at depth 8, beside #18:
| seed 1 | #18 | now | | seed 13 | #18 | now |
| --- | --- | --- | --- | --- | --- | --- |
| cold | 369/261 | 264/232 | | cold | 1330/707 | 758/627 |
| many | 157/95 | 41/41 | | many | 524/159 | 16/16 |
| size | 16/12 | 3/3 | | resize | nothing | nothing |
| scroll | 2 | 1 | | | | |
| resize | 13/13 | 36/13 | | | | |
`many`, `size` and `scroll` are at their floor: every draw is a marked
widget, or the parent a marked widget deferred to. `resize` at seed 1 and
cold's draws over its distinct count are the reported-share cost above.
- **A change that could move cold layout gets the dump diff**, which the
warm/cold oracle cannot replace -- both of its sides move together:
IRIS_DUMP_SEEDS=400 IRIS_DUMP_DEPTH=5 cargo test --release \
--test layout_dump -- --ignored --nocapture | grep -E '^[0-9]+ [0-9]+ '
once at each commit, then `diff`. Zero differing lines of 34,488 is the
expectation for a cost change.
- **Check the work counters before comparing two commits' times.**
`tests/layout_diagnostics.rs` prints drawn widgets, widget draws and
primitive writes; a comparison is only worth reading when they match.
docs/LAYOUT_LOG.md
-671
View File
@@ -1,671 +0,0 @@
# Layout findings log
What the sessions working on Iris's retained layout found, planner and
worker alike, kept so that nothing here is rediscovered. Each entry says who
found it and when. **Delete this file when transparent frames lands**; what
must outlive it (settled design, the measurement method) is already in
`docs/LAYOUT.md`, and the current plan is in `docs/HANDOFF.md`. Commit ids
are in `/home/bob/repos/iris-layout-experiment` unless said otherwise.
## What the step 5 measurements found (planner, 2026-09-19)
Re-measured `a888717` first: every number in the handoff's table reproduced
exactly (342/288, 118/95, 3/3, 1, 44/13 at seed 1; 1278/982, 429/366 at
seed 13), so the table was trusted. What the counters said beyond it:
- **Most redraws never reached the reuse check.** `many` at seed 1 made 118
draws and only 78 reuse attempts, 4 of them rejected; the other draws
came from `retained_answer` failing before `try_reuse` ran. Mechanism: a
share child is asked in the room and then in its slot, one record holds
one answer, and each ask overwrites it -- so the room ask finds the slot
answer (pinned to the slot's length) and the slot ask finds the room
answer, and each draws. Every ancestor redraw pays it twice per share
child, multiplied down nested shares.
- **#18 avoided the room ask where a hint or rule gave the length.** Its
`Span` called `painter.known_len` first; `3091fb8` dropped that with the
offer machinery. Restored at `f6242aa` as `Painter::size_hint`, which now
resolves a fraction against the asking widget's frame (it returned the raw
rule before, which no caller had noticed because its one caller read
`px`) and pins that frame where it did. `Scroll`, `Masked`, a `Stack`
without a sizing child and the fixture's `Branch` gained `LEFTOVER`
hints. A hinted child is asked once, in its slot; a hinted share with no
room is never asked and is undrawn. Cost (draws / distinct, depth 8):
seed 1 cold 264/232, many 41/41, size 3/3, scroll 1, resize 36/13; seed
13 cold 758/627, many 16/16, resize nothing. `many` is at its floor:
every draw is a marked widget or the parent one deferred to. Cold layout
of 400 trees at depth 5 is byte-identical to `a888717` (the new
`tests/layout_dump.rs`), and the three long fuzzers agree.
- **What is left is reported shares.** With hints in, the hottest widget
in seed 1's resize frame is a `Span` drawn 8 times, two more drawn 4 and a
`Branch` drawn 4: a span whose children report `leftover` has no hint, is
asked in the room and in its slot, and both draw because the room drawing
divided the room (`far` read, pinned to that length). Under the root's
resize redraw that is 2 per level of nesting. It is also the whole of
cold's draws over its distinct count.
The design that removes it, not implemented: keep two answers on the
record, the room answer from a measuring ask and the slot answer from the
placing ask, with the drawing belonging to the placing ask. A measuring
ask (`Painter::measure`, a new call the parent makes explicitly, so
"which ask is the measurement" is stated rather than inferred) reuses the
room answer when its holds contain the room and touches no drawing;
otherwise it draws in the room and records the answer as the room's. The
placing ask is `widget_at` as now. A local redraw of a twice-asked child
re-asks both questions, marks the parent if either answer moved, and puts
the drawing back otherwise -- which also retires `re_asked`'s deferral.
The cost is a second `Option<(Size, LayoutHolds)>` on `ActiveData` and a
measuring path through `draw_inner` that skips `try_reuse` and `relocate`.
This is the one-record-two-questions bookkeeping the earlier plans died
of, made explicit at the call site instead of recovered from the ask; it
is worth doing only once an app screen shows reported shares nested
under a resize, and that screen should be measured first.
- **Dropping the span's `far` pin does nothing** (`wip/inset` at `5b181bc`,
pushed, one commit over `f6242aa`): `Part::Of` replaced by
`Part::Inset { lead, trail }` in window lengths, an exact `in_parent` arm
for a pixel inset (the range moved by the pixels, `Holds::longer_by`) and
a pinned own length plus a window range for one with a fraction, `Pad`
speaking it, and `Span` reading `extent_len` only with shares or in the
negative direction. Suite, oracle and clippy green; counters identical to
`f6242aa` at every phase of both seeds. Why: a symbolic pin survives a
resize because every ask box is symbolically stable across one -- the
viewport a scroll asks its content in is `rel 1`, a slot is a sum of
pixel answers, a stack's sized part is an answer -- so the pin only fails
where an answer changed, which is a real relayout. The worker's "44 to
20" at `a888717` was a deliberately unsound read, and what it saved was
share spans reusing the drawing that divided the room in the slot that
is narrower: wrong layout, not spared work. Not to be repeated; the
branch is evidence.
- **The frame pin on a rule that is a fraction costs nothing** to speak of:
`reuse outside: a frame length` is 3 of 264 cold draws at seed 1 and 1 of
758 at seed 13. Dropped from the plan.
## What making the frame a length found (worker, 2026-09-19)
Step 2 of the handoff, implemented over `0ef87eb` in
`/home/bob/repos/iris-layout-experiment`. Six of these are corrections to the
plan, not to the code that implemented it, so they are worth keeping even
after the step lands.
- **A narrowed frame is a window length, not a fraction of the parent's
frame.** The plan said both at once: `Pad` narrowing by `rel 1 - 32px` (a
fraction) and `Span` narrowing by `slot.len()` (a window length). The first
implementation resolved every `narrow` against the parent frame, which took
padding off twice for anything inside a padded row -- a `leftover(1)` slot
of a 900 px row inside `pad(16)` gave its child a frame of 418 where the
slot is 434. A slot *cannot* be written as a fraction of the row's frame:
that is a division of two lengths, the same argument that forced region
node entries to be translations. So `narrow` is a window length, `Pad` reads
its own frame with the new `Painter::frame_len` and takes the pixels off,
and nothing resolves a narrow a second time. Pinned by
`layout::a_share_inside_padding_fills_the_slot_it_was_given`.
- **The root resolved its own rule twice**, because `root_layout` passed the
declaration as both the narrow and the declaration: a root of `rel(0.5)` in
a 900 px window came out 225 wide. Pinned by
`layout::a_root_with_a_fraction_rule_is_that_fraction_of_the_window`.
- **Which box arm `in_parent` takes is decided by the box, not the frame.**
The first implementation treated any decided frame -- a narrow, a share, a
declaration -- as a box this widget chose, so a `Pad` (which narrows) lost
its child's box pin, and a rule changed over two zero pads relocated the
column under them instead of dividing it again (seed 59, depth 5,
`resize-size`). Only a declaration places the box inside the part it was
given; a narrowed frame leaves the box exactly the part. Pinned by
`unsettled::changing_a_rule_over_two_pads_divides_the_column_again`.
- **Frame validity is a pin, not a range.** A range of window pixels cannot
say "this answer is a fraction of *that* frame": two frames are different
lengths at the same window size. `LayoutHolds` therefore carries
`frame_len` beside `extent_len`, set where a widget reads its frame
(`frame_len`, which `Pad` does) and where a rule that is a fraction of the
frame is answered with it, and composed up where a length of this frame is
what reached the child. It replaced a `holds_for_frame` helper that
compared expressions and asked `Holds::ANY` whether anything depended on
them, which missed the rule-answered case entirely. **No test in the repo
distinguishes the rule-answered half**: 400 seeds over sixteen cases at
depth 5 agree with it and without it, and hand-built trees for the shape
the seed-30 records showed (a stale `1 rel` answer beside a stack's sizing
child) do not reproduce it either, because a frame that changes almost
always changes a box as well and the box pins catch it. Kept anyway,
because "these two invalidations always coincide" is the kind of unstated
assumption the three earlier plans died of; it is one line at each of two
sites if it is ever shown to be dead.
- **`Scroll` resolved its content length against its own box.** With one unit
its child's answer is a window length, so it becomes pixels against the
window; the two differ wherever a scroll's box is not its frame. Fixing it
then broke `resize`, because resolving a length against the window is a
*read of the window* and nothing recorded it: a viewport 40 px tall inside
a branch's box does not change when the window does, so an end-snapped
scroll kept the offset it had (seed 942, depth 6). `Painter::to_px` takes
the read where the resolution happens and pins the window only where the
length has a fraction in it; `window_px_len`, which returned the number
and recorded nothing, is gone. Pinned as
`unsettled::resizing_under_a_short_scroll_snaps_its_window_tall_content_again`.
- **A local redraw resolves its own frame** from its record's narrow and
declaration (`ask_again`, shared with `place_in`) rather than reusing the
retained frame. That deleted `asked_px` and the assert that the two agree:
they legitimately disagree when an ancestor's frame changed and the
ancestor was reused because nothing under it pinned the frame, and the
freshly resolved one is the right one.
- **The clipping debug assert compared the reported size with the frame**
rather than with the box it is a claim about.
Three rig changes came out of it, all kept:
- The shrinker prints the *shrunk* tree's divergence rather than the grown
tree's, and the divergence now lists, for the failing widget and every
ancestor, warm and cold `frame`, `ask`, `box` and `size`. Reading that
table is what found the box-arm defect; reconstructing a 20-widget plan by
hand would not have.
- The fuzzer's `Branch` states the range it branched on rather than pinning
the window it read the measurement against (`adbedaf`). It is the only
widget in the fixture that reads a length in pixels and chooses a subtree
from it, so with a pin every one of them redrew on every resize: seed 1's
resize was 131 widget draws and seed 13's 828, against 44 and nothing once
it says the range. That is a fixture that cannot tell a change that reuses
well from one that does not, and it looked exactly like a regression in
the change being measured.
- A `size-resize` case: a size change and *then* a resize. Every other case
changes something and compares at once or resizes first, so an answer kept
as a fraction of the wrong length agreed at the window it was made at and
parted from cold at every other one.
Checks at the end of the step (`a30971e`, and `23523ee` before the rig and
the renames): `cargo fmt --all --check`, clippy
with and without `layout-diagnostics`, 122 suite tests, 20 core, 11
generated, the 400-seed depth-5 shrinker over all sixteen cases, the
1000-seed depth-6 oracle, and the 2000-seed depth-4 scan over all sixteen
cases.
What the renders showed, at 1920x1200 against `e44dea3` (#18), with the
examples identical in both checkouts and Vulkan confirmed (the example asks
for `Backends::PRIMARY` and would have aborted on the GL fallback):
- `minimal` and `tabs` are byte-identical, and so is `tabs` after the
reference replay -- which changes 94% of the screen, so it is a real
comparison and not two pictures of nothing.
- `random` differs in two pixels, each by one grey level: glyph
antialiasing, nothing deleted.
- `text` differs only in its third block, the `rel(0.5)` padded paragraph in
a row: the block is the same width (945 px) and wraps the same lines, and
sits one pixel further right. Its box is 948 px, so the drawing is centred
with a 1.5 px remainder and the two expressions round that half-pixel the
other way. The first two blocks, which have no declared width, are
identical -- padding coming off the frame does not move a text that wraps
at the box it was given.
- A live resize of `random` to 900x1200 is byte-identical to a cold render
at 900x1200, which is the warm/cold rule checked in pixels rather than in
regions.
One thing is unexplained and worth knowing about: seed 30 at depth 5
(`shuffle-all-but-first`) failed once, with a span keeping an answer of
`1 rel` for an axis whose frame had become `78 px`, and then stopped failing
before the pin was added -- reverting each part of the change since that run
did not bring it back. Either an edit fixed it that I could not identify, or
the rig is not as deterministic as it looks. If a seed ever fails once and
not again, suspect the second.
## What the step 1 review found (worker, 2026-09-19)
The pre-submit review of the six commits `4328eac..a30971e` as one diff.
Build, lint and every fuzzer were already clean, and the review still found
one wrong layout.
- **A scroll placed content that fits into a box the length of the
window.** `Scroll` asked for its content box as
`Place::Fill(Part::From(content))` with `content` defaulting to
`UiSpan::FULL`, and only replaced it when the content is scrolled or
longer than the viewport. A `Part::From` span is in window lengths, so
`UiSpan::FULL` -- `rel 0` to `rel 1` -- is the whole window rather than the
whole box. The 300 px viewport of a 400 px window put its content in a
400 px box anchored at the viewport's start: a 50 px child centred at
275..325 where #18 has it at 225..275. It is the common case, not a corner
one -- any content shorter than its viewport takes it.
Nothing in the repository caught it. The warm/cold oracle cannot: both
sides are wrong the same way, and it compares warm with cold rather than
with what is right. The render set cannot either: every scroll in
`minimal`, `tabs`, `random`, `text` and `view` is either window-length on
its axis or has content longer than its viewport, so all five are
byte-identical before and after the fix. Only a hand-written expectation
found it, which is the argument for keeping some.
Fixed at `e8a5792` by saying the whole of a box as `Part::All`, the one
expression that cannot mean anything else -- and, being the place the
child was already asked in, one that makes the placement a no-op. Pinned
by `scroll::content_that_fits_is_placed_in_the_viewport_and_not_in_the_window`.
- **`Part::From`'s own documentation still called its spans frame lengths**,
which is what the scroll above read them as, and two comments in
`in_parent` still described the window ranges as frame ranges converted
through a frame. Corrected at `a888717`. When a unit changes, the comments
naming the old one are not decoration: this one cost a layout defect.
Checks re-run at `a888717`: `cargo fmt --all --check`, clippy with and
without `layout-diagnostics`, 123 suite tests, 20 core, 11 generated, the
400-seed depth-5 shrinker over all sixteen cases (73 s, 34,488 widgets), the
1000-seed depth-6 oracle (187 s), the 2000-seed depth-4 scan (347 s, 82,203
widgets), all five renders at 1920x1200, the `tabs` replay, and a live
resize of `random` to 900x1200 against a cold render there. The renderer was
confirmed as Venus on the host's RX 7900 XT rather than llvmpipe. Every
render is byte-identical to the same render at `a30971e`.
## What the one-ask protocol found (planner, 2026-09-18, third session)
Branch `wip/one-ask` at `3091fb8` over `4328eac`. The change is described in
`docs/HANDOFF.md`; this is what building and fuzzing it turned up, in order.
- **The step 3 assertion was the wrong check.** At the worker's stopping
point every generated case stopped on `assert_eq!(extent, info.part)` in
`place`, on a `Text` asked at 45 px in `Within(All)` whose longest word is
89.5 px: its drawing holds for `[45, 45]` and its answer box is 89.5 px
wide. No leaf can promise its drawing holds for a box chosen from its
answer; the plan's sentence "every drawing must hold for the answer box it
supplies" was the defect, not `Text`. Removing the check *and the redraw
it guarded* -- placement is re-expression, unconditionally -- made the
suite green including both decided-box pins, with nothing else changed.
- **Six retained tests encoded the paired draw** (`settled + 2`, "drawn
once" expecting two): they now expect one draw. Two more failed because
`Stack` had been changed to ask non-sizing children with `Fill(All)`; they
need `Within(All)` so a smaller answer sits inside the stack, with the
frame narrowed to the sizing answer. That is the whole of the suite churn.
- **Seed 2 at depth 4 (`repaint-some`, `region-node`)** shrinks to a `Stack`
sized by one text. A local redraw of the text asked it again in the box
the stack was *asked* in and left it filling that; cold puts it in the box
the stack's *answer* chose. A local redraw must put the fresh drawing
back at the retained place of the parent's answer box, always, not only
when the asked and placed places differ.
- **Six seeds at depth 5** (60 `repaint`, 308 `every-size`, 20
`region-node`, 248, 384, 162 `reorder`) shrink to one shape: a zero `Pad`
round a widget reading `extent_len` (a span with a share rect, or a
`Branch`), as a share child of a span beside a fixed sibling. Seed 248
reproduced: **cold** was wrong. In the measuring pass the inner span read
`far = rel 1 - 17.6px` and wrote slots in it; in the placing pass the pad
got a narrowed frame of the same pixel length, was reused (its contract
had no pin, because `in_parent`'s `Part::Of` arm drops a child's
`extent_len`), and its child was re-expressed into a frame where `rel 1`
means 17.6 px less. The inner span's own record had the pin; only the
composition lost it. Composing the pin through `Of` -- exactly where the
part is the whole box less pixels, `pinned - part.px`, and pinning the
parent's own length otherwise -- fixed all six; the planner's earlier
version of the same experiment broke seeds under the old protocol because
the second draw was still there.
- Cost after the fix (widget draws / distinct widgets, depth 8): seed 1
cold 331/288, many 78/78, size 3/3, scroll 1, resize 40/15; seed 13 cold
1179/982, many 335/333, resize nothing drawn. The table with the baselines
is in the handoff. `many` at seed 13 touches 333 distinct widgets because
a share child is asked twice per span draw, and `resize` at seed 1 draws
40 because positive-direction spans without shares pin `far`; both are
step 4 there.
- At `3091fb8` the 2000-seed depth-4 scan agreed over all fifteen cases
(82,203 widgets, 269 s) and the 1000-seed depth-6 oracle found **seed 946,
`shuffle-all-but-first`**, which shrinks to nine widgets: `Span{LEFT}[Rect,
Span{RIGHT}[Rect, Span{DOWN}[Wrapped, Rect, Wrapped].height(159), Rect]]`,
the column emptied to its first text. Reproduced with the records
printed: the column reports a share along the row while its rect has room
(both texts wrap at the whole row, 35 px each, under 159) and a fixed
width once asked in its third of the row (each text 88 px, no room, rect
undrawn, cross-axis answer a pixel width). So the row asked it twice and
built its slots on the first answer and its cross size on the second.
Emptying the column changes only the first; the local redraw re-asked the
second, found it unchanged, and told nobody. Cold has the column 874 px
wide. Fixed in `0ef87eb` by deferring any widget its parent asked twice
in one draw to that parent, pinned as
`unsettled::emptying_a_column_the_row_asked_twice_asks_the_row_again`
(checked red without the deferral, green with it). Cost: `many` at depth
8 went from 78/78 to 110/92 draws/distinct at seed 1 and from 335/333 to
424/364 at seed 13; still under #18's 157/95 and 524/159 draws. At
`0ef87eb` the 1000-seed depth-6 oracle (148 s) and the 2000-seed depth-4
scan (262 s) both agree over all fifteen cases.
## What the step 3 plan got wrong (planner, 2026-09-18, third session)
Three things, each a different kind of mistake:
1. **It kept the question the machinery could not answer.** The offer bit
existed because a widget was drawn twice, once where asked and once where
its answer placed it, and something had to say which draw was the
measurement. Both plans of 2026-09-18 tried to define that bit better
(the worker's four rules, the planner's two experiments). The repair is
to have one draw, after which the bit has nothing to name. A plan that
proposes bookkeeping for a distinction should first ask whether the
distinction has to exist.
2. **It stated the contract at the wrong box.** "Every drawing must hold for
the answer box it supplies" reads as a tightening and is actually
unsatisfiable for any widget that reads its box. The right statement is
the opposite: the answer box is never a question, so nothing is demanded
of the drawing there; `Holds` is about the ask box only.
3. **It narrowed frames by region.** The worker's share frames were
`extent_part(axis, From(span))`, a position in the span's frame; a frame
given as a position does not move when its part moves, and `re_ask` had
no way to put it back. A narrowed frame must be a *length* of the
parent's frame, placed into the part by alignment on every placement,
exactly as a declared rule already was.
The requirements did change under the plans (transparent frames, then
shares narrowing frames, then `Scroll`'s content length), which explains
some churn but not the three above; those were reasoning errors that a
five-widget reproduction with the retained records printed would have
caught in minutes. Print the records before proposing the next rule.
## The two shrinker seeds that were open on `wip/transparent-frames` (planner, 2026-09-18) -- closed by `wip/one-ask`
Both were the rule in `draw_inner` that decided which answer places a widget's
box: `measured = if info.offer() { fresh } else { retained }` with
`offer() := place == offer_place`. That bit is meant to say "this ask is a
measurement" and has no consistent value once a container's body runs in
more than one box.
- **Seed 2, `repaint`, depth 5** shrinks to `Stack{sized by child 0}[x:
leftover] > [OneLine, Pad{0} > Span{DOWN}[Rect, OneLine], Branch]`. The
stack's box is one text line tall. The span measured in the stack's room
has leftover room, draws the rect and reports `leftover` across; in the
stack's box it has none, undraws the rect and reports the text's width.
Cold kept the first because the stack's placing evaluation reused the
pad's retained answer, which looked valid only because `in_parent`'s
`Part::Of` arm drops the span's `extent_len` pin. Warm re-asked the span in
the box and got the second. The correct layout is the second: the stack's
non-sizing children belong in the stack's box, and the measuring pass drew
them in a box the stack never has.
- **Seed 108, `reorder`, depth 5** shrinks to `Span{LEFT} > [Span{LEFT} >
[Span{DOWN} > [Branch{probe Rect, wide Rect, narrow Wrapped, 483},
Wrapped], Rect], Rect]`. A draw trace shows the `DOWN` span evaluated at
900, 450, 600 and 300 px across in one cold layout (room of the outer
span, slot inside the inner span's measuring pass, slot of the inner span,
slot inside that), the wrapping text re-shaped at each. In the last one
the branch's narrow text is drawn in its real 300 px box and answers 286
px; `draw_inner` discards that for the retained 438.9 px answer from the
450 px evaluation because the branch's `below` place embeds
`extent_len(Y)`, which moved when the sibling text's height changed, so
`place != offer_place`; the text is redrawn 438.9 px wide in a 300 px box.
Warm does the same with a different stale answer (286 px from the first
frame). The same shape is `unsettled::a_widget_under_a_region_node_is_asked_in_the_box_that_node_was_offered`.
Two one-line experiments, both reverted, run at `49cec82` with the fast
oracle, the debug suite and the shrinker at 400 seeds of depth 5:
| change | seed 2 | seed 108 | else |
| --- | --- | --- | --- |
| compose the pin through `Part::Of` (`len.px - part_len.px` where `part_len.rel == ONE`) | fixed | fails | seed 220 `reorder` and one suite test fail |
| `measured = answer.0` always, and every ask records `answer`/`offer_part` | fails | fixed | seeds 184, 246, 292, 372 and two suite tests fail, all under `Scroll` |
The second experiment's failures are the case the gate was secretly holding
up: a container re-drawn in a box its own answer derived. The specific one
is `Scroll`, whose `apply_leftover` content length adds a whole viewport on
top of the content's pixels (a row reporting `600 px + leftover` in a 900 px
viewport gets a 1500 px content box), so the wrapping text inside is offered
900 px more room than it was measured in and re-wraps; the layout drifts one
iteration per evaluation, and warm and cold differ by how many they ran.
Bryan: that content length was always wrong; a share is the room left in the
viewport, so `content = max(viewport, px + rel*viewport)`. With it the
content box never re-offers room and `Scroll` needs no own-answer
evaluation.
The `many`/`resize` cost is the same mechanism in miniature: four
evaluations of a ten-widget subtree in one cold layout, from `Within(cursor..
far)` measuring rooms followed by `Fill(slot)` placing, with the span's
`extent_len` pin forcing a redraw at every box length.
## Original step 3 plan, preserved verbatim (planner, 2026-09-18)
This is the original idea from ai-app-2 `280fad7`, kept here even as the
handoff's stop notice changes with new evidence:
> ### 3. Evaluate children in parent-decided boxes
>
> Change placement so an answer-derived box never runs a container body.
> Placing becomes reuse-or-translate. Remove the offer/measurement gate and
> its retained bookkeeping only as each caller stops needing it; do not leave
> a parallel old path.
>
> Three current widgets must stop depending on measuring boxes they will never
> own:
>
> - `Span`: do not read `extent_len` unconditionally. Slots depend on `far`
> only when shares exist (the decided slot fills its part) or for negative
> direction; compute negative-direction slots from `total`. Pin the extent
> length only in those cases.
> - `Stack`: draw non-sizing children in `From(0..size)` on an axis where the
> sizing child's answer is `px`/`rel`, and `All` where it is `leftover`,
> instead of drawing them in `All` of the measuring room.
> - `Branch` in the random rig: express "the rest of my box" as
> `Of(40px..FULL)` rather than reading `extent_len(Y)`.
>
> Every drawing must hold for the answer box it supplies. The two focused tests
> from step 1 and the existing region-node and decided-box tests must pass here.
The surrounding seven-step implementation plan remains in `docs/HANDOFF.md`;
the quoted text above is the part whose interpretation is now blocked, so it
must remain available even if the live handoff is rewritten.
## Step 3 experiment after verifying the tests (worker, 2026-09-18)
The initial 17 suite failures mixed obsolete expectations with real defects.
`ReadsBox`, `ReadsWidth`, and `Measured` used pixel reads to compute only their
answer; failing `Counted` cases installed a box dependency without drawing
anything. Those fixtures encoded the old second draw inside the answer box.
Giving the painter answer-only pixel reads preserved their invalidation and
geometry claims while removing the obsolete paired-draw count. Separate
fixtures still exercise actual drawing dependencies with ordinary pixel
reads.
The experiment then reached a green 114-test suite. Its substantive findings:
- answer validity and final-drawing validity must remain independent all the
way through `draw_inner`; returning their old combined contract falsely
pinned wrappers such as `Pad` to provisional boxes;
- a repeated ask of one child replaces that child's provisional final-drawing
contract rather than intersecting both drawings forever;
- provisional `Span` asks fill their measuring room; `Within` immediately
tries to put the child in its own answer before the span decides anything;
- only an actually allocated share pins final span slots to `far`; an
overfull row with a requested but unallocated share uses fixed slots;
- a narrowed share frame is `Part::From(slot)` of the span's *extent*, not a
zero-based span in the forwarded frame;
- a local redraw re-asks the question that produced the retained answer and
marks the parent to restore any distinct final placement.
The release fast oracle then failed all eleven ordinary cases during cold
layout. Most failures were `Text`: its provisional drawing was pinned to a
pixel width that did not include the smaller box its answer selected. A
generated `Stack` showed the same class through a symbolic `FULL` pin. This
is stronger evidence than the original suite failures: the generated tests
are not asserting draw counts or the old two-draw protocol, and no warm state
exists yet. The open decision is whether the guarantee quoted above applies
to every leaf and how `Text` can establish it, or whether answer-derived leaf
evaluation remains legal while container evaluation does not.
## What the plan of 2026-09-18 got wrong (planner, same day)
The plan claimed the retained model rests on the frame's length being the
same on every ask. Answers depend on the *part* (the box offered), not the
frame, for every widget that reads its box, so transparent frames fixed
fraction resolution and moved the second geometry from the frame to the
extent rather than removing it. Step 6 ("redraw without the deferral; both
seeds must pass now because the frame no longer changes under the widget")
followed from that claim and was false for the same reason. The plan also
carried `offer_place` ("from the first ask of the parent's draw at the
offer") forward from the old protocol without noticing it is undefinable
once a container is evaluated in several boxes, and its `Inset` sketch
assumed a narrowed frame and a placed extent could coexist, which
`frame_and_extent` collapses ("a narrowed frame is its own extent") and which
the extent-in-frame-coordinates representation cannot express in general
(`(0.5*B - 20)/(B - 20)` is not `rel + px`). The plan also wrote "a frame is
narrowed only by a declared length, an inset or the root" and the code's
`declared_lens` filters `leftover` out; Bryan: a share was always meant to
narrow the frame like the other two.
The worker executed the plan as written, found exactly this ("whether a
given draw is a measurement cannot be recovered from the ask"), tried four
bookkeeping rules (first-ness alone, place matching the retained offer place,
box matching the measured box, a retained flag on the drawing), each fixing
some seeds and breaking others, restored the deferral and stopped. That was
the right call. The plan as handed over is at ai-app-2 `152bed7` if the
wording is ever needed.
## What the worker measured on `wip/transparent-frames` (worker, 2026-09-18)
Widget draws / distinct widgets / update ms from `tests/layout_diagnostics.rs`
at depth 8; draw counts are deterministic so these are single runs.
`e44dea3` is #18's head, `34cafb6` the commit the branch starts from,
`49cec82` its head.
| seed 1 | e44dea3 | 34cafb6 | 49cec82 | 3091fb8 | 0ef87eb |
| --- | --- | --- | --- | --- | --- |
| cold | 369/261/10.6 | 463/274/13.3 | 516/288/12.0 | 331/288/12.3 | 331/288/12.0 |
| repaint | 1 | 1 | 1 | 1 | 1 |
| many | 157/95/0.33 | 263/108/0.59 | 187/119/0.52 | 78/78/0.13 | 110/92/0.23 |
| size | 16/12/0.018 | 3/3 | 3/3/0.010 | 3/3/0.008 | 3/3/0.008 |
| scroll | 2/0.002 | 1 | 1/0.004 | 1/0.001 | 1/0.001 |
| resize | 13/13/0.019 | 22/15/0.032 | 24/76/0.090 | 40/15/0.096 | 40/15/0.096 |
| seed 13 | e44dea3 | 49cec82 | 3091fb8 | 0ef87eb |
| --- | --- | --- | --- | --- |
| cold | 1330/707/20.3 | 2940/982/28.3 | 1179/982/24.6 | 1179/982/24.6 |
| many | 524/159/1.09 | 1091/423/2.39 | 335/333/0.62 | 424/364/0.87 |
| resize | nothing drawn | 2215/510/6.56 | nothing drawn | nothing drawn |
The last two columns are single runs from the third session; `cold` and
`resize` at seed 1 are after the `Part::Of` pin composition, which added
18 and 19 draws respectively over the version without it, and `many` at
`0ef87eb` includes the deferral of twice-asked share children.
Three findings, each applied on the branch:
- **Lazy `Within` placement costs more than it saves.** Leaving a child's
answer to be placed at the end of the parent's draw puts the drawing in the
part first and in the answer's box after; where it does not hold for both
that is two drawings. Seed 1's resize went from 391 widget draws to 29
with it removed.
- **An inset said in frame lengths makes a container read its own box.**
"Less eleven pixels at the end" needs the length, and a container whose
box is its own answer then depends on its own answer: `Pad` drew
sixty-four times in one resize frame at seed 13. `Part::Of` says it as a
part of the box and composes without a length.
- **Pin one axis at a time.** `extent_len` pinning both axes made a span
dividing one hold for one length of the other.
`Part::Of`'s `in_parent` arm drops a child's `extent_len` pin (only `All`
composes one), which is what let seed 2's pad reuse an answer for a box its
span had never been measured in.
Reference renders at `49cec82` against `34cafb6`: `view` and `minimal`
byte-identical; `tabs` 2,332 pixels; `text` shows `wtext(..).width(rel(1.0))`
inside `.pad(16)` at the window's edges. Under the decided share rule that
example is correct as written: once the share narrows the frame, `rel(1.0)`
is the padded share and fits. The clipped render records the current bug.
`random` moves where nested spans do.
## Defects landed before transparent frames, and their lessons (2026-09-17 review)
- **A report is a fraction of the containing widget** (`ffd79f3`). A
report used to come back composed through the box it was offered, and a
span offers each child the room from its cursor, so a nested span taking
half of what it was given took a quarter of a row whose first half was
spoken for. Consequence: a span can overflow itself without bound, so its
`fixed <= 0` branches are ordinary and boxes of negative length reach
children; nothing yet states what a widget may assume about one.
- **An answer is not an answer while anything under it is dirty**
(`0e0d4af`, superseded by `a0693ac`). `dirty_size_under` was what made an
answer an answer until the settling walk made the state it guarded against
unreachable. Found at seed 564, depth 6, `shuffle-every-other`.
- **A frame settles strictly bottom-up** (`a92c6ac`, `a0693ac`). A widget
that cannot settle defers to its parent rather than drawing the parent
from inside itself; `update` marks the root for a resize instead of
drawing it top-down, and only where the new output falls outside what the
root's answer holds for. Sound by induction on depth. Bryan: "then that
entire category of issue can't even occur."
- **A text is handed back a box its own line fits in** (`4bd8607`): report
`ceil` of the shaped size; the two tolerances that were holding it
together (`BREAK_EPSILON_PX`, a nearest-step `Holds` start) both went.
- **A subtree that changes hands is recorded on both sides** (`e44dea3`):
the old parent's child list is repaired and the subtree's depths re-walked
where its top moved. The fuzzer never re-parents, which is why nothing
generated reached it.
- **A span's leftover boundary is its own inverse** (`53b00c6`): forty
lines became twelve and one `div` left layout.
- **`wip/stack-fraction-twice`** (a stack sized by a child reporting
`rel(0.5)` applied it twice) is closed by transparent frames:
`placed_extent` takes the reported length *from* the part rather than
composing it into the part. No oracle could see it; the branch's test pins
it. **`wip/padding-outset-and-inset`** is superseded: padding goes outside
what it pads, and its `rel(0.5)`-under-inset failure was the same second
application.
- Four findings from the frame/extent prototype still shape the code: an
answer and a drawing each retain their dependencies (`answer_under` and
`under`); a wider contract does not invalidate an existing guarantee; no
measurement is different from a measured zero (`ActiveData::answer` is
optional); the settling walk takes the deepest mark from a `BTreeSet`
keyed by depth and what ends it is the mark set.
## Failed hypotheses worth not repeating
- **The placement pin was blamed for the `many` gap and is not the cause**
(2026-09-17). Disabling it still redrew 487 distinct widgets a frame at
seed 13 against 159; the per-widget trace showed local redraws deferring
to their parents in chains to the root because a span handed its children
its own placement as their frame. Transparent frames fixed that part.
- **`wip/local-reask`** re-asked a dirty widget at its offer instead of
deferring, under the old protocol, and diverged at seeds 532 and 398 of
depth 6. Superseded; the branch can be deleted.
- **An offer composed back up the move chain** fell back to `FULL` under a
region node and was resolved against that node's *placed* box, so
everything under a `Scroll` was re-asked at the content's width. Pinned by
`unsettled::a_widget_under_a_region_node_is_asked_in_the_box_that_node_was_offered`.
- **Four bookkeeping rules for "which draw is a measurement"** (worker) and
**the two experiments above** (planner): each fixes some seeds and breaks
others. The bit is undefinable; `wip/one-ask` removes the second draw, so
there is nothing left for it to name.
- **Asserting that a placed drawing holds for its answer box** (worker's
step 3): unsatisfiable for `Text` and for any widget that reads its box
and reports something else. The answer box is not a question.
- **Choosing between a fixed and a relative child in pixels at the span's
current width** admits multiple self-sizing fixed points; seed 13 settled
differently warm and cold under it. The same circularity is what a cap
containing `leftover` would put into `SizeRule::Max`.
- **A tolerant endpoint on the span's leftover split** retained zero-height
children at seed 16. Pinned by
`unsettled::a_box_that_only_rounds_past_its_fixed_children_leaves_nothing_over`.
- **`Scroll` returning the final placed answer** rather than the first
box's advanced one fixed-point iteration (seed 86). Moot once the content
box never re-offers room, but keep the test until that is measured.
- **A `git bisect`** once named a commit that could not be the cause; read
the tree rather than the bisect when that happens.
## Fuzzer coverage
- Seeds 1121 and 1839 at depth 4 failed on every commit before `4bd8607`
and nothing routine reached them: the fast oracle takes ten seeds, the
shrinker 400 at depth 5, the long oracle 1000 at depth 6. The scan that
found them (2000 seeds at depth 4 over all fifteen cases, 261 s) should be
run after any layout change. `Rng::new` is `seed | 1`, so an even seed and
the odd one above it are one tree.
- The shrinker panics per thread at the first failing seed, so a run
reports at most one seed per chunk of about 58; a seed listed as new after
a change may have been hidden behind another in the same chunk. Check a
single seed against the unpatched code before calling it new (seed 220
was checked this way and is new under the pin experiment).
- Depth finds things and so does breadth; widen one axis at a time and
record which.
- The cases are what a change is *then* compared at, and until 2026-09-19
every one of them compared at the window the change was made in. A
`size-resize` case was added because a length kept as a fraction of the
wrong box agrees with cold at that window and parts from it at every other
one. There is still no case that changes a tree twice with a resize
between, and no case that resizes twice.
## Smaller open items (carried from the old handoff)
- An undrawn `leftover` child still contributes its gap, so a vanished
child leaves a double gap.
- Nested spans pass `leftover` weight up, so three leftover children in one
inner span beside one in another get three quarters to one quarter. No
other layout system does that; confirm it is wanted.
- `Fixed::div` by zero answers `MIN`/`MAX` while `ratio` answers `ZERO`;
both are caller bugs under `debug_assert`, but the fallbacks differ.
- The comment on the `local == UiRegion::FULL` shortcut in `widget_at` says
composing through `FULL` "is not quite the identity in f32"; on the grid
it is exact and the shortcut is performance only.
- `docs/LAYOUT.md` §4, §5 and the density section name `Painter::place`,
`SetSize`, `desired_width`, `apply_rest`, `Len::dp`, `Aligned` and
`MaxSize`, none of which exist. Do not restore `OnResize::Translate` or
`OrthoSize`.
- `tabs` changed twice across `d3b0ebf` with nobody looking; take the oracle
as the reference and the five renders as a spot check.