Record frame-extent performance gains and alignment verification

docs/HANDOFF.md

@@ -422,8 +422,9 @@ back.
 The original proposal was implemented by Claude as `5fcace1` on
 `wip/region-and-placement` in `/home/bob/repos/iris-pr18`. It kept the fraction
 reference stable but failed five layout cases and three draw-count cases.
-The first correction is `efb416b`; the exact-recomposition continuation is
-`2ed5503`, pushed to origin's `wip/region-and-placement`, in the isolated checkout
+The first correction is `efb416b`, exact recomposition is `2ed5503`, and the
+current performance/alignment continuation is **`7601aa2`**, pushed to origin's
+`wip/region-and-placement`, in the isolated checkout
 `/home/bob/repos/iris-layout-experiment`; **it is not on PR #18**. The original
 checkout is unchanged. Keep the experiment's build directory separate: Cargo
 can accept a different worktree's artifacts as fresh when a target directory
@@ -436,7 +437,7 @@ report must not replace the reference against which that report was obtained.
 Text wraps at the available extent, not the whole fraction reference.
 
 **The offer includes availability.** A span can keep the same reference frame
-while offering a text less room after an earlier sibling. `known_len` therefore
+while offering a text less room after an earlier sibling. `measure_len` therefore
 accepts the placement as well as the frame. The first measurement's
 `offer_placement` survives later placing evaluations, and local redraw asks
 that original question before restoring the assigned slot. Frame-length
@@ -484,52 +485,96 @@ at arbitrarily wider widths. New tests compare actual primitive and mask
 geometry and assert reuse for fixed-frame resizing and widening unwrapped
 text (including explicit newlines and empty content).
 
+**Measure the offer without placing an intermediate answer.** `Painter::measure_len`
+replaces the hint/cache query and its duplicated drawing fallback. It still runs
+`Widget::draw` when needed, but leaves that drawing in the offer until the parent
+assigns the child's slot. Placing the child's own answer first was wasted work:
+the parent immediately replaced that placement. The regression counts three
+rather than four evaluations for a numeric-size leaf in a span, checks its actual
+primitive bounds, and repeats after resize. There is no second layout body on
+`Widget`. `Holds::ANY.through(...)` now returns `ANY` directly: an unrestricted
+range needs no inverse division, including for a negative fraction.
+
+**A chosen slot does not cancel a child's alignment.** Bryan caught this in the
+`tabs` image: a fixed 100 px square under a flexible wrapper sat at the slot's
+near edge. `ask_box` now aligns a declared frame inside the chosen slot, just as
+it does inside an unassigned offer. A regression asserts the actual coordinates
+before and after resize. Warm/cold equality could not find this because both
+were wrong. The corrected 900 px image still has 88 px between blue and red and
+78 px between red and orange: the enclosing `.pad(10)` adds 10 px before the
+red slot. Red is centered in that slot, 5 px right of the visible black gap's
+center. Do not describe those two centers as the same thing or compensate with
+an arbitrary offset. The parked padding API change is still separate.
+
 Ordinary verification passes: formatting, workspace clippy with
-`layout-diagnostics`, workspace tests, and the fast generated cases. The
-suite contains 98 tests. A `tabs` resize from 1920x1200
-to 900x1200 matches a cold 900x1200 screenshot with zero differing pixels.
-With exact equality, the final release oracle passes 2000 seeds at depth 4
-and 1000 seeds at depth 6, the shrinker passes 400 seeds at depth 5, and the
-debug oracle passes 120 seeds at depth 4 with assertions enabled. Every scan
-runs all fifteen scenarios. The fixed-frame reuse regression was also run
-against `efb416b`: it fails there with two child draws instead of one.
+`layout-diagnostics`, workspace tests (100 suite tests and 21 core tests), and
+the fast generated cases. The corrected `tabs` resize from 1920x1200 to
+900x1200 matches a cold image with zero differing pixels. The release oracle
+passes 2000 seeds at depth 4 and 1000 at depth 6; the shrinker passes 400 at
+depth 5, and the debug oracle passes 120 at depth 4 with assertions enabled.
+Every scan runs all fifteen scenarios with exact equality. The measurement
+regression fails with four draws instead of three when the optimization is
+disabled.
 
-**Do not land this as a performance improvement.** The old stretching and
-wrapper reuse tests pass, and one numeric-size test improves from three draws
-to two. The continuation reduces widget draws from `efb416b` by 6.9% cold,
-9.0% with many dirty widgets, 7.6% for a size change, and 8.7% for resize.
-Against the last working `e44dea3`, it still does substantially more nested
-layout work (seed 1, depth 8):
+**Still not ready to replace PR #18.** The prototype does substantially more
+nested layout work than `e44dea3`, even after removing intermediate placements
+(seed 1, depth 8):
 
-| phase | widget draws before | prototype | primitive writes before | prototype |
+| phase | widget draws at e44dea3 | current | primitive writes at e44dea3 | current |
 | --- | ---: | ---: | ---: | ---: |
-| cold | 369 | 498 | 9635 | 9704 |
+| cold | 369 | 484 | 9635 | 9179 |
 | repaint | 1 | 1 | 1 | 1 |
-| many | 157 | 282 | 4721 | 6398 |
-| size | 16 | 254 | 106 | 5725 |
+| many | 157 | 274 | 4721 | 5873 |
+| size | 16 | 247 | 106 | 5200 |
 | scroll | 2 | 2 | 0 | 0 |
-| resize | 13 | 300 | 0 | 6255 |
+| resize | 13 | 292 | 0 | 5730 |
 
-These are instrumented work counts, not elapsed-time speedups (50 frames per
-phase except cold, which is one; 578 generated widgets, 191 finally active).
-Uninstrumented release binaries, run directly with `perf stat -r 3 -e
-instructions:u`, also improve over `efb416b`: 2000 size-change frames fall
-from 14.3923B to 13.8914B instructions (3.5%); 2000 resize frames fall from
-16.9756B to 16.1331B (5.0%). These whole-process totals include fixture setup;
-reported run-to-run variation is at most 0.24%. The many-dirty workload falls
-from 16.7962B to 16.0030B instructions (4.7%, 2000 frames). Repaint and scroll
-keep their draw counts but incur small CPU increases: 0.8849B to 0.8913B
-(+0.7%) and 2.4332B to 2.4479B (+0.6%) over 100,000 frames each. These are
-CPU-work measurements, not phone frame times. Keeping measurement validity separate from final drawing validity
-and deferring to an already dirty parent were also tried; neither improved
-this rig's draw counts, so those changes were removed.
+These are instrumented work counts, not speedups (50 frames per phase except
+cold, which is one; 578 generated widgets, 191 finally active). Against
+`2ed5503`, the changes remove 7 widget draws and 525 primitive writes per size
+frame, and 8 draws and 525 writes per resize frame.
 
-Raw placement reads in containers and repeated offer/placement evaluations
-remain conservative. Retaining local coordinates also costs memory per
-primitive. The experiment demonstrates selective reuse with one `Widget::draw`
-body, not a finished performance replacement. Optimize the container protocol
-under the same correctness tests before adopting it. The parked `Inset`/`Outset`
-API changes are not included in this experiment.
+**Measure cycles as well as instructions** (Bryan, 2026-09-17). The following
+are whole-process medians from seven alternating before/after runs of direct,
+uninstrumented release test executables under `perf stat -e
+cycles:u,instructions:u`. The baseline is `39f7b08`: `2ed5503` with the same
+centering fix, so both sides draw the corrected layout. CPU-heavy scans were stopped
+during measurement. Counts include fixture setup; size, resize and many run
+5000 frames, repaint and scroll run 1,000,000. These are VM CPU measurements,
+not phone frame times.
+
+| phase | cycles before → after (billions) | instructions before → after (billions) |
+| --- | ---: | ---: |
+| size | 12.8008 → 11.0344 (-13.8%) | 34.3660 → 29.9450 (-12.9%) |
+| resize | 14.9368 → 13.0376 (-12.7%) | 39.8487 → 35.0032 (-12.2%) |
+| many | 14.8875 → 12.9579 (-13.0%) | 39.6021 → 34.7499 (-12.3%) |
+| repaint | 2.7416 → 2.6626 (-2.9%) | 6.0266 → 6.0197 (-0.1%) |
+| scroll | 9.7848 → 8.4829 (-13.3%) | 21.7203 → 20.9527 (-3.5%) |
+
+The size/resize/many cycle samples have nonoverlapping before/after ranges,
+with each range less than 1.3% of its median. Additional depth-6 fixtures,
+10,000 frames and seven alternating pairs each: seed 3 improves cycles by
+7.0% for size and 8.4% for resize; seed 13 improves resize by 9.2%, while its
+size samples overlap and establish no cycle improvement.
+
+**The remaining gap has not been established as an unavoidable correctness
+cost.** A temporary failure trace of one frame on seed 1, depth 8 counted
+39/108 failed reuse attempts in size, 65/151 in resize, and 45/109 in many
+where only the raw placement snapshot differed: both numeric frame and extent
+ranges still held. Containers read that snapshot to construct child positions,
+so discarding the check would be wrong; a way to retain those child positions
+relative to the extent is the next protocol question. There is also necessary
+width-dependent work: the hot text at widget 199 visits 68 px and 89.53613 px
+and reports different heights (316.80078 and 228.80078). It is evaluated
+repeatedly at those widths, so that fact alone does not justify all the repeats.
+
+Keeping measurement validity separate from final drawing validity and deferring
+to an already dirty parent were previously tried without improving draw counts.
+Measuring first in local redraw and then restoring its slot saved only one draw
+in `many` and added another request on simple repaints; that trial was removed.
+Retaining local coordinates still costs memory per primitive. Optimize under
+the exact oracle before adopting the prototype; the `Inset`/`Outset` changes
+remain parked.
 
 ## How layout is decided