@woosh/meep-engine 3.42.0 → 3.44.0

package/src/engine/physics/cloth/MEASUREMENTS.md

@@ -4114,3 +4114,798 @@ clip keying the cape carries it as its rest.
   is two disjoint ranges of one row, which the partial upload shader spec already
   lands under the emulator; and the worker path writes nothing after a release,
   because a join for a cloth unlinked or reseeded since its dispatch is dropped.
+
+## 62. Attachment points: holds instead of an anchor
+
+M7 took the fixed set off the `Cloth` entity and gave it to `ClothPin`s (PLAN §19). Three
+commits before this one carry the parts that change no behaviour on their own: the rows from a
+particle graph, the solver's kinematic targets and sourced leash rows, and a rename. Every
+number below was taken on this machine with 23 other node processes running.
+
+### The design was reviewed before it was built
+
+Three adversarial reviewers read the first draft against the code — the solver and the
+pathology budget, the ECS and the worker, the topology and the output. Their findings changed
+the design in these places, each folded into PLAN §19:
+
+| Finding | Change | Measured or argued |
+|---|---|---|
+| Up to four hard tethers per particle, one per island (Chaos's rule), go infeasible when two holders separate past the fabric between them, and the idle slots cost a sqrt each per iteration | One tether, to the nearest held particle | argued; the cost is below |
+| A leash centre written once a step sits a step behind a held particle that moves inside the step | `leash_source` + offset, re-centred every substep | measured: `cloth_held.spec.js` |
+| `x += w·(T − x)` with `w = (i+1)/n` is not equal shares for `n ≥ 3` | `x += (T − x)/(n − i)` | measured: same spec, 1–4 substeps |
+| `slack = 0` times an unreachable particle's `d_geo = ∞` is NaN | radius `∞` where `d_geo` is; leash loops test `radius === ∞` first | measured: a zero-floor leash row went NaN on the previous solver |
+| A cloth a holder drags has no velocity, so contact discovery and the gather do not reach for what it is dragged into | `KINEMATIC_TRAVEL` in both reaches | measured: `cloth_held.spec.js` |
+| A hold taken mid-run switches on tethers the fabric already exceeds | tethers start at the larger of `d_geo` and where the particle is, and only come in | measured: catching a falling rope, no particle faster than it was falling |
+| Removing the frame particle costs every strand its first free link | the `Cloth` entity stays particle 0, held by a pin on itself | measured: equivalence below |
+| `Cloth` on a bone: transient, not found by `instance_of`, a GPU parent under the subtree path | `Cloth`/`ClothRig` stay on the model root; the rig's frame is a proxy joint | argued from the code |
+| A saved rope reloads without its pin and falls | `ClothPinSerializationAdapter` | measured: round trip |
+| Self-collision is blind to a held particle, and `closest_free_pair` goes stale on release | held–free pairs; the pair measured over every particle | measured: two specs rewritten red-to-green |
+
+Kept against the review: the subtree's walk order as the particle numbering (a single chain is
+numbered by depth whatever the ids; a branching subtree's siblings ascend by entity id), and a
+hold beginning where its particle is.
+
+### What it costs
+
+`bench_holds.mjs` (scratchpad): the previous build and this one in one process, each building the
+same rig through its own `ClothSystem`, the rope tops pinned in the new arm, a walk with a turn
+in it driving the frame every step, 400 warm-up frames then 2,000 measured, alternating which
+arm goes first.
+
+| scene | min | p10 | median |
+|---|---|---|---|
+| 8 strips × 16, silk (129 particles) | 1.028 | 1.012 | 1.014 |
+| 32 strips × 32, rope (1,025) | 1.017 | 1.023 | 1.028 |
+
+Before the per-step hold pass learned to skip when no holder had moved in the frame, the same
+rows read 1.046 / 1.046 / 1.032 and 1.055 / 1.052 / 1.062. What is left is the substep loop that
+re-centres sourced leash rows and the two leash rows the frame's own particle now carries.
+
+### It is the same cloth
+
+`equiv_holds.mjs`: both arms stepped together, the worst particle-to-particle distance printed.
+A nudged copy of the previous build against itself (`chaos_control.mjs`, one particle moved
+1e-7 m) is the control.
+
+| step | M7 against the anchor | the anchor against itself, nudged |
+|---|---|---|
+| 1 | 1.9e-7 m | 1.3e-7 m |
+| 5 | 2.3e-6 | 4.6e-7 |
+| 10 | 1.3e-5 | 1.8e-5 |
+| 30 | 1.8e-2 | 1.6e-2 |
+| 60 | 1.9e-1 | 2.1e-1 |
+
+The difference starts at rounding — stage one's row order and a geodesic summed in double — and
+grows exactly as fast as a nudge does in a rig driven this hard. Chaos, not a change.
+
+### One spec read the row order
+
+Building the rows per kind rather than per node changed the order a particle's rows are summed
+in. `cloth_collision.spec.js`'s replaced-collider test then saw the particle it picks slide
+14.7 mm in the step its contact is rediscovered, from a rope at rest (0.23 mm/s) — it had held
+the fresh friction anchor to half a millimetre. The defect it guards puts the anchor 700 mm off,
+so the tolerance is 50 mm. With the old row order restored the spec passed, so the slide is the
+order, not the graph.
+
+### Gates
+
+Cloth suite 327 (from 297), slow 7, the whole physics suite 1,861, extended physics including the
+allocation budgets, `check-types` 18 against a budget of 19. The worker parity gate gained a run
+in which a second holder drags a rope's tip, lets go at step 80 and takes hold again at 160:
+bit-identical, every step, for 240.
+
+---
+
+## 63. Sheets: a proxy from triangles, and a flag that turns on its pole
+
+The third stage of M7 (PLAN §19.5). `cloth_proxy_from_mesh` builds a proxy from a skin with one
+joint per vertex and the triangles over them, the rows take its hinges, and a joint with
+triangles round it is turned by its surface. The numbers below are from `cloth_sheet.spec.js`'s
+own scenes, with a probe printing what its assertions bound.
+
+### Every library fabric comes to rest on a sheet
+
+A 6 × 6 sheet, 0.1 m apart, lying flat and held along one edge by a pin on a rail entity. It
+swings down about the held edge, which gravity restores.
+
+| fabric | asleep at step | far edge dropped, of 0.5 m |
+|---|---|---|
+| silk | 568 | 0.50 |
+| cotton | 73 | 0.48 |
+| denim | 76 | 0.37 |
+| leather | 72 | 0.29 |
+| rope | 512 | 0.50 |
+| chain | 280 | 0.50 |
+| hair | 66 | 0.46 |
+
+### A flag on a pole does not, and it is the solver's
+
+Stand the same sheet up and hold it along its hoist — a vertical column — on a pole. With three
+centimetres of curve across it, so that it buckles out of its plane rather than folding inside it
+(§7 P4), rope and chain never sleep inside 4,000 steps; silk, cotton, denim, leather and hair
+sleep at 1,261, 751, 572, 226 and 976.
+
+It is not jitter. The flag turns about the pole as a rigid body, and its angular momentum about
+the pole, `Σ (r × v)·ŷ` per unit mass, holds to four figures from step 300 to step 3,000: −0.1894
+at rope's settings, about ten degrees a second. A turn about a vertical hold line is the one
+motion nothing restores — gravity is parallel to the axis, and every row either ends on it or is
+an equal and opposite pair — so the true answer is a flag that stays where it stops, and a
+steady turn is something the step puts there.
+
+| rope on the flag, one change at a time | `Ly` at step 3,000 |
+|---|---|
+| as shipped: 2 × 4, `damping` 0.1 | −0.1894 |
+| `damping` 1 | −0.0203 |
+| no tethers (every radius ∞) | −0.1836 |
+| no tethers and no strain-limit row | −0.1757 |
+| no hinges (`bend` 0) | −0.2226 |
+| no curve | 0 — it folds in its own plane instead |
+| the curve mirrored | +0.1894 |
+
+Ten times the damping rate is a tenth of the speed: a constant torque, balanced rather than
+removed. None of the rows M7 added carries it. On the solver directly — a `ClothState` of the
+flag's hard stretch rows and nothing else, the hoist at inverse mass zero, damping applied by
+hand between steps:
+
+| iterations | held along a vertical column | held along the top row |
+|---|---|---|
+| 2 | −0.0843 | 0 |
+| 4 | −0.1998 | 0 |
+| 8 | −0.3499 | 0 |
+
+The solver as it stood before holds existed (`3654c0949`) prints the same numbers to every
+digit. Through `ClothSystem` the torque goes to zero with the iteration count, but not
+monotonically — 0 at 1, −0.0039 at 2, −0.1894 at 4, −0.3981 at 8, −0.0670 at 16, +0.0562 at 32,
++0.0020 at 128 — so this is AVBD stopped short of convergence, and the step's fixed point at a
+fabric's budget is a steady turn rather than rest. Two repairs inside the step were tried, and
+neither is one:
+
+| sweep order, rope | 2 iterations | 4 | 8 |
+|---|---|---|---|
+| colours forward (shipped) | −0.0039 | −0.1894 | −0.3981 |
+| direction alternating each sweep | −0.0792 | +0.2403 | +0.0669 |
+
+| warm start `α`, rope at 2 × 4 | `Ly` | peak speed, m/s |
+|---|---|---|
+| 0.25 (shipped) | −0.1894 | 0.065 |
+| 0.6 | −0.1354 | 0.044 |
+| 0.9 | +0.4459 | 2.18 |
+| 1 | −88.1 | 119 — it blew up |
+
+The fabrics that sleep do so because a slack row catches the turn, not because it stopped. At
+step 3,000 the fly end of a silk flag rests 123° round the pole from where it was hung, cotton's
+74.5°, denim's 8.7°. Recorded in PLAN §18 rather than fixed: the levers are §5.2's, and every
+library entry's measurements rest on them.
+
+### Joint normals are exact; the twist is not measured
+
+A 15 × 15 cotton sheet held by nothing falls half a metre onto a 0.4 m sphere and drapes; its
+middle rests 7.6 mm off the surface, cotton's thickness being 8 mm. Every interior joint's turn
+takes the rest normal onto its surface's area-weighted normal to within a hundredth of a degree.
+That is the construction — the swing is normal onto normal — so the spec now bounds it at half a
+degree, and what it guards is the plumbing: the winding's sign, the frame, the write-back. The
+twist about the normal comes from one fixed neighbour and nothing measures it. It matters to
+tangent space — a normal map, an anisotropic sheen — and not to a vertex normal.
+
+### What a sheet costs to write back
+
+One joint per vertex is one main-thread write per vertex per step, and under `WorkerClothSystem`
+it is the part that stays on the main thread. Cotton held at its hoist and moving; the whole step
+and `cloth_write_back_rig` alone timed in alternating blocks of ten, the median of ten blocks
+each, two runs:
+
+| sheet | joints | whole step | write-back | a joint |
+|---|---|---|---|---|
+| 11 × 11 | 121 | 0.85 ms | 0.075 ms | 620 ns |
+| 21 × 21 | 441 | 3.17 | 0.27 | 600 |
+| 41 × 41 | 1,681 | 12.6–13.0 | 1.09–1.12 | 650–670 |
+
+Linear, and about 8.5 % of the main-thread step at every size. At the library's 400-particle
+garment budget that is 0.25 ms of main thread every fixed step, against 0.18 ms for the main
+thread's whole share of the §17 workload (§41). Where the 620 ns goes — the reconstruction,
+composing the local pose, the change event — is not broken down.
+
+### Gates
+
+Cloth suite 338 (from 327): `cloth_proxy_from_mesh.spec.js` 7, `cloth_sheet.spec.js` 4. The whole
+physics suite with its slow, bench and extended specs (`--mode all`): 185 files and 1,891 tests,
+green. `check-types` 18 against a budget of 19.
+
+---
+
+## 64. Three more sheets, and a fold resting on a fold that never comes to rest
+
+The three sheet scenes M7's third stage was approved with and did not land: a draped sheet on a
+sphere that turns, a pile of ribbons dropped on a floor, and a circle skirt on a body that turns.
+The sphere and the skirt are specs in `cloth_sheet.slow.spec.js` — seconds of simulation each, so
+they run with the extended set — and their numbers are below. The ribbons
+found a defect in §9's self-collision row instead: a self-collision pair that carries weight adds
+energy to the step. PLAN §18 records it. Every number here comes from `ClothSystem` scenes driven
+the way the specs drive them, read back from the joints' world transforms.
+
+### A sheet on a turning sphere is carried round by friction
+
+§63's scene: 15 × 15 cotton at 0.1 m, nothing holding it, draped for 600 steps on a 0.4 m sphere
+with friction 0.5. Then the sphere is turned about its vertical axis at a quarter turn a second for
+four seconds, one whole turn. The "turn" is the sheet's mean change of azimuth, summed a step at a
+time, over the particles more than 10 cm from the axis.
+
+| | measured |
+|---|---|
+| the sheet's turn when the sphere stops, at 360° | 337.4° |
+| two seconds later | 359.8° |
+| closest any particle comes to the surface, at any step | 0.57 mm outside it |
+| the middle's height above the sphere's top | 7.5 mm, as before the turn |
+
+The turn reaches the cloth through the friction anchor. `cloth_contact_find` stores a contact's
+anchor in the collider's own frame, so it turns with the collider. The sheet lags by what friction
+lets slip while it is spun up, then catches up. A frictionless sphere is no control: the sheet slides
+off it while it drapes, and is 25 m down by step 600.
+
+### A circle skirt on a turning body flares, and settles
+
+The skirt is cut flat: an annulus 32 particles round and 7 out, 6 cm apart from a 15 cm waist, so
+the hem hangs 36 cm below the waistband. It is cotton, and its waistband is pinned by a `ClothPin`
+on a waist entity at 1 m. The body is a vertical capsule, radius 0.18 and cylinder 0.5, centred at
+0.6 m, so the hips are wider than the waist and the skirt drapes over them. The body and the waist
+are turned together once a second for two seconds, then stopped.
+
+| phase | hem radius | hem height | hem off the vertical |
+|---|---|---|---|
+| draped, asleep at step 164 | 0.293 m | 0.673 m | 24° |
+| the second second of the turn, mean | 0.402 | 0.791 | 50° |
+| the turn's peak | 0.488 | 0.881 | |
+| ten seconds after it stops | 0.278 | 0.673 | |
+
+After the stop, no particle moves more than 0.60 mm over the last five seconds. The hem's height
+comes back to within 0.7 mm, and its radius to 14 mm inside where it was, because its folds fall
+differently. Nothing goes inside the hips at any step: the closest is 4.6 mm off the surface,
+against cotton's 8 mm.
+
+It settles and does not sleep. Before the turn the folds did not touch, and it slept at step 164.
+After the turn the search finds eight self-collision pairs every step. The cloth holds a jitter of a few
+tenths of a millimetre per 300 steps, at up to 0.037 m/s, and never passes the sleep test in 4,000
+steps. That is the defect below.
+
+### A fold resting on a fold never comes to rest
+
+The ribbons are silk strips 16 × 3 at 3 cm, all in one cloth. The self-collision radius is 15 mm,
+so the diameter is one spacing. They lie on a floor plane with friction 0.5. The lowest ribbon starts
+5 mm up, each next one a gap above it, turned by an angle about the pile's middle. A run lasts 2,400
+steps or until the cloth sleeps. Where a gap is 3.2 cm, §37's ceiling takes the radius down to
+12.8 mm, because the bind pose has the ribbons that close. The ratio was varied in a probe build whose
+`cloth_step` read it from the probe. "Crossings" is §31's metric, as §33 counts it: edges passing
+through triangles they share no vertex with, over the whole cloth, read at the end of the run.
+"Peak" is the fastest particle at any step. Free fall from the top ribbon's height bounds what the
+drop alone can give it.
+
+```
+  scene                        free fall   CLOTH_SELF_RATIO: asleep at step / peak m/s / crossings at the end
+
+                                                  1                 2                 4                 8
+
+  2 parallel,  gap 3.2 cm       0.85 m/s    143 / 0.54 /  0    156 / 0.63 /  0    353 / 0.77 /  0     -  / 1.51 /   0
+  2 parallel,  gap 8 cm         1.29        174 / 0.93 /  0    183 / 0.93 /  2     -  / 0.93 /  0     -  / 3.00 /  14
+  2 at 90°,    gap 3.2 cm       0.85        101 / 0.63 /  0    107 / 0.64 /  0     -  / 0.65 /  0     -  / 1.55 /   0
+  2 at 90°,    gap 8 cm         1.29        198 / 1.08 /  0     83 / 1.08 /  0     -  / 1.08 /  0    159 / 1.08 /   0
+  3 at 90°,    gap 3.2 cm       1.16        211 / 0.79 /  2    296 / 0.79 /  5   1165 / 0.79 /  0     -  / 4.40 /  14
+  3 at 90°,    gap 8 cm         1.80        253 / 1.50 /  0     -  / 1.50 /  8     -  / 2.24 /  0     -  / 2.94 /  19
+  3 at 30°,    gap 3.2 cm       1.16        205 / 0.99 /  0    199 / 0.95 /  0    248 / 0.94 /  0     -  / 1.72 /  11
+  3 at 30°,    gap 8 cm         1.80        257 / 1.81 /  0    224 / 1.69 /  4    237 / 1.71 /  0     -  / 2.33 /  17
+  6 at 30°,    gap 3.2 cm       1.80        304 / 1.63 / 20    398 / 1.66 / 55    264 / 1.72 / 20     -  / 4.64 / 139
+  6 at 30°,    gap 6 cm         2.45        367 / 2.79 /  0     -  / 3.04 / 38   1664 / 4.30 / 84     -  / 5.74 /  44
+
+  "-" is "still awake at step 2,400". The shipped ratio is 8.
+```
+
+**At the shipped ratio no pile comes to rest, and a pile of three or more gains energy.** Every
+three- and six-ribbon pile ends awake. Each moves 1.3 to 3.8 times faster than its drop could make
+it, and ends with 11 to 139 crossings. Nine three-ribbon piles were tried at 8, fanned 30°, 60° and
+90° apart from gaps of 3.2, 5 and 8 cm, and all nine cross and none sleeps. Two ribbons do not
+sleep either. Two crossed at 90° were dropped from gaps of 4 to 12 cm in 1 cm steps, and only the 6
+and 8 cm drops sleep, at steps 127 and 159. The others creep 3 to 56 mm every five seconds for as
+long as they run, and the 10 cm drop ends with 8 crossings. One ribbon alone sleeps at step 32.
+Two ribbons with the flag clear pass through each other and sleep at steps 41 and 42.
+
+**At ratio 1 every scene in the table sleeps**, by step 367 at the latest, and the draped and
+twirled scenes below sleep too. Where the energy comes from was located by elimination rather than
+derived. Between the columns only the self row's stiffness changes. Between the rows what changes
+is how many self pairs carry weight. In a stack a pair stays in contact, so it is a spring that is
+on every step. That is the case M0's dissipative band was measured on, and §32's sweep did not
+contain it: §32 drove a folding sheet whose contacts come and go, and its "no energy growth" holds
+there. Averaging each particle's share of the Jacobi scratch over its pairs, at ratio 8, was tried
+and is not the fix. Two ribbons crossed at 90° then sleep, at steps 77 and 78, but the parallel
+pair and every three- and six-ribbon pile still never sleep.
+
+**The same defect reaches the scenes above.** The draped sheet of §63 never sleeps on its sphere at
+ratio 8, though nothing turns it. The search finds 14 to 18 pairs every step where its hanging
+edges fold onto themselves, and over 4,000 steps those edges wander up to 50 mm in half a second.
+At ratio 1 it sleeps at step 386. The skirt sleeps at ratio 1 too, 267 steps after its twirl stops.
+
+**"No crossings at rest" does not mean nothing passed through.** A ribbon can go right through
+another while landing, then lie under it with nothing crossing. At ratio 1, four of the nine
+three-ribbon piles end asleep in the wrong order. The 30° fan from 8 cm is one of them: asleep at
+step 257 with no crossings at the end, and its top ribbon on the floor beneath the other two. Only
+the 60° fan, from 3.2 or 5 cm, is clean at every step at all three of ratios 1, 2 and 4. From 5 cm it keeps its order and sleeps at
+steps 189, 165 and 212; at 8 its first crossing comes at step 452 and it never sleeps. That pile,
+asserting no crossing at any step and the order kept, is the proof a change to the ratio needs. It
+was kept out of this change because it is red on the shipped one.
+
+**Ratio 1 does not stop a pile landing through itself.** Six ribbons 30° apart still cross while
+they land: 109 crossings in the worst step from a 3.2 cm gap, and 20 left at the end. That residue
+is §35's: the pair set is found from where the previous step left the particles, and the search
+carries no travel term.
+
+**What ratio 1 costs is §32's own table.** On the folding sheet the mean is 3.42 crossings a
+sample against 2.03 at 8, and the worst is 51 against 71. Not changed here: the ratio is §32's
+measured choice for silk, cotton and hair, and changing it wants §32, §33 and the library bench
+measured again, at each candidate ratio, against these standing contacts.
+
+### Gates
+
+Cloth suite 338, unchanged; the extended cloth specs 9 (from 7): `cloth_sheet.slow.spec.js` 2, whose
+`grid()` also builds an annulus. The whole physics suite with its slow, bench and extended specs
+(`--mode all`): 186 files and 1,893 tests, green. `check-types` 18 against a budget of 19.
+
+---
+
+## 65. The self-collision row ships at 2, so a fold resting on a fold comes to rest
+
+§64 found that at `CLOTH_SELF_RATIO = 8` a self-collision pair that carries weight adds energy to the
+step, and that at 1 everything it tried sleeps. Ratio 1 costs §32's folding sheet, which is what 8
+was chosen on, so this measures the ratios between them against both. The ratio was fed to
+`cloth_step` through a probe that was never committed. Everything below is today's code, and §32's
+own numbers were re-measured rather than quoted.
+
+### The folding sheet, at 48 gusts
+
+§32's sheet and §33's library bends, through the same `fold()` as `cloth_m4_measure.mjs`. Twelve
+realisations were not enough: at twelve, silk read 1.44 crossings a sample at ratio 6 and 2.64 at 8,
+and no column was in order. So each row here is 48 realisations: §32's twelve integer gust phases
+and the quarter phases between them, which is the same family of gusts sampled more finely. The
+figure is the mean crossings a sample, ± its standard error.
+
+```
+  ratio   sheet (bend 1e-7)   SILK 0.04     HAIR 0.20     COTTON 0.25     mean of 4
+
+      0     9.50 ± 0.59        9.28 ± 0.63   8.96 ± 0.57   8.95 ± 0.51      9.17
+      1     3.62 ± 0.40        3.80 ± 0.34   4.13 ± 0.37   3.41 ± 0.30      3.74
+    1.5     3.44 ± 0.33        3.27 ± 0.32   3.44 ± 0.44   2.80 ± 0.31      3.24
+      2     2.77 ± 0.27        3.43 ± 0.34   3.05 ± 0.35   2.67 ± 0.33      2.98
+      3     2.65 ± 0.30        2.56 ± 0.24   2.65 ± 0.29   2.89 ± 0.29      2.69
+      4     2.89 ± 0.28        2.31 ± 0.29   2.13 ± 0.24   2.16 ± 0.24      2.37
+      6     1.90 ± 0.23        2.40 ± 0.32   2.10 ± 0.24   2.23 ± 0.24      2.16
+      8     1.98 ± 0.22        2.52 ± 0.28   2.01 ± 0.29   1.84 ± 0.22      2.09
+```
+
+At 48 the trend §32 chose 8 on is there: a stiffer row passes the sheet through itself less while it
+moves. The row removes 73% to 79% of the crossings at 8, 63% to 71% at 2, and 54% to 62% at 1. The
+worst realisation lies between 62 and 134 crossings at every ratio, with no trend. At rest the
+sheet is clean at every ratio, and its kinetic energy at the end of the run moves between 4.5e-3
+and 6.4e-3 J with no trend either. This sheet still cannot see the defect.
+
+The row's code has moved since M4. At twelve realisations, with no row, it reads exactly §32's
+11.39 / 146 and §33's 8.08, 10.62 and 9.36. With the row it reads 1.84 / 63 at 8 and 3.27 / 67 at 1,
+against §32's 2.03 / 71 and 3.42 / 51.
+
+### Piles of ribbons, 75 per fabric
+
+§64's ribbons: strips 16 × 3 at 3 cm, all in one cloth through `ClothSystem`, on a floor with
+friction 0.5, run for 2,400 steps or until the cloth sleeps. Each pile is 2, 3 or 4 ribbons fanned
+0°, 30°, 45°, 60° or 90° apart and dropped from gaps of 3.2, 4, 5, 6 or 8 cm. That is 75 piles, for
+silk, cotton and hair, the three entries that set the flag. "Clean" means no crossing at any step.
+"Swapped" means two ribbons whose middles still lie over each other ended in the wrong order: one
+landed right through the other.
+
+```
+  fabric  ratio   asleep   clean: asleep / of   landed through: asleep / of   swapped
+
+  SILK      1     73          32 / 32                  41 / 43                  27
+  SILK   1.25     72          39 / 39                  33 / 36                  23
+  SILK    1.5     69          33 / 34                  36 / 41                  20
+  SILK      2     68          38 / 38                  30 / 37                  16
+  SILK      3     66          42 / 43                  24 / 32                  13
+  SILK      4     50          33 / 38                  17 / 37                   7
+  SILK      8      3           2 / 16                   1 / 59                   1
+  COTTON    1     74          35 / 36                  39 / 39                  24
+  COTTON  1.5     67          41 / 44                  26 / 31                  15
+  COTTON    2     74          45 / 45                  29 / 30                  13
+  COTTON    3     70          49 / 49                  21 / 26                  10
+  HAIR      1     75          47 / 47                  28 / 28                   8
+  HAIR    1.5     73          50 / 52                  23 / 23                   7
+  HAIR      2     74          51 / 52                  23 / 23                   7
+  HAIR      3     68          56 / 58                  12 / 17                   5
+```
+
+**A clean pile comes to rest from 1 to 3 and stops doing so above.** Over the three fabrics, the
+clean piles left awake are 1 of 115 at ratio 1, 6 of 130 at 1.5, 1 of 135 at 2 and 3 of 150 at 3. At
+4, five of 38 clean silk piles never sleep, all of them two-ribbon piles. At 8, 14 of 16 never
+sleep. Clean piles at 1 to 3 move no faster than their drop can make them, at most 1.07 times their
+free-fall speed, except one hair pile at ratio 1 at 1.28. Clean piles at 8 reach 1.82 times it, and
+§64's piles up to 3.8. At 1.5 the six awake
+are all four-ribbon stacks with 78 to 133 standing pairs, still moving at 0.15 to 0.32 m/s. Whether
+one particular pile sleeps does not follow the ratio. The four-ribbon 45° fan from 4 cm lands clean
+and sleeps at 1.25, 2 and 3, and never at 1.5. The fraction over 75 piles is the measurement.
+
+**At 3 and below, every crossing starts while the pile is landing.** All 406 of them start by step
+119, two seconds in, and 95% by step 64. That is §35's missing travel term, not the energy defect.
+At 8, ten of the 59 piles that cross do so first after step 100, as late as step 1,862. A softer row lets more piles land through themselves (110 of 225 at 1,
+90 at 2, 75 at 3) and swaps more of them (59, 36, 28). In exchange it lets more of those tangled
+piles come to rest anyway (108 of 110 at 1, 82 of 90 at 2, 57 of 75 at 3).
+
+### The draped sheet, the skirt and the turning sphere
+
+§63's sheet, 15 × 15 cotton draped on a 0.4 m sphere, run up to 4,000 steps. The step it sleeps at:
+
+```
+  ratio   1     1.25   1.5   1.75   2     2.25   2.5   3     4     5      6     8
+  asleep  386   299    309   357    299   300    723   298   326   never  462   never
+```
+
+Its middle rests 7.5 mm above the sphere's top at every ratio, and its corners hang 0.58 to 0.78 m
+below the middle. At 8 the search finds 17 pairs a step and the hanging edges
+wander 116 mm. At 5 it finds 12.
+
+The circle skirt from §64 twirls identically at every ratio: the hem 0.109 m further out and
+0.118 m higher over the second second. After the stop it sleeps within 600 steps at 1, 1.25, 1.5,
+1.75, 2, 2.25, 2.5, 3 and 4, and never at 5, 6 or 8, where the search finds six or seven pairs a
+step. At 2 it sleeps 270 steps after the stop, its hem 0.9 mm from its old height and 13 mm in
+from its old radius.
+
+The sheet a turning sphere carries round (§64) is carried at every ratio, to between 335° and 343°
+of the sphere's 360° when the sphere stops. Two seconds later it stops at a different angle at
+each ratio, by how far short of a whole turn it ends:
+
+```
+  ratio      1    1.25   1.5   1.75   2     2.25   2.5    3     3.5    4     5     6     8
+  short °   0.6   0.1    2.4   0.3    8.4   1.3   -2.8   9.1   0.8   -0.6  -0.1  -0.6   0.3
+```
+
+At 2 the whole sheet sits behind the sphere: the cap by 8.6°, the sides by 8.9° and the hanging
+edges by 8.0°. That is a net slip, not the folds rearranging. Nine spins at each of 1, 2, 3 and 8 —
+0.2, 0.25 and 0.3 turns a second, for 3, 4 and 5 seconds — share one 600-step drape per ratio, and
+they agree with each other: 2.8° to 8.4° short at 2, 1.4° to 9.1° at 3, within 1° at 1 and 8. So where the sheet stops
+is set by which folds the drape formed, and that changes with the ratio in no order. The spec
+asserted 2° from 359.8°, one realisation at 8. It now asserts what holds at every ratio: the sheet
+turns further after the stop than before it, and ends within 10° of the whole turn.
+
+### What the step costs
+
+`cloth_library_bench.mjs`'s sheets for the three entries, at their own budgets, driven by its
+folding gust. Each pair of arms is stepped alternately in one process, the order alternating every
+frame, and the figure is each arm's fastest step over three rounds of 1,500.
+
+```
+  entry    8 against 8   8 against 2   8 against 1     pairs a step at 8 / 2 / 1
+
+  SILK        1.026         0.995         0.987          372 / 452 / 534
+  COTTON      1.003         1.016         1.022          365 / 426 / 542
+  HAIR        0.982         0.984         1.013          366 / 439 / 533
+```
+
+Every ratio is inside the control's 3%: the cost does not move. The search finds 17% to 22% more
+pairs a step at 2, but §34 says the cost is in the search, not the pairs.
+
+### A pair on its own does not do it
+
+Two particles with no rows between them, pressed together by equal and opposite accelerations, so
+the self-collision row is all that holds them apart, come to rest at every ratio from 0.5 to 8. So
+does a column of three. The energy needs the cloth's own rows around the pair, and this section does
+not claim a mechanism for it. §64's "averaging a particle's Jacobi share over its pairs is not the
+fix" stands.
+
+### Why 2
+
+It is the stiffest ratio at which cloth resting on itself reliably comes to rest. The clean piles
+sleep, 134 of 135, and so do the draped sheet, at step 299, the skirt after its twirl, and §64's
+proof pile, at step 166 with no crossing at any step. Above 2.25 the draped sheet's sleep turns
+erratic, and at 4 clean two-ribbon piles stop sleeping. Below 2 the folding sheet only gets worse:
+3.24 crossings a sample at 1.5 and 3.74 at 1, against 2.98. Between 2 and 3 the folding sheet
+cannot tell the two apart, 2.98 against 2.69 with a combined error of about 0.2. Across all three
+fabrics, 216 of 225 piles sleep at 2 against 204 at 3.
+
+What it costs is §32's sheet: 2.98 crossings a sample against 2.09 at 8, and the row removes 63% to
+71% of the unaided crossings rather than 73% to 79%. Ratio 1 would be the answer if a pile that
+lands through itself must come to rest anyway. It sleeps 222 of 225 piles, but the folding sheet
+gets 25% worse than at 2 and 59 piles swap against 36.
+
+### Gates
+
+Cloth suite 338, unchanged: `cloth_sheet.spec.js`'s draped sheet now asserts it sleeps. The
+extended cloth specs 10 (from 9): `cloth_sheet.slow.spec.js` gains the ribbon pile, and its skirt
+asserts it sleeps after the twirl. All three are red on 8, where none of them sleeps. The whole
+physics suite with its slow, bench and extended specs (`--mode all`): 186 files and 1,894 tests,
+green. `check-types` 18 against a budget of 19.
+
+---
+
+## 66. A flag stops turning on its pole: the turn about a hold line is a block of the descent
+
+§63's flag turned about its pole for as long as it was run. A side session measured why, at
+`d9a51a8ce`, and built the repair as a probe; this section is the repair as landed. The first two
+subsections are the side session's measurements at that commit. Between it and this one's parent the
+self-collision row went from 8 to 2 (§65), so everything from "What landed" on was measured again:
+"before" is the parent, `0e20799be`, on this machine, with about 30 other node processes running.
+
+### Why it turned
+
+These are the side session's measurements, on a bare `ClothState` flag: 9 × 9 at 0.1 m, curved by
+`0.0005·x²`, hard stretch rows only, the hoist column held, 2 substeps, rope's damping applied between
+steps.
+
+- **Only the vertex updates change the angular momentum about the pole.** The predict, the dual
+  updates and the integrate change it by exactly zero: gravity is parallel to the axis, and the holds
+  sit on it.
+- **Over a substep it changes by `h·Σ x_prev × R`**, `R` the step energy's gradient left when the
+  iterations stop. At 2 × 4 the identity predicts −2.6036e-3 a substep and −2.6036e-3 is measured.
+  Implicit Euler's own term is +1.2e-5. The moment sits on the middle colours: colour 1 +1.49e-2,
+  colour 2 −1.78e-2, the last ≈ 0.
+- **It never stops** because the warm start keeps `α·γ = 0.25` of each hard multiplier, so every
+  substep re-solves three quarters of the hanging weight and a cloth at rest is never a fixed point.
+  Carrying a quarter reads −0.1588; carrying half, −0.0874.
+- **It is not sweep order.** At 8 iterations:
+
+| sweep order | `L_y` |
+|---|---|
+| forward (shipped) | −0.4656 |
+| backward | −0.4629 |
+| alternating each step | −0.4640 |
+| alternating each substep | −0.5013 |
+| forward then backward each iteration, twice the cost | −0.4034 |
+
+### What was tried first, and dropped
+
+AVBD's `α` is two mechanisms. Eq. 18 regularises the hard constraint, `C*(x) = C(x) − α·C(xᵗ)`; Eq. 19
+is the warm start, `λ ← αγλ`. §5.2 has only Eq. 19. The reference 2D code carries `λ` whole and adds one
+`α = 0` pass that moves positions and not velocity, which is post-stabilisation. All three were built in
+the probe and measured on the 345 cloth gates of the time:
+
+| variant | stops the turn | gates failed | and |
+|---|---|---|---|
+| Eq. 18 on the stretch table, `α` 0.98 | yes: every flag sleeps, rope at 459 | 6 of 345 | silk on the bench's sheet blows up at step 722 |
+| Eq. 18 on every hard row | only at `α` ≥ 0.98; rope −0.0153 at 0.9 | not run | cotton and hair lying sheets never sleep at `α` ≥ 0.8 |
+| post-stabilisation | after 1,300–2,200 steps of creep | 10 of 345 | diverges in the same scene |
+
+The six Eq. 18 failures: P6's TeleportKeep shift 0.24 against `< 0.2`, P7's locomotion at the velocity
+clamp, the write-back lag 0.18 against `< 0.15`, a rig offset of 0.1013 against `0.1 ± 0.0005`, and two
+collision specs. The divergence is where Eq. 18's rows meet rows still on the old warm start: it
+corrects only `1 − α` of an old error a substep, which makes a hard row soft against a transient.
+
+### What landed
+
+`cloth_hold_line.js`, called from `cloth_step`. PLAN §5.2 has it in the listing.
+
+- **The block.** In the last iteration of each substep, after the sweep and before the dual update: one
+  Newton step of the step's energy along the rotation about the line through the held particles,
+  `θ = −g/H`, then every free particle turned by `θ`.
+- **What enters.** `g = Σ (a × r_i)·∇E_i` and `H = Σ tᵀH_i t + max(0, Σ ∇E_i·(a(a·r_i) − r_i))`, with
+  `∇E_i` and `H_i` the vertex update's own assembly of the inertial term, the leash rows, the backstop
+  rows and the contacts. The internal rows are invariant under the turn and are left out.
+- **When it runs.** When the held particles lie on one line at that substep, at two distinct points or
+  more. The predict lists the held particles as it carries them, so the test is a pass over the held
+  particles. A particle is on the line within `2⁻¹⁶` of the largest held coordinate.
+
+What the probe had that this does not: a scan of every particle for the held ones each substep, a
+fixed 1e-4 m collinearity tolerance, and the rotation pass run when `θ` is zero.
+
+### The bare flag
+
+`cloth_hold_line.spec.js`'s scene, read at step 3,000.
+
+| iterations | `L_y` before | `L_y` now | worst stretch row's `C/L₀`, both |
+|---|---|---|---|
+| 1 | 0.0002 | 0.0000 | 6.0e-2 |
+| 2 | 0.0519 | 0.0000 | 4.5e-2 |
+| 4 | −0.1588 | 0.0000 | 3.4e-2 |
+| 8 | −0.4656 | 0.0000 | 2.4e-2 |
+| 16 | 0.3783 | 0.0000 | 1.7e-2 |
+
+Before, `L_y` is the same to four figures at steps 1,000, 2,000 and 3,000: a steady turn. The same
+sheet held along its top row, a line gravity restores, reads the same `L_y`, worst `C/L₀` and fly-end
+position to every digit printed in both builds. Run in every iteration rather than the last, the
+block reads 0.0000 to 0.0001 — no better.
+
+**The tolerance.** Turn the same flag 0.7 rad about `(1, 2, 3)`, turn gravity with it, and move it
+`d` metres from the origin. It is the same flag, but its hoist no longer lies along an axis, so its
+`Float32` coordinates carry rounding off the line. At 4 iterations, over 1,000 steps:
+
+| `d`, m | worst held particle off the line, m | over the largest held coordinate | `L_y` before | `L_y` now |
+|---|---|---|---|---|
+| 0 | 7.8e-9 | 1.2e-8 | −0.1586 | −0.0003 |
+| 1 | 5.8e-8 | 4.2e-8 | −0.1587 | 0.0006 |
+| 10 | 6.5e-7 | 6.3e-8 | −0.1558 | 0.0064 |
+| 100 | 3.7e-6 | 3.6e-8 | −0.2266 | −0.0478 |
+| 1,000 | 3.7e-5 | 3.7e-8 | −0.5235 | −0.5690 |
+| 4,000 | 1.5e-4 | 3.7e-8 | 0.1921 | −0.0723 |
+
+Relative to its own magnitude the rounding never passes 6.3e-8, half a unit in the last place, and
+`2⁻¹⁶` is 240 times that. In metres it grows with `d`, and the probe's fixed 1e-4 m refuses the line at
+4 km. Through `ClothSystem`, the hoists of the 6 × 6 flags below were sampled at every step of every
+run, driven and turned poles included: none was off its line by more than 3e-16 of its size. The cloth
+steps in its own frame, and those hoists were authored along an axis of it.
+
+**Far from the origin the turn comes back.** At 10 m the turn holds at 0.004 to 0.007 through 3,000
+steps. At 100 m it grows: −0.025 at step 500, −0.048, −0.065, −0.074, −0.093 and −0.136 at 3,000,
+against −0.18 to −0.23 before. The block turns the flag 5.6e-6 rad a substep at the origin and 7.2e-6
+at 100 m, the median over the last 500 steps. That is 2.8 to 3.6 µm of travel at half a metre. A unit
+in the last place is 0.06 µm at the origin, 0.95 µm at 10 m and 7.6 µm at 100 m, so at 100 m the
+correction is smaller than the rounding of the coordinates it moves. PLAN §18 keeps this open.
+
+### Flags through `ClothSystem`
+
+§63's sheet: 6 × 6 at 0.1 m, three centimetres of curve, held by a `ClothPin` on a pole. "Turn" is the
+fly end's azimuth about the pole at step 3,000, from where it was hung.
+
+| fabric | asleep at, before | `L_y` at 3,000, before | turn, before | asleep at, now | `L_y` at 3,000, now | turn, now |
+|---|---|---|---|---|---|---|
+| silk | 618 | −0.0038 | −24.8° | 507 | −0.0004 | 30.7° |
+| cotton | never | 0.0117 | −57.0° | 336 | −0.0005 | 29.5° |
+| denim | 480 | 0.0021 | 21.7° | 246 | 0.0000 | 29.8° |
+| leather | 227 | 0.0000 | −0.6° | 216 | 0.0000 | −0.6° |
+| rope | never | 0.1809 | 103.9° | 416 | 0.0004 | −15.6° |
+| chain | never | 0.0406 | 94.6° | 250 | 0.0001 | 0.2° |
+| hair | 799 | 0.0022 | 59.4° | 527 | 0.0003 | 33.8° |
+
+Denim, leather, rope and chain are the side session's numbers to every digit. Silk, cotton and hair
+self-collide, and at the row's old stiffness of 8 the probe read 1,164, 2,685 and 742 before and 593,
+597 and 446 after. Asleep, the top of the fly end has come down 0.11 to 0.22 m; leather's has come
+down 3 mm, because a leather flag stands out from its pole like a board. `cloth_sheet.spec.js` holds
+the rest claim for every fabric.
+
+`cloth_sheet.spec.js`'s first flag, 9 × 9 with `0.0005·x²` of curve, asleep at: silk 1,613 → 649,
+cotton never → 377, denim 472 → 307, leather 195 → 195, rope never → 431, chain never → 268, hair never
+→ 303.
+
+### What else it moves
+
+**Held along a line gravity restores.** §63's lying sheet, held along one edge, asleep at:
+
+| fabric | before | now |
+|---|---|---|
+| silk | 568 | 598 |
+| cotton | 73 | 68 |
+| denim | 76 | 68 |
+| leather | 72 | **165** |
+| rope | 512 | 507 |
+| chain | 280 | 271 |
+| hair | 66 | 65 |
+
+The far edge drops as far in both. The flag on an entity turned 30° about x and then 40° about y, so
+that its hoist leans 30° from gravity, sleeps at: rope 474 → 574, silk 574 → 571, chain 264 → 312.
+The regressions are leather's lying sheet, more than twice as long, the leaning rope and chain, a fifth
+longer, and silk's lying sheet, 5% longer. Why is not measured.
+
+**A driven pole.** The 6 × 6 flag, the pole swayed 0.3 m at 1.5 Hz, bobbed, turned about its own
+vertical at half a turn a second, or teleported a kilometre. Each runs to step 900, then stops; the
+teleport is at step 600. "Peak" is the fastest particle while driven; "asleep" counts from step 900.
+
+| fabric | motion | peak, before | asleep, before | peak, now | asleep, now |
+|---|---|---|---|---|---|
+| rope | sway | 7.40 m/s | never | 6.45 | 1,397 |
+| rope | bob | 7.89 | never | 6.71 | 1,383 |
+| rope | spin | 0.94 | never | 0.94 | 929 |
+| rope | teleport | 0.94 | never | 0.94 | never |
+| silk | sway | 6.47 | 1,903 | 6.30 | 1,542 |
+| silk | bob | 6.52 | 2,141 | 8.18 | 1,498 |
+| silk | spin | 2.10 | 2,218 | 2.04 | 1,488 |
+| silk | teleport | 1.21 | never | 1.21 | never |
+| leather | sway | 4.48 | 1,172 | 4.16 | 1,189 |
+| leather | bob | 4.61 | 1,196 | 5.11 | 1,163 |
+| leather | spin | 2.78 | 1,005 | 2.63 | 1,003 |
+| leather | teleport | 0.24 | 901 | 0.23 | 901 |
+| hair | sway | 8.89 | 1,269 | 5.29 | 1,235 |
+| hair | bob | 7.14 | 1,194 | 6.51 | 1,198 |
+| hair | spin | 2.50 | 1,684 | 2.48 | 1,162 |
+| hair | teleport | 0.92 | never | 0.89 | never |
+| chain | sway | 6.06 | never | 6.75 | 1,205 |
+| chain | bob | 6.20 | never | 5.92 | 1,168 |
+| chain | spin | 0.75 | never | 0.75 | 929 |
+| chain | teleport | 0.75 | never | 0.75 | never |
+
+Every run is finite. The peaks move both ways, by as much as hair's sway, 8.89 → 5.29 m/s, and silk's
+bob, 6.52 → 8.18. After the teleport neither build sleeps within 2,400 steps, except leather; the cause
+is not measured.
+
+**Wind.** The flag in a steady 6 m/s breeze across it: the fly end's azimuth about the pole at steps
+600, 1,200 and 1,800 is the same in both builds to 0.4° — silk 94.7° → 95.1°, cotton 77.2°, rope
+100.6° → 100.8°, hair 70.2°. Cotton sleeps at step 1,333, where before it did not within 1,800.
+
+**One held point, or a line the cloth hangs on.** §53's sleep table and the P5 12-link swing sleep at
+the same step in both builds, every cell: a rig held at one point has no line, and the block never
+runs. So does a sheet hung from one corner, and its `L_y` is the same: silk awake at 3,000, `L_y` at
+most 0.004. An 8-link rope held at its top two particles hangs and swings in a plane through its own
+hold line, which has no moment about it, so the block takes no step. It sleeps at the same step in
+both builds, leaning 2% or not.
+
+### Two hands, and the one scene the block keeps awake
+
+The 6 × 6 flag held by two hands at its two top corners — two held points, so always a line — either
+still from the start, or after the hands see-saw 0.15 m at 1.2 Hz for 900 steps and stop. Asleep at:
+
+| fabric | still, before | still, now | after the see-saw, before | after the see-saw, now |
+|---|---|---|---|---|
+| silk | 328 | 395 | 1,504 | 1,547 |
+| cotton | 234 | 242 | 1,378 | **never** — 1.71 mm per 30 steps at 8,000 |
+| denim | 141 | 142 | 1,076 | 1,081 |
+| leather | 91 | 84 | 1,034 | 1,031 |
+| rope | 286 | 268 | never — 4.62 mm | never — 2.06 mm |
+| chain | 175 | 199 | 1,122 | 1,102 |
+| hair | 185 | 183 | 1,152 | 1,198 |
+
+Cotton after the see-saw is the block's doing, and not the history's. With a copy of the block that can
+be switched at step 901, cotton sleeps at 1,378 with it off throughout, at 1,409 switched off at 901,
+and never while it is on after 901, whether or not it was on before. Its turn there is about 7e-7 rad a
+substep, median −6.8e-7 over steps 3,000 to 6,000. The corner line is horizontal, so gravity restores
+the turn; the block removes what is left of it rigidly each substep, and the sweeps put it back near a
+corner, unevenly. Rope never sleeps after the see-saw in either build, and silk and hair sleep at
+about the same step whichever way the block is switched.
+
+### The adversarial review
+
+Before this landed, a review drove the side session's probe of the block, at `d9a51a8ce` and at the
+self-collision row's old stiffness, against the cases a rigid turn could get wrong. Its numbers for
+silk, cotton and hair predate §65. What it found:
+
+- **Every block step reduced the gradient along the turn**: none of more than 50,000 calls left it
+  larger.
+- **Wind.** Across, along and against the wind, the fly end's mean direction was within 0.3° of the
+  probe's shipped arm and its flutter as large. Into a headwind silk swung round to trail at −171°,
+  where shipped stalled side-on at −94°.
+- **Contacts.** A flag pressed onto a sphere by a 4 or 8 m/s wind penetrated 0.38–1.45 mm at worst,
+  against 0.50–1.88 mm.
+- **A leaning pole, 1° to 60°.** Finite at every lean, and asleep at every lean but the chain's at 1°,
+  still awake at 2,400 steps: a slow decaying swing towards hanging downhill, where shipped had come to
+  rest 13° short of it.
+- **A curtain** held along a straight top, in wind: the same mean deflection to 0.1 mm, and the same
+  flutter. **Two hands** moving apart and together, or one twirling: finite, and asleep as shipped is.
+- **Placement.** Before the last sweep instead of after it, the flag still turns: rope −0.162, chain
+  −0.040. In every iteration, the same open items as in the last, at 13–17% of a step.
+- **Considered and not done:** running the block only on hold lines near the vertical would take away
+  both open items. It needs a tolerance constant and a test against gravity.
+
+The two-corner see-saw, a cross wind, a pole leaning 30° and a rope on its own hold line are measured
+again on the parent elsewhere in this section.
+
+### What it costs
+
+`ab_cost.mjs` (scratchpad), from the side session's harness: the parent and this tree in one process,
+the library's claims, 600 warm-up frames, then 1,500 interleaved frames × 3 rounds with the order
+alternating per frame. The control is the parent against a second checkout of the parent. "Sheet" is the
+bench's sheet, top row held, which never leaves its plane: the turn is exactly zero, the builds stay
+bit-identical, and what is priced is the line test and the `g`, `H` pass. "Flag" is the same sheet held
+along its left column with a centimetre of curve and the fabric's damping applied between steps, as
+`ClothSystem` applies it. Its turn is not zero, so the rotation pass runs and the builds part.
+
+| entry | budget | control, min / median | sheet, min / median | flag, min / median |
+|---|---|---|---|---|
+| silk | 2 × 4 | 1.004 / 1.001 | 1.029 / 1.031 | 1.032 / 1.044 |
+| cotton | 2 × 4 | 1.006 / 1.000 | 1.035 / 1.031 | 1.036 / 1.044 |
+| denim | 2 × 2 | 1.008 / 0.999 | 1.046 / 1.050 | 1.069 / 1.061 |
+| leather | 2 × 2 | 1.000 / 1.001 | 1.050 / 1.051 | 1.059 / 1.061 |
+| rope | 2 × 4 | 1.000 / 1.002 | 1.003 / 1.002 | 1.005 / 1.007 |
+| chain | 2 × 6 | 1.007 / 1.002 | 0.997 / 0.997 | 0.998 / 1.008 |
+| hair | 2 × 4 | 0.995 / 1.003 | 1.036 / 1.031 | 1.037 / 1.038 |
+
+The control on the flag read 0.995 to 1.006. A first run of the same harness the same day, against
+`5e75c1654` — whose step differs from the parent's only in the self-collision row, which neither scene
+has — read 1.034–1.077 at the minimum on the sheet and 1.042–1.073 on the flag, with its controls at
+0.975–1.016. **3–9% on a sheet held along a line, then, and 3–7% in this run.** It is dearest at 2
+iterations, where one block is a larger share of the step. Rope and chain are held at one point and pay
+for the predict's list and nothing else. The flag without damping is no measure: the parent's flag spins
+up, and a flag spinning taut costs more to step than one at rest, so the block's arm read 0.63 to 0.83
+of the parent's.
+
+Run in every iteration, the block costs 1.110–1.133 at the minimum on the sheet and 1.086–1.157 on the
+flag; the first run read 1.099–1.132 and 1.126–1.183.
+
+### Gates
+
+Cloth suite 342 (from 339 at the parent): `cloth_hold_line.spec.js` 1, the rest claim for every
+fabric's flag in `cloth_sheet.spec.js`, and a flag on a swaying pole in `WorkerClothSystem.spec.js`'s
+parity gate. The first two are red on the parent: the bare flag turns at 0.0519 at 2 iterations, and
+cotton's, rope's and chain's flags never sleep. The parity run is the first in the gate whose holds
+are more than two points on a line; until it, only the run in which a hand holds the rope's tip
+reached the block. A mutant that changes `θ` by a part in a thousand in one arm and not the other
+fails those two runs and no other. The extended cloth specs 10, unchanged. The whole physics suite
+with its slow, bench and extended specs (`--mode all`): 187 files and 1,898 tests, green,
+`cloth_pathology` among them. `check-types` 18 against a budget of 19.