# Sweep36 constructive supported sweep

Status: scoped native construction, full capture and complete visual review passed.
The candidate remains unadopted. Publication/browser evidence is recorded separately
in PUBLICATION.md; this report describes the versioned native and visual package.

Sweep36 removes the height-weighted warp that distorted Sweep34's lower abdomen
and thighs. The original Body shape now moves rigidly. Each original thigh
construction receives its own moved hip and fixed ankle before the existing
bounded hip composition. The two feet remain planted. This is a separate
editable candidate; accepted originals and prior candidates remain preserved.

## What changed

The original profile owners determine the moving surfaces. Body, shoulders,
local analytic arms, neck attachment and tail share one rigid transform. The
original180 upper ellipsoid and198 lower taper reconstruct each thigh, ending
at the currentG paw attachment. No per-frame mesh correction follows the union.
The mathematical guarantees are narrower and more useful than a blanket visual
PASS: exact Body isometry, positive upper affine/ellipse construction, unchanged
transverse scales and positive monotone concave scalar taper. The actual lower
attachment and composed groin are not globally convex even at rest.

The whole action was revised within those shapes. Three6/10/14degree turn cases
were paired with coherent forward travel, lowering and pitch. All75pilot views
were inspected; the middle10degree supported case retained a readable sweep
without the larger hip demand of14degrees. Its single `Body contribution`0–1
control drives the combination. The six helper amplitudes are driven outputs.
This is a coordinated performance choice, not a pixel-only fit of the original
front video or an arbitrary skin warp. The broom proportions and local arm
articulation are retained from Sweep34; their world motion follows the new Body.

## Evidence completed before full capture

- The independent source-profile model reproduces the actual neutral thighs
  within5.48e-7scene units. Three native whole-action pilots match the independent
  moving construction within2.43e-6. Neutral composition remains within4e-6.
- The public coupled family was sampled at17contributions and489times, giving
  16,626part evaluations. Upper segment ratios span0.967384–1.029728; the normalized
  lower centerline acceleration gauge never exceeds its neutral0.566782value.
- Nineteen final native time/contribution probes and two reopened probes pass;
  maximum native-versus-math error is3.021e-6. At these detailed samples, all11,463
  concealed cap vertices on each side are enclosed, with zero cap/paw triangle
  crossings. Source meshes are closed and manifold, with no nonadjacent triangle
  intersections in the recorded topology sample. These are sampled guarantees.
- Three public-control edits return exactly to the default. Saving and reopening
  with scripts disabled preserves the rest geometry; edit/reset and the four-second
  endpoint also match exactly. Five helper-binding parity times preserve the chosen
  pilot's geometry exactly. The complete native has an exact durable Mac backup.
- The selected render snapshot matches the original full-data renderer at three
  poses and three views with exactly identical decoded RGB values. It preserves
  evaluated geometry and required shader attributes, freezes evaluated shader
  driver values on disposable copies and leaves the original native untouched.

The earlier field audit measured local stretches up to2.1485under the discarded
warp. Those infinitesimal field stretches differ from an entire constructive
segment's length ratio; this report does not equate the two measurements or
reduce visual acceptance to either number. The original Body distance preservation
and actual moving contours are the relevant new evidence.

## Existing arm embedding

The original club arms embed their medial surfaces in the Body, including some
material beyond0.5of the arm length. A blanket no-arm/body-intersection assertion
therefore rejected the neutral source itself. A read-only seven-event audit and
an independent convex-hull calculation confirmed the geometry. This was not
resolved by hiding the measurement or labelling the whole region a shoulder.

The current check compares every sampled distal material vertex with the same
local arm articulation against the original neutral composition. Body-core depth
must remain rigidly identical within2e-5. Composition depth may increase by at most
0.0007, one tenth of the original0.007remesh grid. The final-native probes showed
at most0.000214increase. Absolute depths and triangle-pair counts remain visible
in the evidence. This is a non-regression check on an inherited interface, not a
zero-collision or arbitrary-control certificate. Broom surfaces are independently
checked for crossings and contained vertices outside their intended distal grip.

## Finite object-driver precision

The full capture exposed a7.717526e-5ideal head-matrix residual near rest, equal
to an authored0.004422degree pitch. An empty-scene fixture reproduced positive
driver outputs below1e-4 becoming exactly zero. Copy Rotation itself preserves
tiny angles. The same behavior affects tiny positive carrier pitches near the
start/end. The native remains unchanged. Raw ideal residuals are retained; a
separate input check admits only this reproduced zero-snap case, and the actual
neck/head transform must still match those evaluated inputs within5e-6. All
source-shape, head-rigidity and contact checks retain their original strict
criteria. Six critical native times pass with this explicit distinction. See
CONSTRUCTION.md and the precision fixture/audit/preflight evidence.

## Complete action and visual review

All240native action times and116between-frame samples passed source construction,
fixed-paw, cap, tool-contact and interface checks. The maximum source-math error
was3.277e-6; both paw errors were exactly zero. Actual broom-floor error was at
most9.500e-8. Concealed cap triangles had no paw-surface crossings. Maximum arm
composition-depth increase was0.000236, below the documented0.0007remesh tolerance.
The four-second endpoint is exact. The full capture contains98distinct complete
geometry/shader states; image reuse required exact state equality.

All1,440six-view sequence images,144new orbit images,28body-only images and9public
control images were visually inspected. The earlier lower-body wedge is absent
through the turn, sweep and return. Rounded side/quarter contours, planted feet,
visible attachments and facial appearance remain coherent. The75pilot views are
retained for the three whole-action hypotheses. This is main-agent visual review,
not independent sign-off or owner acceptance.

The first22times had inherited a render-only shared-shader error from an earlier
capture method. Their132old images were preserved, and all22actual native render
states were re-evaluated before replacing pixels from an exactly equal verified
state. The native and original geometry/contact probes remain unchanged. The
corrected complete sequence was re-inspected. All1,621final PNGs were transferred
with exact SHA256 verification; six240frame60fps movies passed complete decoding
and framing checks. Capturing and prefix-repair queue jobs finished successfully.

The comparison offers synchronized before/after playback, six views, frame stepping,
three speeds, two complete72angle orbits, trial sheets and the editable native ZIP.
Results and limits travel with the archive. The former Sweep34 media are reused
byte-for-byte; prior originals and adopted releases remain unchanged.

## Scope and reproducibility

Native: `final/sweep36-supported.blend`,221222462bytes, SHA256
`fe4a51d47aebf9f5eadbfb8f0bf2eaca9905f2c71c16f9cdba06c6a7b2c123c7`.
Evidence root: `output/rig_candidates/chores214_20260925/v2/sweep36_constructive/`.
Source: `scripts/research/chores214_20260925/sweep36_constructive/`.
Read CONSTRUCTION.md for equations, source ownership, failed assertions, capture
repairs and exact-replica retirement. RIG_GUIDE.md describes editing and restoration.

Body contribution0–1 is one sampled coupled family. General independent control
combinations, physical force/balance recovery, stepping, pickup/release and a
skeletal runtime export remain outside this change. The heavy source-composition
node graph is not a real-time runtime certificate. Owner adoption and the inherited
fifteen-issue/eleven-gate full-motion qualification remain open. Implementation and
review were performed by the main agent; no independent sign-off is claimed.