The exp05 verdict ("weights alone cannot clear the bar on this mesh") was
measured against a hard partition that was itself causing much of the tearing.
exp01-exp05 gave every vert to exactly ONE digit -- 08_finger_weights.py had an
explicit `elif lo in FING: continue # other digit: hard wall` -- which
guarantees the fused inter-digit bridges tear by the full finger separation,
because a single edge ring absorbs the whole gap.
Replace that with graded blending (WEB_BLEND_R0): a vert's fraction toward its
nearest other digit ramps 0 -> 0.5 as r = d_own/(d_own+d_other) goes 0.35 -> 0.5,
so both sides of the equidistance valley reach 50/50 and the field is continuous
across the boundary. Applied after the smoother (whose hard wall would erode beta
exactly where it must survive) and evaluated in the neighbour digit's own
arc-length frame. WEB_BLEND_SKIP excludes the thumb: its transform is opposition,
not curl, so its frame does not correspond to a finger's.
Measured against an identical baseline (same input, blend the only variable),
real tears (>=1mm rest length) drop 23-60% with NO regression on flat, the pose
that ships: fist_r 1178 -> 530, grip_r 561 -> 222, fist_l 979 -> 633,
grip_l 420 -> 322; total >5x 2487 -> 1307; p99.9 better on every pose;
flat unchanged at 0/1. fin_bones confirms the mechanism rather than just the
count -- the middle<->ring and pinky<->ring families leave the top classes while
the thumb/palm ones are untouched to the edge (182 -> 182, 148 -> 148).
This does NOT make fist/grip shippable: 222-633 real tears still reads as a
destroyed hand in clay renders, and the residual is now ~53% thumb-pad-fused-to
-palm, which is topology and needs mesh surgery or the v02 rebake. Flat and
relaxed are the shippable poses; fist/grip belong to the morph lane for now.
Also here:
- README: the solver's input is v02/..._exp03.glb, NOT exp01. exp01 is pre-hand
-fit (converter steps 2b/2c); its finger groups sit on the wrist and overlap
the real finger by 1.6cm, so a solve from it silently zeroes every _02/_03 bone
-- rigid stick fingers and a torn flat -- while weight sums stay 1.0 and every
assert passes. Cost three wasted bakes and one false "the solver regressed".
- Seed assert demanded >=100 seeds while the radius loop caps at
SEED_AXIS_R_MAX, which left pinky (88 seeds at 16mm) can never satisfy; the two
constants were mutually unsatisfiable. Now >=60, and it is documented as a
sanity gate rather than a quality bar.
- Detwist poses tested at last: real but marginal (fist_r 38.7x -> 28.7x,
grip_r 30.0x -> 17.3x, left hand flat). A knob, not a fix.
- edge_stretch_cmp.py / skin_bone_territory.py / handpose_trim_hand_obj.py:
judge tears by rest length and absolute posed growth, not raw ratio; audit
whether a bone owns any verts at all (thumb_01 owns ZERO in exp05); and trim
an arm-sized skin dump to the hand before rendering.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The plan this track executed against. Its weld diagnosis was real but turned
out to be 1 of 6 defects, and its stretch-only acceptance bars were satisfiable
by a no-op morph — see hand-shapes/README.md and b57b0f2 for what actually
happened.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The HTML references screenshots relatively, so it only renders inside a
checkout -- mailing the file on its own gives broken images. The PDF is one
self-contained file, which is what actually gets sent to Ozlem.
- tools/humans_to_pdf.py: renders .humans/*.html via headless Chrome (or
Edge), so the output is what the browser shows. HTML stays the source.
- @media print block on the page: the wide tables and the pipeline diagram
scroll on screen but would be cut off at the paper edge; this reflows
them and keeps figures, callouts and table rows off page breaks.
Verified at A4 content width -- 6 pages, both screenshots embedded.
- *.pdf -> LFS. Each regeneration is a whole new ~2 MB blob, and unlike
png/jpg there are no PDFs tracked raw yet, so there is no history to
rewrite by adding it now.
Also corrects working-files.md: it claimed .gitattributes does not LFS-track
.blend, which was true when written and is no longer -- .blend/.zprj/.obj/
.npy/.npz have since been added. The note now covers what is actually still
excluded (png/jpg, deliberately) and why that makes QA renders scratch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bulk import of the working lanes that were living untracked on the PC.
Content:
- characters/ Lena/male body lanes, bakes, texture work, run logs
- clothing/ garment pipeline, configs, gates, contract docs
- garments/ MD-authored garment sources (.zprj/.zpac)
- UAL-Lib/ Universal Animation Library 2 source (.blend/.fbx/.glb)
- tools/ blender_bridge, iclone_bridge, md_bridge, tailor, glm_agent
- docs/, plans/, dev/, .agents/plans/
Repo hygiene:
- .gitattributes: LFS now covers .blend, .zprj, .zpac, .obj, .npy and the
Reallusion .iAvatar/.ccAvatar/.ccRestore containers. Without this the
~3.8 GB in this commit would land as raw blobs. .png/.jpg are left out
on purpose — ~250 are already tracked raw and converting them would
rewrite every one without shrinking history.
- .gitignore: exclude /accurig/ (~1 GB AccuRig program files, redistributable
from Reallusion, nothing authored here) and /dev/null/ (git-lfs hook copies
dropped by a `>/dev/null` redirect on Windows).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two pieces of work.
1. Keep milestones, not steps. Staged NN_*.py lanes were saving a full
~80 MB .blend per attempt, so Lena's lane reached 3.0 GB of which 2.2 GB
was 30 .blend files -- five snapshots to land one crotch fix, five more
for the bra. The .py recipes are the real history; the blends are cache.
- .agents/rules/working-files.md: four-tier policy (KEEP / MASTER /
SCRATCH / UNKNOWN) + how to work a lane.
- tools/prune_lane.py: classifies a lane and prunes the scratch tier.
Dry-run by default. Reads a per-lane .lanekeep manifest, flags
binaries byte-identical to a registered original (canonical name
always survives a duplicate pair), and never auto-deletes a .blend
with no step script beside it -- those cannot be rebuilt.
- .gitignore: scratch patterns can never be committed.
Dry run on characters/female/lena_nude reports 2.3 GB reclaimable.
Not applied -- that lane had a live Blender session at the time.
2. .humans/marvelous-designer.html: how we author garments in Marvelous
Designer, written for people rather than agents -- the six-step process,
what has been made, the traps that cost hours, and what is still
unsolved. Matches the .humans/ HTML convention in ariki-game.
Also committing the docs the AGENTS.md knowledge map and the new page
reference, so they are not dangling: the clothing-lane architecture page,
the marvelous-designer skill, and the two screenshots the page embeds.
characters/ is deliberately untracked and stays that way -- it holds GBs of
blends and GLBs, and .gitattributes does not LFS-track .blend.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Root AGENTS.md is now a thin entry point + knowledge map; the batch-workflow
detail it duplicated already lived in .claude/skills/animation/SKILL.md.
Folded docs/ (dances registry + iclone-bridge stub), root plans/, and
devops-reports/ into .agents/wiki/ and .agents/plans/ per the per-repo .agents/
convention. Added .agents/SOUL.md, .agents/AGENTS.md, wiki/ARCHITECTURE.md,
wiki/architecture/, and wiki/master-plan.md (derived from the closed plans +
the dance registry's live ceremony-slot table). Fixed doc-path references in
exchange/GUIDE.md, the animation skill, and the iclone-bridge stub. All moves
via mv (git detected as renames) — no content deleted.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
jeremy