A 3D model pipeline, pushed by GPT-6.2-Astra

agent (cont.)

+0:28
  • exec_commandcat docs/vroid.md && cat docs/map.md (+4 more)

    command

    cat docs/vroid.md && cat docs/map.md
    
    sed -n '1,500p' docs/portrait-guide.md
    
    sed -n '501,1000p' docs/portrait-guide.md
    
    sed -n '1001,1540p' docs/portrait-guide.md
    
    rg --files -g '!.git' | sort && du -h -d 2 ~/.local/share/lowmarch | sort -h | tail -80 && ls -la ~/out && git status --short --branch

    result, 612 lines

    Script completed
    Wall time 0.1 seconds
    Output:
    Warning: truncated output (original token count: 34284)
    Total output lines: 1962
    
    # VRoid base models
    
    **Status (2026-11-09):** current. The files carry 57 expression shapes (2.14 exports) or 56 (older), not the 54 below. Licence terms were read by hand on 11-09: see "Licence, read 2026-11-09" below, which replaces the licence notes further down (led #1443).
    
    ## Licence, read 2026-11-09
    
    Read by hand from pixiv's own pages on 2026-11-09 (headless browser, no login). Page copies and text dumps are in
    `~/citadel-render/workers/handoff/overnight-1109/lowmarch-prep/pages/`. This is a reading of the terms, not legal advice.
    The English terms of use are a translation; pixiv says the Japanese original prevails.
    
    Sources, with the date each page gives:
    - **[G]** VRoid Studio guidelines, edited 2023-12-21: <https://vroid.com/en/studio/guidelines>
    - **[T]** VRoid Studio terms of use, revised 2026-06-24 (articles renumbered then): <https://policies.pixiv.net/en.html#vroidstudio>
    - **[AZ]** "VRoidPreset_A - Z", the page the 2.14 samples point at, updated 2024-12-26: <https://vroid.pixiv.help/hc/en-us/articles/4402394424089>
    - **[SC]** "Do VRoid Studio's sample models come with conditions of use?", undated: <https://vroid.pixiv.help/hc/en-us/articles/4402614652569>
    - **[F]** FAQ "Can I use the models created with VRoid Studio Stable Ver. for commercial purposes?": <https://vroid.pixiv.help/hc/en-us/articles/4405813333657>
    - **[β1 to β4]** <https://vroid.pixiv.help/hc/en-us/articles/360012381793>, `.../360014900273`, `.../360014900113`, `.../360014900233`
    - **[SF]** Sakurada Fumiriya <https://vroid.pixiv.help/hc/en-us/articles/360014788554>; **[SS]** Sendagaya Shino <https://vroid.pixiv.help/hc/en-us/articles/360013482714>
    - **[H]** VRoid Hub terms <https://policies.pixiv.net/en.html#vroidhub>, and the official VRoid Project account's model licences
      (read from `hub.vroid.com/api/character_models/<id>`, the same JSON the model page shows)
    
    ### What we hold, and whose terms apply
    
    | Our folder | What it is | pixiv's terms |
    |---|---|---|
    | `avatarsample_d` to `_g` (OpenGameArt) | Not the 2.14 D to G. They are β Ver AvatarSample_1 to 4: Sendagaya Shibu, Vivi, Vita, Victoria Rubin (file titles match the Japanese β pages) | CC0 [β1 to β4]. The official Hub copy of β1 also allows everything, credit unnecessary [H] |
    | `sakurada_fumiriya`, `sendagaya_shino` | VRoid's mascots | CC0 [SF] [SS]. Official Hub copies allow everything, credit unnecessary [H] |
    | `hairsample_male`, `hairsample_female` | alpha-era hair samples | CC0 [SC] |
    | `base_male`, `base_female` (OpenGameArt) | alpha-era bases, Studio 0.14 exports | **Unclear.** [SC] lists "Base models" under CC0, as a picture of the 2021 bases. [G] says "All content provided by pixiv, including the base models when creating a new avatar, is not CC0." Either way [T] Art. 13 licenses them for any purpose |
    | `samples_2_14/avatarsample_a` to `_z` | VRoidPreset A to Z, from Studio 2.14 | [AZ]: not CC0, any use, no credit, a list of prohibited acts (below) |
    | `lowmarch_bodies/base_masc`, `base_fem` | Studio 2.14 new-model bases | Provided Content under [T] Art. 13: any purpose, not CC0 [G] |
    | `faces/face_sexy.vrm`, our armory bodies | our own outputs | Ours, except the pixiv content inside them, which keeps its own terms [T] Art. 11 |
    
    The armory bodies are built on `base_masc`/`base_fem` with A to Z faces and hair, so [T] and [AZ] are the terms that
    matter for them.
    
    ### 1. A character made in VRoid Studio from scratch
    
    Allowed: ship it in a game, sell the game, show it on a public page, post renders on a public blog, change it in code.
    - Rights: "The copyright and other rights for textures and hair meshes you created from scratch on VRoid Studio or other
      tools belong to you, the creator." The body and clothing meshes and presets are pixiv's, licensed "so that you can use
      it for as wide a range of purposes as possible." [G]
    - Licence: pixiv grants "a non-exclusive, perpetual, worldwide, non-transferable, non-sublicensable license to use,
      duplicate, archive, modify, and display Provided Content ... for (1) any purpose and (2) any use, except where
      different license conditions are specified and displayed for individual items" [T] Art. 13.
    - Games, sale, web pages, renders: "You can use images and videos of models created with VRoid Studio and generate
      profit from videos, books, ... games, applications, software, collections of materials, websites" and "You can sell
      the model data exported from VRoid Studio." [G]
    - Changing it in code: "You can edit and use the model data exported from VRoid Studio using software other than VRoid
      Studio." [G]
    - Limits: items with their own clause keep it ("When using items that display special licensing terms, please observe
      those specific terms" [G]); nothing may suggest pixiv backs it ("refrain from creating any publication or video
      production that leads people to believe it was created in collaboration with VRoid Studio or pixiv Inc." [G]); and
      the character-creator rule in 3.
    
    ### 2. The sample avatars
    
    - **β 1 to 4 (our old D to G), Sakurada Fumiriya, Sendagaya Shino:** all of the above, with no conditions. "The CC
      license* for this model is CC0. This model can be edited and used freely." [β1] [SF] [SS]
    - **A to Z (2.14):** game, sale, public page, blog renders and code changes are allowed. "This model's .vroid and VRM
      files can be used by anyone in any kind of activity, be it for-profit or not." Examples given: "Edit its texture and
      parameters, and create an original 3D character" and "Edit its texture and parameters, upload it to BOOTH for sale to
      others." [AZ] Prohibited, among others [AZ]:
      - "Distributing a sample model or its VRM file to a third party by setting the 'Redistribution and alteration'
        conditions of use to 'Copyright Disclaimer (CC0)'."
      - "Development of a character creation service using data contained in the sample model, or diversion of data
        contained in the sample model to an existing character creation service."
      - "Redistributing the .vroid or VRM file of the sample model (or other data such as image data within the model) in
        exchange for a fee."
      - "Acts that can lure people into thinking that the Company supports or recommends some product/event."
      - pixiv may narrow the terms later: "it's possible that the conditions of use may be limited if there are incidents".
    - **Unclear:** whether a paid game that ships a sample's textures unedited (our faces keep each sample's iris,
      highlight, brow and eyeline drawings as they are) counts as redistributing "image data within the model" for a fee.
      Edited and sold is allowed; how much editing is enough isn't said.
    - **Metadata against the page:** for our use, pixiv's page wins. The licence fields in our 2.14 exports
      (`Redistribution_Prohibited`, `OnlyAuthor`, commercial `Disallow`) were set at our export, not by pixiv: "Users will
      be licensed to set licenses freely for VRM data constituting Output Items, as long as they do not breach the license
      conditions for Provided Content" [T] Art. 12. They bind whoever receives our file, not us. pixiv's own terms for its
      items apply to us: "Where the Company has specified and displayed license conditions for individual items of Provided
      Content, Users will use such Provided Content in accordance therewith" [T] Art. 13. No page states a precedence rule
      in so many words; this is a reading of those two articles. The OpenGameArt files carry pixiv's own metadata (a VRoid
      Hub licence allowing everything, or CC0), which agrees with the pages.
    - If we ever hand out VRM files built on A to Z, their licence field must not be CC0 [AZ].
    
    ### 3. A player-facing character creator
    
    Confirmed: it needs a separate licence from pixiv, and pixiv may refuse.
    - "Users are not licensed to use Provided Content, or 3D models or Output Items containing Provided Content, to create
      applications having functions that allow avatars, items, or other 3D models to be created by deforming meshes or
      combining meshes or textures. When a User wishes to create an application containing such a function, they must
      receive a separate license from the Company; provided, however, that this does not apply when the User is creating
      the application for personal use." [T] Art. 13; also a prohibited act in [T] Art. 7.
    - "Please contact us if you'd like an application creation license. It's possible that we may decline granting you a
      license." [G] [F]
    - [AZ] prohibits a creator built on A to Z data outright and names no licence route.
    - So our own build tools are fine (personal use), and baked characters in a game are fine. A creator in the game, built
      on anything from Studio 2.x or A to Z, needs pixiv's licence.
    - **Unclear:** whether CC0 parts alone (β 1 to 4, the mascots, the hair samples) would escape this. CC0 waives
      copyright, but [T] is a contract we agreed to by using Studio, and our exports pass through it. Not settled by the
      pages.
    
    ### 4. Credit lines
    
    None required.
    - A to Z: "There is no need to credit the original creator when using this sample model." [AZ]
    - CC0 models: "Copyright is waived, and there is no particular limit when using them." [β1]
    - Official VRoid Hub copies: `credit: unnecessary` [H].
    - Studio outputs: [G] and [T] ask for none (none found). [G] asks only that nothing implies pixiv's backing, and notes
      "VRoid is a trademark or registered trademark of pixiv Inc."
    
    ### Superseded
    
    The licence notes further down (the table's "Licence in the file", the yomox9 D to Z note, "That page is behind a
    Cloudflare check and wasn't read") predate this read. Where they differ, this section holds.
    
    Since 2026-10-08 our portraits (and later battle sprites) are rendered from VRoid sample models
    through our own pipeline. VRoid supplies a professionally built anime head, hair, body, skeleton and
    54 expression shapes. Pose, archetype reshaping, hair swaps, colours, shading, outlines and the
    reduction to GBA pixels are ours.
    
    ## Where they live
    
    `~/.local/share/lowmarch/vroid/<name>/*.vrm` (override with `LOWMARCH_VROID`). They are 11–17 MB
    each and stay out of the repo. Source: <https://opengameart.org/content/vroid-studio-cc0-models>.
    
    | Model | Licence in the file |
    |---|---|
    | Base_Male, Base_Female, HairSample_Male, HairSample_Female | CC0 |
    | AvatarSample D–G, Sakurada Fumiriya, Sendagaya Shino | VRoid Hub licence: everyone, commercial use allowed, modification allowed, redistribution allowed, credit unnecessary |
    
    The last two are VRoid's mascot characters. Use their parts (hair), not their look. Our characters
    are built from the CC0 bodies.
    
    **AvatarSample A–Z (researched 2026-10-09):** exported by `tools/winvroid` into `samples_2_14/avatarsample_<x>/` (apart from the older D–G above, which are other characters).
    All 26 were exported on 2026-10-10 (VRM 0.0, VRoid Studio 2.14, 15 to 22 MB each). They are clothed
    exports, so VRoid has deleted the body skin under their clothes: use their heads and hair, not their bodies.
    
    **VRoid's own bases, unclothed (2026-10-10):** `lowmarch_bodies/base_masc.vrm` and `base_fem.vrm`: a new
    character from the Masc or Fem base at default settings, every outfit category set to none, exported as
    VRM 0.0 (`python3 tools/winvroid/winvroid.py bases`). They keep all their skin (10,934 body triangles;
    the old `base_male/Base_Male.vrm` export has 7,340). The briefs and band are painted on the skin texture,
    not geometry. `tools/mass.py` builds on `base_masc.vrm`.
     VRoid Studio 2.1.0 and later ship 26
    sample avatars. The VRoid Project account on VRoid Hub publishes A–C with VRoid Hub's licence:
    - everyone may use them;
    - corporate commercial use is allowed;
    - modification and redistribution are allowed;
    - no credit is needed.
    
    D–Z are on VRoid Hub only as exports by another user (yomox9, character 2175138617022572467). Their
    files carry the same terms in VRM 1.0 metadata (`commercialUsage: corporation`,
    `allowModificationRedistribution`, credit unnecessary), pointing at pixiv's "AvatarSample A-Z" help page.
    That page is behind a Cloudflare check and wasn't read. Take them from VRoid Studio itself, or the
    official account where it has them.
    
    What the set offers:
    - about 9 male and 17 female faces, all young and slim, two of them adult men (T, V);
    - different eye drawings, hair and outfits.
    
    **VRoid Studio's guidelines** (vroid.com/en/studio/guidelines, read 2026-10-09):
    - **Allowed:** models made in VRoid Studio may be sold, edited in other software, and used in games.
    - **Restricted:** "You cannot create an application that can generate or output 3D models, avatars,
      or items consisting of deformed or combined meshes and textures that were created in VRoid Studio"
      without a separate licence from pixiv. Tools meant only for your own use are exempt.
    - **For us:** our build pipeline (the graft, the costume kit) is for our own use, and shipping baked
      characters in the game is fine. A player-facing character creator built on VRoid parts would need
      pixiv's licence.
    
    ## Heads (`tools/heads.py`, 2026-10-13)
    
    The armory's bodies all wore VRoid's base face. A head is now one of the 26 samples' faces on our base:
    - **One template:** every VRoid face mesh has the same skin, eye-white and mouth triangles and texture coordinates
      (4,232 skin triangles in each), so a face moves vertex for vertex. The irises, highlights, brows and eyelines are
      each sample's own geometry: that is where its eye drawing lives, and it comes along whole.
    - **Only the face is taken:** eyes, nose, mouth, cheeks and jaw, fading out up the forehead, toward the ears and toward
      the seam with the body. VRoid's face mesh is the head's front; the back of the skull is in the body mesh, and the
      seam between them runs from the neck to the crown. Taking a sample's whole front set its skull against the base's
      back, and VRoid's hair base (the short cap) couldn't sit on both. The base's skull, ears, seam and cap stay.
    - **Also taken:** the sample's 57 expressions (the same names, in the same order, in every model), its skin weights
      mapped by bone name, its eye bones (so the eyes turn about their own centres), and its eye, brow, eyeline and mouth
      materials. The face skin keeps the base's texture, so the skin tone matches the body.
    - **Sculpted in the file:** the face controls of `tools/vroid_portrait.py` (lower-face length, jaw, chin, cheek, brow
      ridge, nose, eye size and tilt, iris, mouth), with one rule for the face and the neck so the seam holds. The checks
      then measure the face that is shown.
    - **Blends:** a face spec's `sample` may be several samples with weights; the face's shape is their weighted mean
      (one template, so a blend stays coherent from every angle) and the first gives the eye drawing and expressions.
    - **Not fitted (2026-10-13):** `heads.fitted` can move the lower face's length, the cheeks and the eye size until a
      face archetype's ratios pass (`docs/proportions.md`), but the armory doesn't use it. Those ratios were measured on
      the study's paintings, mostly seen from the front; held to them, faces were stretched (lower faces up to 17%
      longer, eyes 20% larger) and read too sharp from the side and too wide from the front. VRoid's faces all set the
      eyes low (0.56 to 0.64 of the head down from the crown), the young look; the study's adults have them at 0.40 to
      0.58. The armory's faces are VRoid's as built, chosen and blended by look, and the ratios are reported.
    - **Then the cage** (`tools/cage.py`, a spec's `face`) builds the body on the headed base and sizes the head to its
      reference's, so head size stays held in metres.
    
    What the face rests in (a frown, half-closed eyes) and age lines are drawn when the figure is built, not in the file.
    
    ## Tools
    
    - `tools/vrm.py`: VRM/glTF loader (meshes, materials, textures, morph targets, skins) and a numpy
      z-buffer rasterizer with alpha testing.
    - `tools/vroid_portrait.py`: `Portrait(model, morphs, pose, shape, hair_from)` and `.render(yaw)`
      give a 96×80 char grid; `image(char, palette=...)` gives a PNG; `smooth_view()` shows the 3D model.
    - `tools/outfit.py`: costume pieces generated in code and fitted to the posed body (`COSTUMES`).
    - `art/cast/cast.py`: one entry per character (base model, hair donor, shape controls, colours, costume).
    
    ## Notes
    
    - VRoid models face −z; `FLIP` turns them toward the camera. Character's left is −x.
    - They stand in a T-pose. `BUST_POSE` lowers the arms through linear blend skinning.
    - Expression names end in `Fcl_EYE_Close`, `Fcl_MTH_A`, `Fcl_ALL_Joy` and so on; match the suffix
      exactly (`_Close` also prefixes `_Close_L`).
    - The eyeline texture is a thick graded band over the iris. Only texels darker than `LID_CORE`
      count as the lash line, or the eyes read as half shut at 1×.
    
    ## Review tooling
    
    `python3 review/vroid-02/build.py aldric` builds the model viewer: `bust`/`battle` GLB models
    (`tools/export_glb.py`), pixel frames every 15°, and `camera.json` so the three.js orthographic
    camera matches the pixel camera exactly. `tools/poses.py` holds poses (BUST, BATTLE_IDLE, `fist`),
    `tools/terrain.py` holds battle ground (`plains`).
    
    ## Hair (`tools/heads.py`, 2026-10-13)
    
    A head spec's `hair` names a sample whose hairstyle goes on. Every VRoid head matches point for point (the face skin,
    and the 562 vertices of skull in the body mesh, by texture coordinates), so the style is warped from its sample's
    head onto ours (a similarity plus the matched points' residuals spread by inverse distance) and held 1.5 mm off the
    skin. It keeps its hair base (HairBack), which replaces the base's; its hair bones (only those its mesh moves with:
    the samples' skins list accessory ears too, and the weightless end bones that say where a lock's last bone points);
    its spring groups; and its head's collision sphere, warped onto our head. `tools/cage.py` scales the head's sphere
    and the hair's spring radius with the head. Hair colour is the character's.
    
    **Collision (`tools/collide.py`, 2026-10-13).** Spring bones collide only with spheres the model carries
    (`secondaryAnimation.colliderGroups`); every runtime that plays a VRM (UniVRM, three-vrm, the review page) runs
    those and nothing else, so what the hair can't pass through is set here, not by the engine. The samples' body
    spheres were sized for VRoid's slim sample bodies, with none over the back or bust, and v11's hair sank up to
    6 cm into our bodies' upper backs, shoulders and napes. Now, on each finished body, `tools/cage.py` calls
    `collide.hair`:
    - **Spheres fitted to the skin:** medial balls (each touching the skin at a point, as large as fits inside), chosen
      greedily until all the skin the hair can reach (within a lock's length below its joints; the shoulders and upper
      arms whole once the hair reaches them) stands within 1 cm of one. Each rides the bone most of its skin is
      weighted to, the breasts' spring bones included. They stand 2 cm out from the skin: a lock's bone runs down its
      middle, and VRoid's locks stand a median 2.3 cm round it, so a bone held at the skin left the lock in it. That
      comes to 30 to 60 spheres per body.
    - **The hair draped at rest:** each lock's bones, root to tip, keep their own way unless the body is in it. A bone
      is turned until its whole length (not only its tail: the body bulges between two joints) is out of the spheres
      and the mesh it carries stands off the skin, and what no turn can lift (points the head carries) is pushed off.
      That is the bind pose, so the hair starts where it rests, not inside the body.
    - **The check:** `tools/hair_check.py` replays the page's spring bones (Still, Twist, Bounce; arms out or down) and
      counts hair vertices inside the skin.
    
    ## Designing faces in VRoid Studio (first session, 2026-10-13)
    
    - **The face editor** (Face tab) has 89 sliders: eye size, position and slant, eyelids, droopy and scornful eyes,
      iris size, brows, nose, mouth and lips, teeth, ears, jaw, chin and cheeks, and masculine and feminine blends of
      each feature. Its **Face Se…24284 tokens truncated…triangles (a zig-zag with notches) and ended where the neck's own bumpy normals took over. The nose patch did the
    same and turned VRoid's painted nose line into a "T", like a stitch.
    
    **What was built (`shadow_cards`, replacing `shade_marks`):** the shadows are small pieces of geometry with designed
    outlines, as on the Yin and Yang rigs: a crescent on the neck under the jaw and a half-moon under the nose tip.
    - They lie just off the skin and borrow its weights, so they ride and bend with it.
    - They take a new role, "shadow": skin that only ever darkens into its shade tone, in the painted renderer, the
      export and the pixel path.
    - `face_normals` gains the rig's second skeleton, a cylinder round the neck's axis, so the neck itself shades
      smoothly around the card.
    
    **Lessons:**
    - **A cel edge drawn by the mesh is only as smooth as the mesh.** Where the shape of a shadow matters, give it its
      own geometry.
    - **Follow the real edge, not a guess at it.** A parabola for the jaw line left the card's top ragged where it met
      the face; the jaw's lowest edge, measured round the neck's axis and smoothed, hides it.
    - **A card must agree with the shade around it.** Fixed at one tone it showed a darker V inside the neck's own
      half-shade. Letting it only darken (never lighten) merges it where the neck is already in shade.
    - **Take the neck from the skeleton, not from a slice of skin.** In the armory's T-pose the slender woman's neck is
      short: 5 cm under her chin is already shoulders and arms, so a ring cut there put her neck's centre at the front
      of her throat. Half her chin card failed and the rest came out shallow. Her visible "shadow" was mostly the neck's
      own shading, so she and the youth looked unlike each other. The line from the neck bone to the head bone (it leans
      as the neck does) fixed it.
    - **Hold a light constant to see what a shape really is.** Rendering the neck fully lit, so only the cards shaded,
      showed at once that her card was nearly missing, and that the light was not to blame.
    - **A chin shadow is the upper neck, not the whole of it:** deepest about a quarter of the eye-to-chin span under
      the chin, ends tapering to points before the jaw's angle, and a soft shade tone rather than the darkest one.
    
    ### 35 · The chin's shadow cast in screen space, 2026-10-30
    
    **What you said:** at angles the chin card still looked wrong; we might need research.
    
    **What the research said:** anime pipelines either paint neck shadows into the texture (Genshin), edit normals and
    shadow thresholds (Guilty Gear Xrd), or cast them in screen space ("depth offset" shadows: Anime Shading Plus,
    Stylized NextGen). They avoid ordinary shadow maps on faces. A painted or card shadow is fixed to the neck, while a
    real chin shadow comes from the jaw's outline and changes as the head turns.
    
    **What was built:** the painted renderer draws its info pass (roles, ids, depth) first. For body skin, the colour
    pass walks from each pixel toward the key light across that pass: [redacted: secret] the face (jaw, chin) or hair stands in front of
    the ray, the pixel drops into the shade band. The chin card is off (`CHIN_CARD`); the nose card stays, since a face
    can't shadow itself this way.
    
    **Lessons:**
    - **A shadow's shape belongs to its caster.** Thrown from the jaw's own outline, the chin's shadow is right at
      every angle with no shape to design.
    - **Unbind what you are about to draw into.** The first version rendered one tile and then black: the info textures
      were still bound for sampling when the next frame drew into them (a WebGL feedback loop).
    
    ### 36 · A skin range, people and fantasy, 2026-10-30
    
    **What you asked:** more range for the models, starting with skin, including fantasy tones (dark elves and the
    like). The pixel renderer is abandoned; the armory's painted renderer is the target.
    
    **What was built:**
    - `cast.SKINS` gains 12 people tones (porcelain to ebony) and 7 fantasy tones (two dark elf, ashen, two orc, sea,
      infernal).
    - `export_glb.skin_ramp` gives each tone its own lit, shade and deep-shade colours.
    - The armory page has a skin picker that tries any tone on the bodies shown (`?skin=<tone>` for sheets).
    
    **Lessons:**
    - **One tint over a light skin can't make a deep one.** Multiplying the light ramp darkened every tone the same way,
      and the stage's blue shade turned brown skin grey. A ramp per tone, with shade hue nudged warm on people (cool on
      fantasy blues and greens) and the stage grade taken at 30% on skin, keeps deep tones rich.
    - **Space the range by eye, not by name.** The older seven bunched at the light end (pale, light, fair look alike)
      and stopped at a mid brown.
    
    ### 37 · Shut eyes shut (from main's 4aa16f2), 2026-10-31
    
    - **A blended face may not shut its eyes.** The youth's lash lines come from another sample than his lid skin, and
      VRoid's close shapes stopped short on him: a strip of iris, then white, then socket showed in joy and the wink.
      `hide_closed_irises` closes a shut eye 30% further, hides its iris and highlight (per eye, so the wink's open eye
      keeps them), and the shut eye's white is dropped.
    - **A closed lid is the model's own open lash, moved** (the rule: lesson 38). VRoid's close shapes move the lid edge further than the
      lash's top, so a closed line came out about 1.3 times the open lash on the youth (a slab beside his thin stroke).
      `keep_lash_weight` moves each UV column of the upper line rigidly with its lid edge, so it keeps the open weight,
      taper and colour. `lid_gap` measures per model whether the lids meet (the edge reaches the painted lower line):
      if so nothing more (closing past 1 folded the woman's lash fan into strays); if not, the lid closes over the whole
      lower line. A shut eye's lower line is dropped, and the lash fan moves with the edge rather than folding.
    - **A big opening needs a dark inside.** VRoid's mouth map read as teeth/tongue/dark looks right on small mouths but
      stretches into a pink blob on a shout; the composite darkens the upper half of a tall opening.
    
    ### 38 · One lash per character (a rule), 2026-10-31
    
    **One lash size per character: open and closed lids share it. A closed lid is that lash moved, never a stack of
    lines or a generic close.**
    
    - **How:** `keep_lash_weight` moves each UV column of a closing eye's upper line rigidly with its lid edge, so the
      closed (or lidded) eye carries the open lash's weight, taper and colour; VRoid's own close stretched it (about 1.3
      times on the youth). `lid_gap` measures per model whether the lids meet: if they do, nothing more; if they stop
      short, the lid closes over the lower line, and a shut eye's lower line is dropped, so the lid is one stroke. The
      lash fan (FaceEyelash) moves with the edge rather than folding past shut.
    - **Where the size is set:** today it is implicit in the VRoid sample that gives the eyes (the first in a head's
      blend, `cast.HEADS`). A designed size goes in the character's shape as `lash` (NEUTRAL default 1, the sample's
      own; per character in its cast `shape`). `lash_size` applies it to the open lash, and every closed lid is moved
      from that open lash, so closed lids follow it automatically; there is no closed-eye lash to set.
    - **VRoid's lash highlights:** the pale specks VRoid paints inside its lash lines are filled at export
      (`export_glb.fill_line_specks`), in open and closed lashes alike.
    - **A shut lid is one solid stroke.** A shut eye's upper line is its own primitive (`lid_fill`), drawn with
      `export_glb.fill_lid`: the lash's main stroke solid from its top ink to its bottom ink, its pale streaks and
      hollows inked, a crease line or loose wisp above it cleared. On an open eye those read as shine and a crease; over
      a closed lid they split it (the brute's hollow lash read as an empty almond, the average's streak as two lines).
      The open eye is drawn as VRoid painted it.
    - **Check it before an armory republish:**
      - `python3 tools/lash_check.py [names]` (meshes; after `tools/cage.py`, under 10 s a body): open lash against
        every lid the recipes close 30% or more, across the lash's middle columns. Flags a weight past 15%, and a shut
        lid that falls into two lines across a pale gap in 2 or more of its 9 columns.
      - `python3 tools/lash_shots.py SHOTS [bodies]` (rendered shots): the same in px on the frontal tile; a body whose
        lash is mostly under hair is "not measurable", which is not a pass.
      - 15%: half the 1.3 times the owner rejected, and above either measure's noise.
    
    ### 39 · Expressions, after the Yin and Yang rigs (the pass before the benchmark), 2026-10-30
    
    Written on master before the expressions benchmark. Its code was replaced by the combined track (B's brows and
    recipes, A's mouth and wink, shut eyes from lessons 37 and 38); the benchmark is in `~/Projects/lowmarch-bench`.
    The lessons below still hold.
    
    **What you asked:** compare the rigs' expressions (their selling point) with ours, then iterate on ours and show
    frontal images of each.
    
    **What the comparison showed:** see the trials folder, `expressions-vs-yin-yang.png`.
    - The rigs pose expressions with macro controls (Action constraints) and draw closed mouths with Line Art.
    - Ours use VRoid's 57 shapes. Joy, wink and the vowels already read.
    - Open mouths were a flat dark blob.
    - Brows hid under the bangs, so angry and sad didn't read.
    - There was no one-sided mouth.
    
    **What was built (the painted renderer and `tools/vrm.py`):**
    - **The mouth from VRoid's own texture, mapped to an anime palette:** maroon cavity, rose tongue, white teeth. The
      rim and a closed mouth's sliver are inked, and face pixels round the mouth get an ink line.
    - **Brows through the bangs:** a brow-only pass shows a brow where hair stands just in front of it (within 6 cm).
    - **VRoid's painted lip line inked** inside the template's lip window.
    - **One-sided shapes:** `Fcl_MTH_Up@L` applies a shape to one half with a soft middle. The smug is one corner up and
      one brow raised.
    
    **Lessons:**
    - **Use the parts you already have.** VRoid's mouth texture had teeth and tongue all along; a flat colour threw them
      away.
    - **Map texture to style, don't show it raw.** The raw texture read washed-out pink. A three-colour palette by
      brightness reads as an anime mouth.

    212 more lines not shown.