.claude/skills/character-pipeline/SKILL.md

@@ -72,9 +72,69 @@ Blender is required (`BLENDER_PATH`, or auto-found under
 | Subdivide cloth panels | `retarget.py::subdivide_cloth_panels` | A skirt being able to bend at all |
 | Retarget clips as rest-relative deltas | `retarget.py::retarget_clip` | Limbs not being twisted by foreign bone roll |
 | Leave cosmetic bones unkeyed | `export_optimize_animation_keep_anim_armature=False` | Physics owning the cloth |
+| Classify every SURFACE | `tools/surface_map.py` | The runtime never re-guessing what a surface is |
 | Write the rig sidecar | `retarget.py::describe_rig` | The runtime never re-guessing anatomy |
-| Cel look | `LevelMaterials.apply_character_look` | The model's own line-work not being re-lit |
+| Cel look, per surface class | `LevelMaterials.apply_character_look` | Hair not reading as a solid dark cap |
 | Cloth + hair | `characters/spring_bones.gd` | Clothes reading as clothes |
+| Per-character judgement calls | `characters/tuning_store.gd` | Art direction not becoming another constant |
+
+## The surface table
+
+The sidecar carries a `surfaces` list saying what each mesh surface IS — `body`,
+`cloth`, `hair`, `accessory`, or `linework` (the model's own ink shell, which is
+not a surface of the character at all). It is keyed on the MATERIAL name, because
+every character in this game arrives with its meshes called `Object_7` through
+`Object_32` while material names survive the glTF round trip intact.
+
+`SkinSurfaces` reads it and `apply_character_look` acts on it: hair takes a much
+thinner outline than the body, cloth a heavier one and a crisper terminator,
+accessories the heaviest. `SkinnedPlayerModel.surfaces_of(cls)` answers the
+question for anything else that needs it.
+
+Backfill a character that predates it, without re-importing:
+
+```bash
+blender --background --python tools/surface_map.py -- \
+    assets/characters/skins/<name>.glb
+```
+
+New imports get it from `describe_rig`, built from the same chains the solver
+uses, so the surface table and the cloth solver can never disagree about which
+bones are a skirt.
+
+## Per-character judgement
+
+Anything derivable from the skeleton is derived. What is left is genuinely an
+artist's call, and it lives in layered JSON rather than in a constant:
+
+| File | Scope | Class |
+|---|---|---|
+| `assets/characters/weapon_holds.json` | character + weapon | `WeaponHoldTuning` |
+| `assets/characters/rig_anchors.json` | character | `RigAnchors` |
+
+Both layer `defaults` → `skins.<skin>._all` → `skins.<skin>.<subject>` through
+`TuningStore`. An absent file means "use what the code derives", so nothing here
+is required for the game to run. Adding a knob is adding a row to a `KNOBS` spec
+table — the lab builds its whole UI from those.
+
+`RigAnchors` is where "the grip sits here in the palm" lives. A hand bone's
+origin is the WRIST; how far down the palm a grip belongs depends on the
+character's hand and cannot be derived. It defaults to identity, and identity is
+exactly the derived mount. Do **not** put a fixed rotation on the weapon mount
+instead — a bone attachment is expressed in the BONE's axes, no two rigs agree on
+those, and that constant is why the hand mount points were once wrong on every
+character.
+
+## The rig lab
+
+```bash
+godot --path . res://debug/rig_lab.tscn
+```
+
+Pick a character, a weapon, a pose or a single clip. Drag sliders for the HOLD
+(character + weapon) and the ANCHORS (character), and save. Click a surface class
+to isolate it — that is how the classifier gets checked: click `hair` and
+anything else still standing was misclassified.
 
 ## Read before you touch anything
 
@@ -129,13 +189,30 @@ is a list of things that cost a debugging cycle to learn.
 ## After importing
 
 ```bash
-godot --headless --path . -s res://debug/spawn_smoke_test.gd     # 29 checks
+godot --headless --path . -s res://debug/spawn_smoke_test.gd      # 29 checks
-godot --headless --path . -s res://debug/cloth_clip_check.gd     # leg-through-cloth
+godot --headless --path . -s res://debug/surface_class_check.gd   # every surface classified
-godot --headless --path . -s res://debug/cloth_settle_check.gd   # idle stability
+godot --headless --path . -s res://debug/character_picker_check.gd # the escape-menu roster
-godot --headless --path . -s res://debug/cloth_perf_check.gd     # ms per character
+godot --headless --path . -s res://debug/rig_anchor_check.gd      # anchors move the weapon
-godot --headless --path . -s res://debug/travel_dir_check.gd     # legs face travel
+godot --headless --path . -s res://debug/cloth_clip_check.gd      # leg-through-cloth
+godot --headless --path . -s res://debug/cloth_settle_check.gd    # idle stability
+godot --headless --path . -s res://debug/cloth_perf_check.gd      # ms per character
+godot --headless --path . -s res://debug/travel_dir_check.gd      # legs face travel
 ```
 
+And LOOK at it, which is where four of the last five real defects were found:
+
+```bash
+godot --path . -s res://debug/roster_capture.gd -- <dir>   # every character, one shot each
+godot --path . -s res://debug/ui_capture.gd -- <dir>       # every menu screen
+godot --path . res://debug/rig_lab.tscn -- shot <png> <skin>
+```
+
+`surface_class_check` fails on any surface that falls through to the heuristic
+rather than resolving from the table. That is deliberate: a model whose names
+stopped matching still RENDERS, because the fallback catches it — it just
+quietly loses its per-class art direction, which is exactly the kind of
+regression nothing else would report.
+
 What "good" looks like on Taila, for calibration:
 
 | Measure | Good | Bad |
@@ -149,15 +226,16 @@ What "good" looks like on Taila, for calibration:
 
 ## Characters currently shipping
 
-| Skin | Source | Meshes | Cloth chains |
+The six in `skins.json`, with what `surface_class_check` reports:
+
+| Skin | Source | Cloth chains | Surfaces |
 |---|---|---|---|
-| taila | rigged, Sketchfab CC-BY | 18 | 35 |
+| taila | rigged, Sketchfab CC-BY | 35 | 18 — body 6, cloth 7, hair 1, linework 4 |
-| kiyoko | VRoid, CC-BY | 13 | 20 |
+| kiyoko | VRoid, CC-BY | 20 | 13 — body 8, cloth 3, hair 2 |
-| aria | VRoid, CC-BY | 15 | 15 |
+| aria | VRoid, CC-BY | 15 | 15 — body 9, cloth 4, hair 2 |
-| momo | VRoid, CC-BY | 5 | 9 |
+| momo | VRoid, CC-BY | 9 | 5 — body 2, cloth 1, hair 2 |
-| hikari | VRoid, CC-BY | 13 | 10 *(zero-length — do not simulate)* |
+| miku | unrigged source, auto-rigged | 0 | 4 — body 3, hair 1 *(one mesh, four surfaces)* |
-| miku | unrigged source, auto-rigged | 1 | 19 *(hair grown)* |
+| mannequin | Quaternius CC0, from the animation library | 0 | 2 — body 2 |
-| mannequin | Quaternius CC0, from the animation library | 1 | 0 *(no costume)* |
 
 ## Known-unsolved
 
@@ -170,3 +248,14 @@ What "good" looks like on Taila, for calibration:
 - **No strafe or backpedal clips.** Direction is conveyed by yawing the hips
   (`SkinnedPlayerModel._update_travel`), which is capped, so a pure sidestep
   still runs its legs ~40° off the direction of travel.
+- **momo's idle pose is wrong** — arms overhead and a pinched waist. Every
+  assertion passes on her: she loads, animates, classifies and mounts a weapon.
+  It shows up only in `roster_capture`. Her source rig is the reason to suspect —
+  its bones are auto-named (`Item_O_Sphere_005`, `Unused_Noname_004`,
+  `Root_001_001`), so the role resolver has almost nothing to go on, and a limb
+  role claimed by the wrong bone would look exactly like this. Start by dumping
+  her resolved roles against her skeleton.
+- **A stray `Icosphere` ships inside every skin GLB** — 42 vertices, no parent,
+  no vertex groups. It rides in from the animation library. Harmless, and now
+  skipped by construction rather than by name in `surface_map`, but the export
+  should not be producing it.

.claude/skills/character-pipeline/references/separation.md

@@ -19,6 +19,7 @@ the convention):
 | Cloth/hair bones carry NO animation keys; physics owns them | `export_optimize_animation_keep_anim_armature=False` |
 | Physics colliders are a small set of capsules: thighs, shins, and a big one at the waist acting as a lid | 5 capsules, measured from the mesh (`_leg_colliders`) |
 | Neighbouring skirt panels are linked sideways | 278 cross-panel distance links from shared vertices |
+| Each surface is TAGGED with what it is, so shading can differ per class | `tools/surface_map.py` writes it; `SkinSurfaces` reads it |
 | Cel shading with a ramp, plus a separate outline pass | `LevelMaterials.apply_toon_recursive` + `apply_character_look` |
 
 Where we differ: their collider capsules and cloth parameters are hand-authored
@@ -26,6 +27,38 @@ per character by a technical artist. We MEASURE them from the model's own
 geometry at build time, because there is no artist in this loop. That is the
 whole reason `<model>.rig.json` exists.
 
+Where there IS an artist in the loop, there is now somewhere to put the answer:
+`debug/rig_lab.tscn` and the layered files behind it (see the SKILL). Measuring
+is the default and hand-authoring is the override, rather than the other way
+round.
+
+## Separation is only half of it — the parts have to be NAMED
+
+Keeping the meshes apart is structural. Knowing which is which is what lets
+anything act on the difference, and until the surface table existed nothing did:
+every surface of every character took one set of shading numbers, calibrated on
+skin, because there was no way to ask whether a surface was hair.
+
+The table lives in the sidecar as `surfaces`, keyed on the MATERIAL name — mesh
+node names are `Object_7` through `Object_32` on every character in this game and
+carry no meaning, while material names survive the glTF round trip intact and are
+what the artist actually chose. It is decided three ways, in descending order of
+how much it trusts them:
+
+1. **the material name.** On VRoid exports it is formal —
+   `N00_000_00_Body_00_SKIN_Instance` carries its own class infix, and every
+   VRoid character here uses SKIN / FACE / EYE / HAIR / CLOTH.
+2. **the weights.** Decisive when the name says nothing: a surface pulled by the
+   skirt chain is a skirt whatever it is called. The threshold is deliberately
+   low (5%), because VRoid welds the whole cap of the hair to the head bone and
+   springs only the strands — kiyoko's hair mesh is 85% head, and a majority rule
+   would call it skin.
+3. **the material flags.** These catch line-work, which is the one class that is
+   not a surface of the character at all.
+
+It is built from the same chains the spring solver uses, so the two can never
+disagree about which bones are a skirt.
+
 ## Why the separation is load-bearing
 
 **Materials.** The body wants skin shading, hair wants an anisotropic-ish ramp

.claude/skills/character-pipeline/references/stylization.md

@@ -23,10 +23,17 @@ rim term lights it brightly exactly where it is supposed to read as ink.
 it flat and unshaded. "Untextured" alone is NOT the test — that made every
 flat-coloured model render as a black silhouette, because Quaternius' mannequin
 has two untextured materials (a yellow body, lilac joints) and both were hidden
-as though they were an outline shell. `_is_line_work()` asks three things
+as though they were an outline shell. The test asks three things instead: is it
-instead: is it named `eyes*`, is it drawn front-face-culled (the classic
+named `eyes*`, is it drawn front-face-culled (the classic inverted-hull setup),
-inverted-hull setup), or is its albedo near-black. An ink shell is black; a
+or is its albedo near-black. An ink shell is black; a flat-coloured character is
-flat-coloured character is any colour at all.
+any colour at all.
+
+That test now runs at BUILD time (`tools/surface_map.classify_linework`) and its
+answer lives in the sidecar. `SkinSurfaces.guess()` is the same rule kept as the
+runtime fallback, for a model with no surface table — and the cull-mode half of
+it is re-run at runtime even when the table exists, because glTF has no way to
+say "draw only the backfaces" and an inverted hull cannot survive the round trip
+as a cull mode. Blender genuinely cannot see it; Godot can.
 
 What it then does:
 
@@ -37,9 +44,29 @@ What it then does:
   with `HL` in the name, which is the glint in the pupil and really is white.
 
 If a newly imported character comes out with a white halo, a black silhouette,
-or black eyes that should have irises, `_is_line_work()` and the name-matching
+or black eyes that should have irises, the line-work test and the name-matching
-below it are where to look. The conventions vary by source and this is the one
+below it are where to look — now in `tools/surface_map.py`, mirrored by
-place they are read.
+`SkinSurfaces.guess()`. **Change both or neither**: a model with a surface table
+would start rendering differently from one without.
+
+## Per-class art direction
+
+Because each surface says what it is, each class takes its own numbers
+(`LevelMaterials.CHARACTER_LOOK`). `body` is deliberately identical to what every
+surface used to get, so the calibration this was all built on does not move. The
+others are departures, each for a reason:
+
+| Class | Outline | Band | Why |
+|---|---|---|---|
+| body | 5.0 mm | 0.16 | unchanged — the baseline |
+| cloth | 5.8 mm | 0.13 | a garment's silhouette is most of what separates a character from the background at range; folds need a defined terminator to read as fabric |
+| hair | 3.4 mm | 0.20 | **the one that matters.** A hair mesh is dozens of near-parallel strands millimetres apart; at the body's 5 mm each strand's hull swallows its neighbour and the head reads as one solid dark cap |
+| accessory | 6.8 mm | 0.10 | small, rigid, usually the most saturated thing on the character — meant to pop |
+
+This required moving the outline from `material_overlay` on the INSTANCE to
+`next_pass` on each surface's material. Miku's body, face and hair are three
+surfaces of one mesh, so an instance-wide overlay can only ever give all three
+the same weight.
 
 Taila's eyes still render as black cards rather than amber irises. Her eye
 surfaces are untextured, and the glTF import hands every untextured surface a

.claude/skills/character-pipeline/references/verification.md

@@ -41,7 +41,36 @@ Two related traps:
 | `travel_dir_check.gd` | stride direction vs. travel direction | < 10° except a capped sidestep |
 | `limb_deform_check.gd` | joint collapse | knee ~0.99 |
 | `verify_character.py` | meshes, bones, weights of a SOURCE model | several meshes, cloth bones present |
+| `surface_class_check.gd` | every surface resolves from the sidecar, not the fallback | 0 fallbacks on all six skins |
+| `character_picker_check.gd` | the escape-menu roster: skeleton, clips, surfaces, and that the pose MOVES | 0 failures |
+| `rig_anchor_check.gd` | a grip anchor physically moves the weapon, and clears | 0 failures |
 | `anim_capture.gd` / `orbit_capture.gd` | renders, for looking | — |
+| `roster_capture.gd` | one photo of every character, from the picker | — |
+| `ui_capture.gd` | one photo of every menu screen | — |
+
+## Assert the consequence, not the plumbing
+
+Three of these exist because the obvious check passes on a broken system.
+
+- `character_picker_check` asserts the skeleton's pose CHANGES over a dozen
+  frames. Asking the model which clip it is playing does not work: that is a
+  variable the class sets on itself, and it reads `"Idle"` just as happily when
+  the animation tree is not ticking at all.
+- `rig_anchor_check` asserts the weapon MOVES by the offset asked for. An anchor
+  system is easy to build so that the sliders move, the file saves and the JSON
+  round-trips while the gun does not budge — the value read into a variable
+  nobody consumed. It measures in the attachment's frame, not the world's:
+  the attachment tracks a bone on an animating skeleton, so a world-space delta
+  is mostly the idle animation.
+- `surface_class_check` FAILS on a surface that falls through to the heuristic
+  instead of resolving from the table. A model whose names stopped matching still
+  renders — the fallback catches it — and quietly loses its per-class art
+  direction. Nothing else would report that.
+
+And four of the last five real defects came from LOOKING, not from asserting:
+a preview showing the back of the character's head, a turntable that carried on
+from the previous character, an unstyled list, and momo's idle pose. Every one
+passed every assertion. Run `roster_capture` and `ui_capture` and open the PNGs.
 
 Run them: