cozyclay 2.1.0 → 3.0.0

package/tools/ardy/setup-text-encoder.py

@@ -1,306 +0,0 @@
-#!/usr/bin/env python3
-"""setup-text-encoder.py - provision the ARDY text-encoder models without a
-Hugging Face account, approval wait, or access token.
-
-ARDY's text encoder is LLM2Vec on top of meta-llama/Meta-Llama-3-8B-Instruct.
-The official Meta repository is gated (manual approval + token). This script
-instead downloads a byte-pinned stack from public, ungated repositories:
-
-  base       NousResearch/Meta-Llama-3-8B-Instruct   (public mirror of the
-             official weights; every safetensors shard is verified against a
-             pinned SHA-256 before use)
-  adapters   McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp
-             McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised
-
-and lays them out for ARDY's `TEXT_ENCODERS_DIR` override
-(ardy/model/llm2vec/llm2vec_wrapper.py), rewriting each adapter's
-`base_model_name_or_path` to the local base directory so nothing ever
-resolves back to the gated meta-llama repository at runtime.
-
-Every revision is pinned to a full commit SHA and every LFS payload to a
-SHA-256, so re-runs are reproducible and a tampered or changed upstream file
-fails loudly instead of silently drifting the embeddings (and with them the
-generated motion).
-
-Licenses: the base weights are Meta Llama 3 Community License (the LICENSE
-and USE_POLICY.md files are downloaded alongside the weights and must stay
-there); the McGill adapters are MIT. This script only downloads for local
-use - it does not redistribute anything.
-
-Usage (typically on the ARDY box; stdlib only, python >= 3.8):
-
-  python3 setup-text-encoder.py                  # into ~/cclay-text-encoders
-  python3 setup-text-encoder.py --dest /srv/enc  # custom destination
-  python3 setup-text-encoder.py --verify-only    # re-hash an existing tree
-
-Afterwards start the encoder service with the printed environment:
-
-  TEXT_ENCODERS_DIR=<dest> python scripts/run_text_encoder_server.py
-"""
-
-import argparse
-import hashlib
-import json
-import os
-import sys
-import time
-import urllib.error
-import urllib.request
-
-HF = "https://huggingface.co"
-
-BASE_DIR = "NousResearch/Meta-Llama-3-8B-Instruct"
-MNTP_DIR = "McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp"
-# The mntp *download* lands in a -src dir: merge-text-encoder.py bakes the
-# base+MNTP merge into the runtime MNTP_DIR path (transformers v5 dropped
-# automatic adapter-dir resolution, so the runtime path must be a full model).
-MNTP_SRC_DIR = MNTP_DIR + "-src"
-SUP_DIR = "McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised"
-
-# (local dir, repo id, pinned full commit SHA, files)
-# Each file: (path, size, sha256-or-None). SHA-256 comes from the HF LFS
-# metadata of the pinned revision; small non-LFS files carry only a size and
-# are already immutable because the revision is a commit hash.
-# Pins match nv-tlabs/ardy issue #9's reproduction stack.
-MANIFEST = [
-    (
-        BASE_DIR,
-        "NousResearch/Meta-Llama-3-8B-Instruct",
-        "53346005fb0ef11d3b6a83b12c895cca40156b6c",
-        [
-            ("LICENSE", 7801, None),
-            ("USE_POLICY.md", 4696, None),
-            ("config.json", 654, None),
-            ("generation_config.json", 187, None),
-            ("model-00001-of-00004.safetensors", 4976698672,
-             "d8cf9c4d0dd972e1a2131bfe656235ee98221679711a3beef6d46dadf0f20b5c"),
-            ("model-00002-of-00004.safetensors", 4999802720,
-             "8d4782b4a69ef03845159ce1a15e272aadaaf134dc138d68f616098e8531729c"),
-            ("model-00003-of-00004.safetensors", 4915916176,
-             "3acdd690e65c24f42a24581b8467af98bd3ca357444580f8012aacd2bd607921"),
-            ("model-00004-of-00004.safetensors", 1168138808,
-             "67e9ad31c8c32abf3a55ee7fc7217b3ecb35fd3c74d98a5bd233e0e4d6964f46"),
-            ("model.safetensors.index.json", 23950, None),
-            ("special_tokens_map.json", 73, None),
-            ("tokenizer.json", 9085698, None),
-            ("tokenizer_config.json", 50977, None),
-        ],
-    ),
-    (
-        MNTP_SRC_DIR,
-        "McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp",
-        "31474e395ada192e8ed1586db6be79fb3b70c9c0",
-        [
-            ("README.md", 3315, None),
-            ("adapter_config.json", 794, None),
-            ("adapter_model.safetensors", 167829552,
-             "de9c8736618a13173c6a1623cdef1b75e86c69317f1073ae82cd516ac36a632d"),
-            ("attn_mask_utils.py", 11132, None),
-            ("config.json", 781, None),
-            ("modeling_llama_encoder.py", 2856, None),
-            ("special_tokens_map.json", 335, None),
-            ("tokenizer.json", 9085671, None),
-            ("tokenizer_config.json", 51042, None),
-        ],
-    ),
-    (
-        SUP_DIR,
-        "McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp-supervised",
-        "baa8ebf04a1c2500e61288e7dad65e8ae42601a7",
-        [
-            ("README.md", 62702, None),
-            ("adapter_config.json", 794, None),
-            ("adapter_model.safetensors", 167829552,
-             "53f8f94ebdf396667ba99dd96e78203edae27bbcdbd1cf5f12b611e1d916b225"),
-        ],
-    ),
-]
-
-# adapter_config.json is rewritten after download (base_model_name_or_path ->
-# local base dir), so its byte size legitimately changes; everything else must
-# keep the manifest size forever.
-MUTATED = {"adapter_config.json"}
-
-CHUNK = 1024 * 1024
-
-
-def log(msg):
-    print("setup-text-encoder: %s" % msg, flush=True)
-
-
-def fail(msg):
-    print("setup-text-encoder: ERROR: %s" % msg, file=sys.stderr, flush=True)
-    sys.exit(1)
-
-
-def sha256_file(path):
-    h = hashlib.sha256()
-    with open(path, "rb") as f:
-        while True:
-            block = f.read(CHUNK)
-            if not block:
-                break
-            h.update(block)
-    return h.hexdigest()
-
-
-def file_ok(path, size, sha, verify_hash):
-    """Is the on-disk file already the pinned artifact?"""
-    if not os.path.isfile(path):
-        return False
-    name = os.path.basename(path)
-    if name in MUTATED:
-        return True  # rewritten in place; checked by rewrite_adapter_config
-    if os.path.getsize(path) != size:
-        return False
-    if sha is not None and verify_hash:
-        return sha256_file(path) == sha
-    return True
-
-
-def download(url, dest, size, sha):
-    """Resumable download to dest + '.partial', verified, then atomic rename."""
-    part = dest + ".partial"
-    os.makedirs(os.path.dirname(dest), exist_ok=True)
-    for attempt in range(1, 6):
-        try:
-            have = os.path.getsize(part) if os.path.isfile(part) else 0
-            if have > size:
-                os.remove(part)
-                have = 0
-            if have < size:
-                req = urllib.request.Request(url)
-                req.add_header("User-Agent", "cozyclay-setup-text-encoder/1")
-                if have:
-                    req.add_header("Range", "bytes=%d-" % have)
-                with urllib.request.urlopen(req, timeout=60) as resp:
-                    if have and resp.status != 206:
-                        # Server ignored the Range header; start over.
-                        have = 0
-                        mode = "wb"
-                    else:
-                        mode = "ab" if have else "wb"
-                    done = have
-                    t0 = time.time()
-                    with open(part, mode) as out:
-                        while True:
-                            block = resp.read(CHUNK)
-                            if not block:
-                                break
-                            out.write(block)
-                            done += len(block)
-                            if size > CHUNK and time.time() - t0 > 5:
-                                log("  ... %d / %d MiB" % (done // 2**20, size // 2**20))
-                                t0 = time.time()
-            got = os.path.getsize(part)
-            if got != size:
-                raise IOError("size mismatch: got %d, want %d" % (got, size))
-            if sha is not None:
-                actual = sha256_file(part)
-                if actual != sha:
-                    os.remove(part)
-                    raise IOError("sha256 mismatch: got %s, want %s" % (actual, sha))
-            os.replace(part, dest)
-            return
-        except (urllib.error.URLError, IOError, OSError) as e:
-            if attempt == 5:
-                fail("giving up on %s after 5 attempts: %s" % (url, e))
-            log("  retry %d/5 for %s (%s)" % (attempt, os.path.basename(dest), e))
-            time.sleep(2 * attempt)
-
-
-def rewrite_adapter_config(path, base_path):
-    """Point the adapter at the local base so nothing resolves to gated repos."""
-    with open(path, "r", encoding="utf-8") as f:
-        cfg = json.load(f)
-    if cfg.get("base_model_name_or_path") == base_path:
-        return False
-    cfg["base_model_name_or_path"] = base_path
-    tmp = path + ".tmp"
-    with open(tmp, "w", encoding="utf-8") as f:
-        json.dump(cfg, f, indent=2)
-        f.write("\n")
-    os.replace(tmp, path)
-    return True
-
-
-def main():
-    ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
-    ap.add_argument("--dest", default=os.environ.get(
-        "TEXT_ENCODERS_DIR", os.path.expanduser("~/cclay-text-encoders")),
-        help="target directory (default: $TEXT_ENCODERS_DIR or ~/cclay-text-encoders)")
-    ap.add_argument("--verify-only", action="store_true",
-                    help="verify an existing tree (sizes + SHA-256), download nothing")
-    ap.add_argument("--print-plan", action="store_true",
-                    help="list every pinned file and exit")
-    args = ap.parse_args()
-
-    dest = os.path.abspath(os.path.expanduser(args.dest))
-    os.makedirs(dest, exist_ok=True)
-    dest_real = os.path.realpath(dest)
-
-    def safe_join(*parts):
-        """Join paths and verify the result stays within base."""
-        full = os.path.realpath(os.path.join(dest_real, *parts))
-        if full != dest_real and not full.startswith(dest_real + os.sep):
-            fail("path traversal detected: %s" % os.path.join(*parts))
-        return full
-
-    base_path = safe_join(BASE_DIR)
-
-    if args.print_plan:
-        for local_dir, repo, rev, files in MANIFEST:
-            print("%s @ %s" % (repo, rev))
-            for path, size, sha in files:
-                print("  %-40s %12d  %s" % (path, size, sha or "-"))
-        return
-
-    total = sum(size for _, _, _, files in MANIFEST for _, size, _ in files)
-    missing = []
-    for local_dir, repo, rev, files in MANIFEST:
-        for path, size, sha in files:
-            target = safe_join(local_dir, path)
-            if not file_ok(target, size, sha, verify_hash=args.verify_only):
-                missing.append((local_dir, repo, rev, path, size, sha))
-
-    if args.verify_only:
-        if missing:
-            for local_dir, _, _, path, _, _ in missing:
-                print("  BAD/MISSING  %s/%s" % (local_dir, path))
-            fail("%d file(s) failed verification under %s" % (len(missing), dest))
-        log("all %d files verified OK under %s" % (
-            sum(len(f) for _, _, _, f in MANIFEST), dest))
-        return
-
-    if missing:
-        todo = sum(m[4] for m in missing)
-        log("destination: %s" % dest)
-        log("downloading %d file(s), %.1f GiB of %.1f GiB total pinned payload"
-            % (len(missing), todo / 2**30, total / 2**30))
-        for local_dir, repo, rev, path, size, sha in missing:
-            url = "%s/%s/resolve/%s/%s?download=true" % (HF, repo, rev, path)
-            log("fetch %s/%s (%d bytes)" % (local_dir, path, size))
-            download(url, safe_join(local_dir, path), size, sha)
-    else:
-        log("all pinned files already present under %s" % dest)
-
-    for adapter_dir in (MNTP_SRC_DIR, SUP_DIR):
-        cfg = safe_join(adapter_dir, "adapter_config.json")
-        if rewrite_adapter_config(cfg, base_path):
-            log("repointed %s -> local base" % os.path.join(adapter_dir, "adapter_config.json"))
-
-    log("downloads verified. Bake the MNTP merge next (one-time, ~5 min):")
-    log("  cd <ardy checkout> && <ardy venv python> merge-text-encoder.py --dest %s" % dest)
-    print()
-    print("Start ARDY's text-encoder service (and any local ARDY run) with:")
-    print()
-    print("  export TEXT_ENCODERS_DIR=%s" % dest)
-    print("  python scripts/run_text_encoder_server.py")
-    print()
-    print("The base weights are Meta Llama 3 (Meta Llama 3 Community License;")
-    print("see %s/LICENSE)." % base_path)
-    print("This product is built with Meta Llama 3.")
-
-
-if __name__ == "__main__":
-    main()

package/tools/ardy/smpl-cskel27.mjs

@@ -1,178 +0,0 @@
-/** Rotation-only GVHMR SMPL -> cskel27 retarget.
- * The argument is the member map returned by readNpz; plain arrays are also
- * accepted for small callers. Arrays may be nested or flat typed buffers.
- *
- * PROPORTIONS. Mocap is the product here, so the take carries the filmed
- * performer's body, not the canonical one: every cskel27 bone is scaled by
- * the ratio of the performer's SMPL rest bone length to the canonical bone
- * (`boneScale`, one factor per joint = the length of the bone ENDING at that
- * joint). posedJoints are grown with FK over those scaled offsets, so foot
- * placement, reach and stride are the performer's; playback applies the same
- * factors to the bones it drives by rotation only (the arm chain). A bone the
- * SMPL skeleton does not segment the same way (Spine..Spine3 are four links
- * against SMPL's three; hand ends; thumb) takes the nearest measured chain's
- * ratio so the body stays continuous. rootPos is the performer's pelvis
- * height as filmed: with the legs at their real length there is no height
- * ratio left to apply, and personScale is 1 by construction.
- */
-import { CSKEL27_JOINTS, CSKEL27_PARENTS } from "../../src/ardy/cskel27.js";
-import { deriveBoneOffsets, forwardKinematics, matMul, matTranspose } from "../../src/ardy/convert.js";
-import { canonicalCskel27Reference } from "../../src/ardy/to-cskel27.js";
-
-const SMPL_PARENTS = [-1, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 9, 12, 13, 14, 16, 17, 18, 19, 20, 21];
-const IDENTITY = [[1, 0, 0], [0, 1, 0], [0, 0, 1]];
-const MAP = {
-	Hips: 0, Spine: 3, Spine1: 6, Spine2: 9, Spine3: 9, Neck: 12, Head: 15,
-	RightShoulder: 14, RightArm: 17, RightForeArm: 19, RightHand: 21,
-	LeftShoulder: 13, LeftArm: 16, LeftForeArm: 18, LeftHand: 20,
-	RightUpLeg: 2, RightLeg: 5, RightFoot: 8, RightToeBase: 11,
-	LeftUpLeg: 1, LeftLeg: 4, LeftFoot: 7, LeftToeBase: 10,
-};
-// Limb rest directions are meaningful (SMPL's A-pose versus cskel27's
-// T-pose), so a constant swing aligns those chains without losing source
-// twist. Torso, pelvis, neck and head are deliberately absent: SMPL's joint
-// regressor forms an anatomical S inside the mesh and its head joint sits
-// forward inside the skull. Those are not visible-pose corrections and must
-// not deform the target character's bind shape.
-const CHILD = {
-	RightShoulder: 17, RightArm: 19, RightForeArm: 21, RightHand: 23,
-	LeftShoulder: 16, LeftArm: 18, LeftForeArm: 20, LeftHand: 22,
-	RightUpLeg: 5, RightLeg: 8, RightFoot: 11,
-	LeftUpLeg: 4, LeftLeg: 7, LeftFoot: 10,
-};
-const BORROW = { RightHandEnd: "RightHand", RightHandThumb1: "RightHand", LeftHandEnd: "LeftHand", LeftHandThumb1: "LeftHand" };
-const CORRECTION_BORROW = { RightToeBase: "RightFoot", LeftToeBase: "LeftFoot" };
-// Bone lengths measured on the performer, keyed by the cskel27 joint the bone
-// ENDS at, as [SMPL from, SMPL to] against [cskel27 from, cskel27 to]. Where
-// the two skeletons segment a limb differently the measurement spans the
-// whole segment on both sides and every link in it takes the one ratio:
-//  - torso: SMPL's three spine joints do not sit where cskel27's four do, so
-//    pelvis -> spine3 is one measurement over Spine..Spine3;
-//  - shoulder girdle: SMPL's collar is a short stub (8 cm) where Mixamo's
-//    clavicle is 16 cm, so spine3 -> shoulder covers Shoulder+Arm together
-//    (per-link it read 0.79 / 0.57 and would have pinched the shoulders);
-//  - neck, head, hand ends, thumb: geometry of the character, not a limb
-//    the mocap measures. cskel27's Neck link (25 cm) already lands on the
-//    Mixamo rig at 16 cm — the skin's neck is short by construction — so
-//    scaling it by SMPL's 0.84 sank the neck into the chest (measured:
-//    Spine2->Neck 16.5 -> 12.5 cm on the rig). Left at 1 like the head.
-const LENGTH_SOURCE = {
-	Spine: [0, 9], Spine1: [0, 9], Spine2: [0, 9], Spine3: [0, 9],
-	RightShoulder: [9, 17], RightArm: [9, 17], RightForeArm: [17, 19], RightHand: [19, 21],
-	LeftShoulder: [9, 16], LeftArm: [9, 16], LeftForeArm: [16, 18], LeftHand: [18, 20],
-	RightUpLeg: [0, 2], RightLeg: [2, 5], RightFoot: [5, 8], RightToeBase: [8, 11],
-	LeftUpLeg: [0, 1], LeftLeg: [1, 4], LeftFoot: [4, 7], LeftToeBase: [7, 10],
-};
-const LENGTH_TARGET = {
-	Spine: ["Hips", "Spine3"], Spine1: ["Hips", "Spine3"], Spine2: ["Hips", "Spine3"], Spine3: ["Hips", "Spine3"],
-	RightShoulder: ["Spine3", "RightArm"], RightArm: ["Spine3", "RightArm"],
-	LeftShoulder: ["Spine3", "LeftArm"], LeftArm: ["Spine3", "LeftArm"],
-};
-
-function data(value) {
-	if (value && value.data !== undefined) return value.data;
-	if (Array.isArray(value)) return value.flat(Infinity);
-	return value;
-}
-function shape(value, expected, label) {
-	const d = data(value);
-	if (!d || d.length !== expected.reduce((a, b) => a * b, 1)) throw new Error(`${label} has the wrong size`);
-	return d;
-}
-function vec(a) { const n = Math.hypot(...a); return n > 1e-12 ? a.map((v) => v / n) : [1, 0, 0]; }
-function cross(a, b) { return [a[1] * b[2] - a[2] * b[1], a[2] * b[0] - a[0] * b[2], a[0] * b[1] - a[1] * b[0]]; }
-function dot(a, b) { return a[0] * b[0] + a[1] * b[1] + a[2] * b[2]; }
-function axisAngle(axis, angle) {
-	const [x, y, z] = axis, s = Math.sin(angle), c = Math.cos(angle), t = 1 - c;
-	return [[t * x * x + c, t * x * y - s * z, t * x * z + s * y], [t * x * y + s * z, t * y * y + c, t * y * z - s * x], [t * x * z - s * y, t * y * z + s * x, t * z * z + c]];
-}
-function axisAngleMat(v) {
-	const angle = Math.hypot(...v); return angle < 1e-12 ? IDENTITY.map((r) => r.slice()) : axisAngle(v.map((x) => x / angle), angle);
-}
-function swing(from, to) {
-	const a = vec(from), b = vec(to), c = Math.max(-1, Math.min(1, dot(a, b)));
-	if (c > 1 - 1e-10) return IDENTITY.map((r) => r.slice());
-	let ax = cross(a, b), n = Math.hypot(...ax);
-	if (n < 1e-10) {
-		ax = Math.abs(a[0]) < 0.9 ? cross(a, [1, 0, 0]) : cross(a, [0, 1, 0]); n = Math.hypot(...ax);
-		return axisAngle(ax.map((x) => x / n), Math.PI);
-	}
-	return axisAngle(ax.map((x) => x / n), Math.acos(c));
-}
-function readAa(flat, frame, joint) { const o = (frame * 23 + joint - 1) * 3; return [flat[o], flat[o + 1], flat[o + 2]]; }
-function sourceGlobals(go, pose, frames) {
-	const worlds = new Array(frames);
-	for (let f = 0; f < frames; f += 1) {
-		const out = new Array(24); out[0] = axisAngleMat([go[f * 3], go[f * 3 + 1], go[f * 3 + 2]]);
-		for (let j = 1; j < 24; j += 1) out[j] = matMul(out[SMPL_PARENTS[j]], axisAngleMat(readAa(pose, f, j)));
-		worlds[f] = out;
-	}
-	return worlds;
-}
-function writeMat(out, offset, m) { out[offset] = m[0][0]; out[offset + 1] = m[0][1]; out[offset + 2] = m[0][2]; out[offset + 3] = m[1][0]; out[offset + 4] = m[1][1]; out[offset + 5] = m[1][2]; out[offset + 6] = m[2][0]; out[offset + 7] = m[2][1]; out[offset + 8] = m[2][2]; }
-function percentile(values, p) { const a = [...values].sort((x, y) => x - y); const x = (a.length - 1) * p, i = Math.floor(x), t = x - i; return a[i] * (1 - t) + (a[i + 1] ?? a[i]) * t; }
-
-export function smplToCskel27Motion(members, { boneScale: boneScaleOption = "derive" } = {}) {
-	if (boneScaleOption !== "derive" && !(typeof boneScaleOption === "number" && Number.isFinite(boneScaleOption) && boneScaleOption > 0)) throw new Error('smplToCskel27Motion: boneScale must be "derive" or a positive number');
-	const go = data(members.smpl_global_orient), poseMember = members.smpl_body_pose;
-	const goShape = poseMember?.shape?.[0] ?? members.smpl_global_orient?.shape?.[0];
-	const frames = Number.isInteger(goShape) ? goShape : go.length / 3;
-	if (!Number.isInteger(frames) || frames < 1) throw new Error("smplToCskel27Motion: invalid frame count");
-	shape(go, [frames, 3], "smpl_global_orient");
-	const pose = shape(poseMember, [frames, 23, 3], "smpl_body_pose");
-	const transl = shape(members.smpl_transl, [frames, 3], "smpl_transl");
-	const rest = shape(members.smpl_rest_joints, [24, 3], "smpl_rest_joints");
-	const fpsValue = data(members.fps); const fps = Number(fpsValue?.[0] ?? fpsValue);
-	if (!(fps > 0)) throw new Error("smplToCskel27Motion: invalid fps");
-	const skeleton = canonicalCskel27Reference();
-	const cpos = skeleton.posed_joints;
-	const restVec = (j) => [rest[j * 3], rest[j * 3 + 1], rest[j * 3 + 2]];
-	const dist = (a, b) => Math.hypot(a[0] - b[0], a[1] - b[1], a[2] - b[2]);
-	const boneScale = new Float32Array(27).fill(typeof boneScaleOption === "number" ? boneScaleOption : 1);
-	for (let j = 1; j < 27 && boneScaleOption === "derive"; j += 1) {
-		const name = CSKEL27_JOINTS[j];
-		if (!LENGTH_SOURCE[name]) continue;
-		const [s0, s1] = LENGTH_SOURCE[name];
-		const source = dist(restVec(s0), restVec(s1));
-		const [t0, t1] = LENGTH_TARGET[name] ?? [CSKEL27_JOINTS[CSKEL27_PARENTS[j]], name];
-		const target = dist(cpos[CSKEL27_JOINTS.indexOf(t0)], cpos[CSKEL27_JOINTS.indexOf(t1)]);
-		if (source > 1e-6 && target > 1e-6) boneScale[j] = source / target;
-	}
-	const offsets = deriveBoneOffsets(cpos, skeleton.local_rot_mats).map((o, j) => [o[0] * boneScale[j], o[1] * boneScale[j], o[2] * boneScale[j]]);
-	const corrections = {};
-	for (const [name, sourceChild] of Object.entries(CHILD)) {
-		const c = CSKEL27_JOINTS.indexOf(name);
-		const targetChild = CSKEL27_PARENTS.findIndex((parent) => parent === c);
-		if (c < 0 || targetChild < 0) continue;
-		const source = MAP[name];
-		const a = [cpos[targetChild][0] - cpos[c][0], cpos[targetChild][1] - cpos[c][1], cpos[targetChild][2] - cpos[c][2]];
-		const b = [rest[sourceChild * 3] - rest[source * 3], rest[sourceChild * 3 + 1] - rest[source * 3 + 1], rest[sourceChild * 3 + 2] - rest[source * 3 + 2]];
-		corrections[name] = swing(a, b);
-	}
-	for (const [name, ancestor] of Object.entries(CORRECTION_BORROW)) corrections[name] = corrections[ancestor] ?? IDENTITY;
-	const worlds = sourceGlobals(go, pose, frames), rotMats = new Float32Array(frames * 27 * 9), rootPos = new Float32Array(frames * 3), posedJoints = new Float32Array(frames * 27 * 3);
-	// The legs now have the performer's length, so the pelvis rides at its
-	// filmed height: no ratio between the two skeletons is left to apply.
-	const rootScale = 1;
-	for (let f = 0; f < frames; f += 1) {
-		const globals = new Array(27);
-		for (let j = 0; j < 27; j += 1) {
-			const name = CSKEL27_JOINTS[j], ancestor = BORROW[name];
-			// body_pose is already a rotation FROM SMPL's zero pose. Torso/head
-			// therefore keep cskel27's bind shape; only visible limb rest-pose
-			// differences use the constant correction built above.
-			globals[j] = ancestor ? globals[CSKEL27_JOINTS.indexOf(ancestor)] : matMul(worlds[f][MAP[name]], corrections[name] ?? IDENTITY);
-		}
-		for (let j = 0; j < 27; j += 1) {
-			const p = CSKEL27_PARENTS[j], local = p === null ? globals[j] : matMul(matTranspose(globals[p]), globals[j]);
-			writeMat(rotMats, (f * 27 + j) * 9, local);
-		}
-		rootPos[f * 3] = transl[f * 3] * rootScale; rootPos[f * 3 + 1] = transl[f * 3 + 1] * rootScale; rootPos[f * 3 + 2] = transl[f * 3 + 2] * rootScale;
-	}
-	const readLocals = (f) => Array.from({ length: 27 }, (_, j) => { const o = (f * 27 + j) * 9; return [[rotMats[o], rotMats[o + 1], rotMats[o + 2]], [rotMats[o + 3], rotMats[o + 4], rotMats[o + 5]], [rotMats[o + 6], rotMats[o + 7], rotMats[o + 8]]]; });
-	const lows = new Array(frames);
-	for (let f = 0; f < frames; f += 1) { const p = forwardKinematics(readLocals(f), offsets, [rootPos[f * 3], rootPos[f * 3 + 1], rootPos[f * 3 + 2]]); lows[f] = Math.min(...[21, 22, 25, 26].map((j) => p[j][1])); }
-	const floor = percentile(lows, 0.1);
-	for (let f = 0; f < frames; f += 1) { rootPos[f * 3 + 1] -= floor; const p = forwardKinematics(readLocals(f), offsets, [rootPos[f * 3], rootPos[f * 3 + 1], rootPos[f * 3 + 2]]); for (let j = 0; j < 27; j += 1) posedJoints.set(p[j], (f * 27 + j) * 3); }
-	return { frames, fps, rotMats, rootPos, posedJoints, boneScale, personScale: 1, rawRootStart: [transl[0], transl[1], transl[2]] };
-}

package/tools/ardy/visual-qa.mjs

@@ -1,196 +0,0 @@
-/**
- * Visual QA driver for CozyClay: drives headless Chrome over CDP with
- * Node's built-in WebSocket. Loads the studio, generates "a person walk"
- * with the default seed (2), then captures screenshots at several playback
- * frames for pose review.
- *
- * Usage: node tools/ardy/visual-qa.mjs <outDir> [url]
- * The default URL is the studio route (`/app/`), not the marketing landing page.
- */
-import { writeFileSync, mkdirSync } from "node:fs";
-
-const OUT = process.argv[2] || "/tmp/visual-qa";
-const URL = process.argv[3] || "http://127.0.0.1:5180/app/";
-mkdirSync(OUT, { recursive: true });
-
-const sleep = (ms) => new Promise((r) => setTimeout(r, ms));
-
-// --- minimal CDP client -----------------------------------------------------
-async function newTab() {
-	const res = await fetch("http://127.0.0.1:9222/json/new?about:blank", { method: "PUT" });
-	const target = await res.json();
-	const ws = new WebSocket(target.webSocketDebuggerUrl);
-	await new Promise((resolve, reject) => {
-		ws.onopen = resolve;
-		ws.onerror = reject;
-	});
-	let id = 0;
-	const pending = new Map();
-	const consoleLog = [];
-	ws.onmessage = (event) => {
-		const msg = JSON.parse(event.data);
-		if (msg.id && pending.has(msg.id)) {
-			const { resolve, reject } = pending.get(msg.id);
-			pending.delete(msg.id);
-			if (msg.error) reject(new Error(msg.error.message));
-			else resolve(msg.result);
-		} else if (msg.method === "Runtime.consoleAPICalled") {
-			const type = msg.params.type;
-			const text = msg.params.args.map((a) => a.value ?? a.description ?? "").join(" ");
-			if (type === "error" || type === "warning") consoleLog.push(`[${type}] ${text}`);
-		} else if (msg.method === "Runtime.exceptionThrown") {
-			consoleLog.push(`[exception] ${msg.params.exceptionDetails.text} ${msg.params.exceptionDetails.exception?.description ?? ""}`);
-		}
-	};
-	const send = (method, params = {}) =>
-		new Promise((resolve, reject) => {
-			const msgId = ++id;
-			pending.set(msgId, { resolve, reject });
-			ws.send(JSON.stringify({ id: msgId, method, params }));
-		});
-	return { ws, send, consoleLog };
-}
-
-async function evaluate(tab, expression) {
-	const res = await tab.send("Runtime.evaluate", {
-		expression,
-		awaitPromise: true,
-		returnByValue: true,
-	});
-	if (res.exceptionDetails) {
-		throw new Error(`evaluate failed: ${JSON.stringify(res.exceptionDetails.exception?.description || res.exceptionDetails.text)}`);
-	}
-	return res.result.value;
-}
-
-async function shot(tab, name) {
-	const res = await tab.send("Page.captureScreenshot", { format: "png" });
-	writeFileSync(`${OUT}/${name}.png`, Buffer.from(res.data, "base64"));
-	console.log(`saved ${OUT}/${name}.png`);
-}
-
-// --- flow -------------------------------------------------------------------
-const tab = await newTab();
-await tab.send("Page.enable");
-await tab.send("Runtime.enable");
-await tab.send("Page.navigate", { url: URL });
-await sleep(6000);
-
-// Prompt Blocks replaced the old single "Generate motion" field. Select the
-// first character, open the Prompt Blocks foldout, then add a block before
-// looking for its prompt input and the top-bar Generate Motion action.
-await evaluate(tab, `(() => {
-	const character = [...document.querySelectorAll('.hierarchy-row')].find((row) => /Character 1|Subject 1|인물 1/.test(row.textContent));
-	character?.click();
-	const foldout = [...document.querySelectorAll('.foldout-head')].find((button) => /Prompt Blocks|프롬프트 블록/.test(button.textContent));
-	foldout?.click();
-	const add = document.querySelector('.tl-track.prompts .tl-track-add') || document.querySelector('.tl-track-add[title*="prompt"]');
-	add?.click();
-	return true;
-})()`);
-await sleep(400);
-const state0 = await evaluate(tab, `(() => {
-	const seed = document.querySelector('input[placeholder="empty = random"]');
-	const promptEl = document.querySelector('input[placeholder*="motion block"], input[placeholder*="모션 블록"]');
-	const gen = document.querySelector('[data-testid=topbar-generate]');
-	return { seedValue: seed?.value ?? null, hasGenerate: !!gen, generateDisabled: gen?.getAttribute('aria-disabled') === 'true', hasPrompt: !!promptEl };
-})()`);
-console.log("initial:", JSON.stringify(state0));
-
-if (!state0.hasGenerate || !state0.hasPrompt) {
-	await shot(tab, "00-no-ui");
-	throw new Error("studio UI not ready");
-}
-if (state0.seedValue === null) throw new Error("Prompt Blocks panel did not open");
-
-// Set the prompt and generate: focus the field, then type via the Input
-// domain so React's controlled input sees real key events.
-await evaluate(tab, `(() => {
-	const el = document.querySelector('input[placeholder*="motion block"], input[placeholder*="모션 블록"]');
-	if (!el) return false;
-	el.focus();
-	el.select?.();
-	return true;
-})()`);
-for (const key of ["a", " ", "p", "e", "r", "s", "o", "n", " ", "w", "a", "l", "k"]) {
-	await tab.send("Input.dispatchKeyEvent", { type: "keyDown", text: key });
-	await tab.send("Input.dispatchKeyEvent", { type: "keyUp", text: key });
-}
-await sleep(300);
-await shot(tab, "01-before-generate");
-const clicked = await evaluate(tab, `(() => {
-const gen = document.querySelector('[data-testid=topbar-generate]');
-	if (!gen || gen.getAttribute('aria-disabled') === 'true') return false;
-	gen.click();
-	return true;
-})()`);
-if (!clicked) throw new Error("Kimodo bridge is unavailable or generation is disabled");
-console.log("generate clicked, waiting for the clip...");
-
-// Wait for "Motion loaded" or the PLAYBACK badge (generation ~12 s).
-let loaded = false;
-for (let i = 0; i < 40; i += 1) {
-	await sleep(1000);
-	const st = await evaluate(tab, `document.body.textContent.includes("Motion loaded") || document.body.textContent.includes("PLAYBACK")`);
-	if (st) { loaded = true; break; }
-}
-console.log("motion loaded:", loaded);
-if (!loaded) {
-	await shot(tab, "02-generation-timeout");
-	throw new Error("motion did not load within 40 s");
-}
-await sleep(1500);
-
-// Scrub to specific frames by clicking the timeline's accessible ruler. The
-// old tick-text selector clicked labels, not the scrub surface, after the
-// timeline was redesigned.
-const frames = [0, 20, 40, 60, 79];
-for (const f of frames) {
-	const point = await evaluate(tab, `(() => {
-		const lane = document.querySelector('[role="slider"][aria-label*="Scrub timeline"], [role="slider"][aria-label*="타임라인"]');
-		if (!lane) return null;
-		const rect = lane.getBoundingClientRect();
-		const max = Math.max(1, Number(lane.getAttribute('aria-valuemax')) || 1);
-		return { x: rect.left + rect.width * Math.min(1, ${f} / max), y: rect.top + rect.height / 2 };
-	})()`);
-	if (!point) throw new Error("timeline scrub ruler not found");
-	await tab.send("Input.dispatchMouseEvent", { type: "mouseMoved", x: point.x, y: point.y, button: "none" });
-	await tab.send("Input.dispatchMouseEvent", { type: "mousePressed", x: point.x, y: point.y, button: "left", clickCount: 1 });
-	await tab.send("Input.dispatchMouseEvent", { type: "mouseReleased", x: point.x, y: point.y, button: "left", clickCount: 1 });
-	await sleep(700);
-	await shot(tab, `10-frame-${String(f).padStart(2, "0")}`);
-}
-
-// Play a couple of seconds for a mid-stride live shot.
-await evaluate(tab, `(() => {
-	const play = [...document.querySelectorAll("button")].find(b => b.textContent.trim() === "▶");
-	play?.click();
-	return true;
-})()`);
-await sleep(1600);
-await shot(tab, "20-playing-midstride");
-await sleep(1600);
-await shot(tab, "21-playing-later");
-await evaluate(tab, `(() => {
-	const pause = [...document.querySelectorAll("button")].find(b => /❚❚|⏸/.test(b.textContent.trim()));
-	pause?.click();
-	return true;
-})()`);
-
-// Camera close-up on the subject for a detailed pose check.
-await evaluate(tab, `(() => {
-	const chip = [...document.querySelectorAll("button")].find(b => /close-up/i.test(b.textContent));
-	chip?.click();
-	return true;
-})()`);
-await sleep(900);
-await shot(tab, "30-closeup");
-
-if (tab.consoleLog.length) {
-	console.log("--- console errors/warnings ---");
-	for (const line of [...new Set(tab.consoleLog)].slice(0, 15)) console.log(line);
-} else {
-	console.log("no console errors/warnings captured");
-}
-tab.ws.close();
-console.log("visual QA capture complete");