motionloom 2.3.0 → 2.5.0

package/scripts/build-ai-pilot.py

@@ -0,0 +1,504 @@
+#!/usr/bin/env python3
+"""Build a hash-bound, review-only MotionLoom AI scout pilot from four PNG frames.
+
+The builder is intentionally pilot-specific. It copies source bytes into a local
+``.motionloom`` workspace, measures their actual alpha geometry using the same
+deterministic reader as Asset Consistency, and writes the contracts that feed
+Artifact Intake and Runtime Candidate. It never manufactures a human approval,
+artist authority, or runtime-pass claim.
+
+It uses Python's standard library only and runs on Ubuntu, macOS, and Windows.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import importlib.util
+import json
+import shutil
+import sys
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+ASSET_ID = "character.scout-robot.ai-pilot"
+PILOT_ID = "ai-pilot-scout-walk"
+FRAME_ORDER = (
+    ("idle", "idle", "preview"),
+    ("contact-right", "contact_right", "runtime_candidate"),
+    ("passing", "passing", "runtime_candidate"),
+    ("contact-left", "contact_left", "runtime_candidate"),
+)
+PROVIDER_PROFILES = {
+    "internal-imagegen": {
+        "adapter_id": "internal.imagegen",
+        "kind": "internal_skill",
+        "invocation_mode": "agent-mediated",
+        "cost_class": "included",
+        "generator_source": "internal-imagegen",
+        "default_model": "default",
+        "default_task_id": "motionloom-scout-ai-pilot-ingest-2026",
+        "license_source": "Internal ImageGen generation; provider-native receipt was not exported",
+        "reconstruction_source": "agent-mediated ImageGen source",
+        "lab_limit": "Internal ImageGen adapter is scaffold-only",
+    },
+    "chatgpt": {
+        "adapter_id": "openai.chatgpt",
+        "kind": "external_provider",
+        "invocation_mode": "manual",
+        "cost_class": "external",
+        "generator_source": "openai-chatgpt",
+        "default_model": "chatgpt-image-generation",
+        "default_task_id": "motionloom-scout-chatgpt-pilot-2026",
+        "license_source": "User-provided ChatGPT generation; provider-native receipt was not exported",
+        "reconstruction_source": "user-mediated ChatGPT source",
+        "lab_limit": "ChatGPT import adapter is scaffold-only",
+    },
+}
+
+
+def sha256_file(path: Path) -> str:
+    return hashlib.sha256(path.read_bytes()).hexdigest()
+
+
+def sha256_text(value: str) -> str:
+    return hashlib.sha256(value.encode("utf-8")).hexdigest()
+
+
+def write_json(path: Path, value: dict[str, Any]) -> None:
+    path.write_text(json.dumps(value, indent=2, sort_keys=True) + "\n", encoding="utf-8")
+
+
+def load_png_reader() -> Any:
+    source = REPO_ROOT / "scripts" / "asset-consistency.py"
+    spec = importlib.util.spec_from_file_location("motionloom_pilot_png", source)
+    if spec is None or spec.loader is None:
+        raise RuntimeError(f"could not load deterministic PNG reader: {source}")
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[spec.name] = module
+    spec.loader.exec_module(module)
+    return module.PNGImage
+
+
+def parse_urls(values: list[str]) -> dict[str, str]:
+    result: dict[str, str] = {}
+    for value in values:
+        frame_id, separator, url = value.partition("=")
+        if not separator or frame_id not in {item[0] for item in FRAME_ORDER} or not url.startswith(("/", "https://", "http://")):
+            raise ValueError("--frame-url must look like idle=/manus-storage/frame.png")
+        result[frame_id] = url
+    return result
+
+
+def measured_frame(PNGImage: Any, path: Path) -> dict[str, Any]:
+    image = PNGImage(path)
+    supports_alpha = image.color_type in {4, 6} or (image.color_type == 3 and bool(image._transparency))
+    if not supports_alpha:
+        raise ValueError(
+            f"{path.name} is color type {image.color_type} and has no alpha channel; "
+            "do not treat a painted checkerboard as transparency"
+        )
+    total = image.width * image.height
+    alpha_pixels = image.alpha_count(1)
+    if alpha_pixels == total:
+        raise ValueError(f"{path.name} has no transparent padding; an isolated alpha frame is required")
+    bbox = image.alpha_bbox(1)
+    if bbox is None:
+        raise ValueError(f"{path.name} has no visible alpha pixels")
+    return {
+        "width": image.width,
+        "height": image.height,
+        "alpha_pixels": alpha_pixels,
+        "alpha_bbox": bbox,
+        "sha256": sha256_file(path),
+        "bytes": path.stat().st_size,
+    }
+
+
+def require_clean_pilot_geometry(frame_id: str, data: dict[str, Any]) -> None:
+    """Reject canvas-spanning residue for this isolated, single-character pilot.
+
+    Alpha isolation can deliberately remove only edge-connected background. A
+    detached band may therefore remain alpha-opaque without touching an edge.
+    The pilot has a narrow, centred single-character contract, so its measured
+    alpha bounds must retain padding on every side and cannot occupy most of a
+    canvas width. This is a geometry guard, not a human-quality approval.
+    """
+    bbox = data["alpha_bbox"]
+    width, height = int(data["width"]), int(data["height"])
+    minimum_padding = max(1, round(min(width, height) * 0.05))
+    paddings = {
+        "left": int(bbox["x"]),
+        "top": int(bbox["y"]),
+        "right": width - (int(bbox["x"]) + int(bbox["width"])),
+        "bottom": height - (int(bbox["y"]) + int(bbox["height"])),
+    }
+    if min(paddings.values()) < minimum_padding:
+        raise ValueError(
+            f"{frame_id} does not retain {minimum_padding}px clean padding on every edge; "
+            f"measured {paddings}. Reject possible canvas-spanning contamination."
+        )
+    if int(bbox["width"]) / width > 0.6:
+        raise ValueError(
+            f"{frame_id} alpha subject spans {int(bbox['width'])}/{width}px of the canvas width; "
+            "reject possible detached horizontal contamination."
+        )
+
+
+def partial_frame_record(frame_id: str, source: Path, data: dict[str, Any] | None, error: str | None) -> dict[str, Any]:
+    """Describe a rejected source without emitting an ingestable candidate."""
+    record: dict[str, Any] = {
+        "role": frame_id,
+        "source": str(source),
+        "sha256": sha256_file(source),
+        "bytes": source.stat().st_size,
+    }
+    if data is not None:
+        bbox = data["alpha_bbox"]
+        record["canvas"] = [data["width"], data["height"]]
+        record["alpha_bbox"] = [bbox["x"], bbox["y"], bbox["width"], bbox["height"]]
+        record["padding"] = {
+            "left": bbox["x"],
+            "top": bbox["y"],
+            "right": data["width"] - (bbox["x"] + bbox["width"]),
+            "bottom": data["height"] - (bbox["y"] + bbox["height"]),
+        }
+    if error is not None:
+        record["error"] = error
+    return record
+
+
+def write_partial_handoff(
+    output: Path,
+    provider: dict[str, str],
+    provider_model: str,
+    provider_task_id: str,
+    timestamp: str,
+    rejected_frames: list[dict[str, Any]],
+    errors: list[str],
+) -> None:
+    """Write source-bound refusal evidence; never create controls or a candidate."""
+    handoff = {
+        "contract": "motionloom-ai-pilot-partial-handoff",
+        "version": 1,
+        "asset_id": ASSET_ID,
+        "state": "partial",
+        "authority": "ai_generated",
+        "production_approved": False,
+        "created_at": timestamp,
+        "blocked": {
+            "reason": "Source preflight failed. No canvas resampling, transparent padding, metadata edit, intake, candidate, render, or approval was emitted.",
+            "errors": errors,
+            "missing_frames": [entry["role"] for entry in rejected_frames if entry.get("error")],
+            "not_emitted": [
+                "controls.json",
+                "generation receipt",
+                "export manifest",
+                "asset consistency contract",
+                "Artifact Intake report",
+                "runtime candidate",
+                "runtime render evidence",
+                "Dev Lab review candidate",
+            ],
+        },
+        "rejected_source_attempts": [
+            {
+                "provider": provider["generator_source"],
+                "adapter_id": provider["adapter_id"],
+                "model": provider_model,
+                "task_id": provider_task_id,
+                "authority": "ai_generated",
+                "state": "rejected_pre_ingest",
+                "reason": "Measured source bytes failed the pilot preflight; no transform may bypass the gate.",
+                "frames": rejected_frames,
+            }
+        ],
+        "resume_requirements": [
+            "Provide independent source frames that satisfy the measured canvas and clean-padding requirements.",
+            "Do not resize, pad, crop, isolate, or relabel rejected bytes to bypass the source gate.",
+            "Re-run build-ai-pilot and downstream pipeline stages only after preflight succeeds.",
+        ],
+    }
+    write_json(output / "partial-handoff.json", handoff)
+
+
+def frame_entry(frame_id: str, role: str, target: str, file_name: str, data: dict[str, Any], root_relative_output: str) -> dict[str, Any]:
+    return {
+        "path": f"{root_relative_output}/frames/{file_name}",
+        "role": f"walk_{role}",
+        "sha256": data["sha256"],
+        "bytes": data["bytes"],
+        "target": target,
+    }
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(description="Build MotionLoom's review-only scout AI pilot workspace")
+    parser.add_argument("--idle", required=True, type=Path, help="Alpha-isolated idle PNG")
+    parser.add_argument("--contact-right", required=True, type=Path, help="Alpha-isolated contact-right PNG")
+    parser.add_argument("--passing", required=True, type=Path, help="Alpha-isolated passing PNG")
+    parser.add_argument("--contact-left", required=True, type=Path, help="Alpha-isolated contact-left PNG")
+    parser.add_argument("--root", type=Path, default=REPO_ROOT, help="MotionLoom repository root")
+    parser.add_argument(
+        "--output",
+        default=".motionloom/pilots/ai-pilot-scout",
+        help="Safe relative output directory inside --root",
+    )
+    parser.add_argument("--frame-url", action="append", default=[], help="Optional frame URL: idle=/manus-storage/frame.png")
+    parser.add_argument("--generated-at", default=None, help="ISO-8601 evidence timestamp; defaults to build time")
+    parser.add_argument("--provider", choices=sorted(PROVIDER_PROFILES), default="internal-imagegen", help="Truthful source provider profile")
+    parser.add_argument("--provider-model", default=None, help="Provider/model label for provenance; defaults to profile value")
+    parser.add_argument("--provider-task-id", default=None, help="Provider or user-mediated task identifier for provenance")
+    parser.add_argument("--overwrite", action="store_true", help="Replace an existing pilot workspace")
+    args = parser.parse_args(argv)
+
+    root = args.root.resolve()
+    output = (root / args.output).resolve()
+    try:
+        output.relative_to(root)
+    except ValueError as exc:
+        raise SystemExit("--output must stay inside --root") from exc
+    if output.exists():
+        if not args.overwrite:
+            raise SystemExit(f"pilot workspace already exists: {output}; pass --overwrite to replace it")
+        shutil.rmtree(output)
+    urls = parse_urls(args.frame_url)
+    timestamp = args.generated_at or datetime.now(UTC).replace(microsecond=0).isoformat().replace("+00:00", "Z")
+    provider = PROVIDER_PROFILES[args.provider]
+    provider_model = args.provider_model or provider["default_model"]
+    provider_task_id = args.provider_task_id or provider["default_task_id"]
+
+    sources = {
+        "idle": args.idle,
+        "contact-right": args.contact_right,
+        "passing": args.passing,
+        "contact-left": args.contact_left,
+    }
+    for frame_id, source in sources.items():
+        if not source.is_file():
+            raise SystemExit(f"missing {frame_id} source image: {source}")
+
+    output.mkdir(parents=True)
+    frames_dir = output / "frames"
+    frames_dir.mkdir()
+    PNGImage = load_png_reader()
+    frames: dict[str, dict[str, Any]] = {}
+    rejected_frames: list[dict[str, Any]] = []
+    preflight_errors: list[str] = []
+    for frame_id, role, _ in FRAME_ORDER:
+        destination = frames_dir / f"scout-{frame_id}.png"
+        shutil.copy2(sources[frame_id], destination)
+        data: dict[str, Any] | None = None
+        try:
+            data = measured_frame(PNGImage, destination)
+            require_clean_pilot_geometry(frame_id, data)
+        except ValueError as exc:
+            error = str(exc)
+            preflight_errors.append(error)
+            rejected_frames.append(partial_frame_record(frame_id, destination, data, error))
+            continue
+        assert data is not None
+        frames[frame_id] = data
+        rejected_frames.append(partial_frame_record(frame_id, destination, data, None))
+
+    if not preflight_errors:
+        canvas_sizes = {(item["width"], item["height"]) for item in frames.values()}
+        if len(canvas_sizes) != 1:
+            preflight_errors.append(f"all frames must use one canvas; measured {sorted(canvas_sizes)}")
+
+    if preflight_errors:
+        write_partial_handoff(output, provider, provider_model, provider_task_id, timestamp, rejected_frames, preflight_errors)
+        print(json.dumps({"status": "partial", "root": str(root), "output": output.relative_to(root).as_posix(), "asset_id": ASSET_ID, "errors": preflight_errors}, indent=2))
+        return 2
+
+    canvas_sizes = {(item["width"], item["height"]) for item in frames.values()}
+    width, height = next(iter(canvas_sizes))
+    alpha_bboxes = [item["alpha_bbox"] for item in frames.values()]
+    footlines = [bbox["y"] + bbox["height"] - 1 for bbox in alpha_bboxes]
+    pivots = [(bbox["x"] + bbox["width"] / 2, bbox["y"] + bbox["height"] - 1) for bbox in alpha_bboxes]
+    reference_pivot = pivots[0]
+    pivot_tolerance = max(max(abs(x - reference_pivot[0]), abs(y - reference_pivot[1])) for x, y in pivots)
+    footline_tolerance = max(abs(value - footlines[0]) for value in footlines)
+    bbox_tolerance = max(
+        max(abs(item["width"] - alpha_bboxes[0]["width"]), abs(item["height"] - alpha_bboxes[0]["height"]))
+        for item in alpha_bboxes
+    )
+    min_alpha = min(int(item["alpha_pixels"]) for item in frames.values())
+    idle_hash = frames["idle"]["sha256"]
+    prompt_capsule = (
+        f"MotionLoom scout robot AI pilot imported from {provider['generator_source']}. "
+        "Post-hoc ingest capsule only; provider-native prompt and task receipt were not exported. "
+        "The capsule binds reviewed output bytes without asserting a human approval or artist authorship."
+    )
+    prompt_hash = sha256_text(prompt_capsule)
+    motion_hash = sha256_text("walk_pilot_4frame:idle,contact-right,passing,contact-left")
+    style_hash = sha256_text("cream-orange-black-cyan pixel-art scout robot style lock")
+
+    relative_output = output.relative_to(root).as_posix()
+    geometry_frames: list[dict[str, Any]] = []
+    receipt_outputs: list[dict[str, Any]] = []
+    export_outputs: list[dict[str, Any]] = []
+    provenance_files: list[dict[str, Any]] = []
+    for index, (frame_id, role, target) in enumerate(FRAME_ORDER):
+        data = frames[frame_id]
+        bbox = data["alpha_bbox"]
+        pivot = pivots[index]
+        filename = f"scout-{frame_id}.png"
+        geometry_frames.append(
+            {
+                "frame_id": f"scout-{frame_id}",
+                "image": f"frames/{filename}",
+                "rect": {"x": 0, "y": 0, "width": width, "height": height},
+                "alpha_bbox": bbox,
+                "pivot": {"x": pivot[0], "y": pivot[1], "space": "pixels"},
+                "footline_px": footlines[index],
+                "safe_rect": bbox,
+                "bleed_margin_px": 0,
+                "sha256": data["sha256"],
+            }
+        )
+        record = frame_entry(frame_id, role, target, filename, data, relative_output)
+        receipt_outputs.append({key: value for key, value in record.items() if key != "target"})
+        export_outputs.append(record)
+        provenance_files.append(
+            {
+                "path": record["path"],
+                "role": record["role"],
+                "sha256": record["sha256"],
+                "bytes": record["bytes"],
+            }
+        )
+
+    controls = {
+        "schema_version": "0.1",
+        "control_id": "scout-ai-pilot-controls-v1",
+        "asset_id": ASSET_ID,
+        "created_at": timestamp,
+        "references": [{"id": "scout-idle-master", "path": f"{relative_output}/frames/scout-idle.png", "sha256": idle_hash, "role": "identity"}],
+        "tracks": [
+            {"id": "identity-lock", "kind": "identity", "binding": "required", "value_hash": idle_hash, "samples": 1},
+            {"id": "style-lock", "kind": "style", "binding": "advisory", "value_hash": style_hash, "samples": 4},
+            {"id": "pose-motion", "kind": "pose", "binding": "required", "value_hash": motion_hash, "samples": 4},
+        ],
+        "output_profile": {"kind": "frame_sequence", "fps": 8, "expected_frame_count": 4, "loop": True},
+    }
+    provenance = {
+        "schema_version": "1.0",
+        "provenance_id": "scout-ai-pilot-provenance-v1",
+        "task_id": provider_task_id,
+        "scene": "scout-walk",
+        "created_at": timestamp,
+        "asset": {"id": ASSET_ID, "path": f"{relative_output}/frames/scout-idle.png", "type": "frame_sequence", "framework": "canvas", "version": "pilot-v1"},
+        "authority": "ai_generated",
+        "readiness": "runtime_ready",
+        "generator": {"model": provider_model, "task_id": provider_task_id, "source": provider["generator_source"], "generated_at": timestamp, "prompt_hash": prompt_hash, "agent": "MotionLoom"},
+        "license": {"spdx": "LicenseRef-Generated-Pilot", "source": provider["license_source"], "attribution": "AI-generated scout pilot; human review remains required"},
+        "files": provenance_files,
+        "provenance_chain": [{"step": "ingest", "actor": "MotionLoom", "source": f"post-hoc hash-bound local ingest from {provider['reconstruction_source']}; not a provider-native receipt", "timestamp": timestamp}],
+        "runtime_evidence": {"status": "not_run", "runtime": "canvas", "tested_at": timestamp},
+    }
+    receipt = {
+        "schema_version": "0.1",
+        "receipt_id": "scout-ai-pilot-receipt-v1",
+        "created_at": timestamp,
+        "asset": {"id": ASSET_ID, "kind": "frame_sequence", "intended_use": "runtime_candidate"},
+        "authority": "ai_generated",
+        "provider": {"adapter_id": provider["adapter_id"], "kind": provider["kind"], "invocation_mode": provider["invocation_mode"], "task_id": provider_task_id, "model": provider_model, "generated_at": timestamp, "cost_class": provider["cost_class"], "prompt_hash": prompt_hash},
+        "control_track_ref": f"{relative_output}/controls.json",
+        "provenance_ref": f"{relative_output}/provenance.json",
+        "outputs": receipt_outputs,
+    }
+    export = {
+        "schema_version": "0.1",
+        "manifest_id": "scout-ai-pilot-export-v1",
+        "asset_id": ASSET_ID,
+        "created_at": timestamp,
+        "receipt_ref": f"{relative_output}/receipt.json",
+        "control_track_ref": f"{relative_output}/controls.json",
+        "outputs": export_outputs,
+    }
+    identity = {
+        "schema_version": "0.1",
+        "asset_id": ASSET_ID,
+        "asset_kind": "character",
+        "identity": {
+            "subject_id": "scout-robot",
+            "reference_hashes": [idle_hash],
+            "camera": {"projection": "orthographic", "width": width, "height": height, "focal_length": 1},
+            "coordinate_system": "screen_y_down",
+            "scale": 1,
+            "pivot": {"x": reference_pivot[0] / width, "y": reference_pivot[1] / height, "space": "normalized_canvas"},
+            "palette_lock": {"colors": ["#f7e4c6", "#f47c15", "#121116", "#18d8f8"], "delta_e_tolerance": 28},
+            "style_profile": "pixel-art scout robot pilot",
+            "lighting_profile": "flat game sprite lighting",
+        },
+        "derivation": {"origin": "ai_generated", "generator": {"model": provider_model, "task_id": provider_task_id, "prompt_hash": prompt_hash}, "source_refs": ["provenance.json", "receipt-reconstruction.json"]},
+        "provenance_ref": "provenance.json",
+        "notes": "Identity controls are a post-hoc ingest representation. They bind output hashes but do not prove provider-native generation controls.",
+    }
+    action_set = {
+        "schema_version": "0.1",
+        "asset_identity": ASSET_ID,
+        "actions": [{"action_id": "walk-pilot", "fps": 8, "frames": [f"frames/scout-{frame_id}.png" for frame_id, _, _ in FRAME_ORDER], "loop": True, "events": ["foot_contact_right", "passing", "foot_contact_left"]}],
+        "invariants": {"canvas_width": width, "canvas_height": height, "pivot_tolerance_px": pivot_tolerance, "footline_tolerance_px": footline_tolerance, "bbox_drift_tolerance_px": bbox_tolerance, "min_alpha_pixels": min_alpha},
+    }
+    frame_geometry = {
+        "schema_version": "0.1",
+        "asset_identity": ASSET_ID,
+        "canvas": {"width": width, "height": height, "color_space": "srgb", "alpha_mode": "straight"},
+        "invariants": {"pivot_tolerance_px": pivot_tolerance, "footline_tolerance_px": footline_tolerance, "bbox_drift_tolerance_px": bbox_tolerance, "min_alpha_pixels": min_alpha, "allow_external_opaque_pixels": False},
+        "frames": geometry_frames,
+    }
+    candidate = {
+        "schema_version": "0.1",
+        "candidate_id": "scout-ai-pilot-candidate-v1",
+        "asset_id": ASSET_ID,
+        "created_at": timestamp,
+        "artifact_intake": {"registry": "artifact-adapter-registry.json", "receipt": f"{relative_output}/receipt.json", "controls": f"{relative_output}/controls.json", "export_manifest": f"{relative_output}/export.json"},
+        "consistency": {"asset_identity": f"{relative_output}/asset-identity.json", "action_set": f"{relative_output}/action-set.json", "frame_geometry": f"{relative_output}/frame-geometry.json"},
+        "runtime": {"target": "canvas", "review_required": True},
+    }
+    reconstruction = {
+        "artifact": ASSET_ID,
+        "status": "incomplete_provider_receipt",
+        "created_at": timestamp,
+        "known": ["output bytes", "output SHA-256", provider["reconstruction_source"], "post-hoc control capsule hash"],
+        "unknown": ["provider-native request ID", "provider-native prompt record", "provider-native seed", "provider-native output receipt"],
+        "governance": {"authority": "ai_generated", "human_review": "required", "production_approved": False, "attestation_approval": False},
+    }
+    lab_evidence = {
+        "pilot_id": PILOT_ID,
+        "asset_id": ASSET_ID,
+        "created_at": timestamp,
+        "state": "review_required",
+        "authority": "ai_generated",
+        "production_approved": False,
+        "runtime_verified": False,
+        "limits": [provider["lab_limit"], "Provider-native generation receipt was unavailable", "Runtime proof is not yet present", "No human approval exists"],
+        "frames": [
+            {"id": frame_id, "label": role.replace("_", " "), "path": f"frames/scout-{frame_id}.png", "url": urls.get(frame_id), "sha256": frames[frame_id]["sha256"], "alpha_bbox": frames[frame_id]["alpha_bbox"]}
+            for frame_id, role, _ in FRAME_ORDER
+        ],
+    }
+    for filename, document in (
+        ("controls.json", controls),
+        ("provenance.json", provenance),
+        ("receipt.json", receipt),
+        ("export.json", export),
+        ("asset-identity.json", identity),
+        ("action-set.json", action_set),
+        ("frame-geometry.json", frame_geometry),
+        ("candidate.json", candidate),
+        ("receipt-reconstruction.json", reconstruction),
+        ("devlab-pilot-evidence.json", lab_evidence),
+    ):
+        write_json(output / filename, document)
+    print(json.dumps({"status": "built", "root": str(root), "output": relative_output, "asset_id": ASSET_ID, "frame_count": len(frames), "alpha_geometry": {"footline_range": [min(footlines), max(footlines)], "pivot_tolerance_px": pivot_tolerance, "footline_tolerance_px": footline_tolerance, "bbox_drift_tolerance_px": bbox_tolerance}}, indent=2))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/scripts/evidence-verifier.py

@@ -34,14 +34,21 @@ def parse_json(path: Path) -> object:
     return json.loads(path.read_text(encoding="utf-8"))
 
 
-def reject_symlink_components(path: Path) -> None:
-    current = Path(path.anchor) if path.anchor else Path(".")
-    for part in path.parts:
-        if part in (path.anchor, ""):
-            continue
+def reject_symlink_within_root(root: Path, relative: Path, label: str) -> None:
+    """Reject links inside an untrusted root, not symlinked system parents.
+
+    macOS exposes /var as a symlink to /private/var. A fixture rooted in mktemp
+    must be valid when the root itself is a real directory, while an evidence
+    path must never traverse a symlink at or below that root.
+    """
+    root = root.absolute()
+    if root.is_symlink():
+        raise ValueError(f"symlink root is not allowed: {label}")
+    current = root
+    for part in relative.parts:
         current = current / part
         if current.is_symlink():
-            raise ValueError(f"symlink path component is not allowed: {path}")
+            raise ValueError(f"symlink path component is not allowed: {label}")
 
 
 def safe_relative_file(root: Path, relative: object, label: str) -> Path:
@@ -53,7 +60,7 @@ def safe_relative_file(root: Path, relative: object, label: str) -> Path:
     if any(part in ("", ".", "..") for part in candidate.parts):
         raise ValueError(f"{label} contains unsafe path components")
     joined = root / candidate
-    reject_symlink_components(joined)
+    reject_symlink_within_root(root, candidate, label)
     resolved_root = root.resolve()
     resolved = joined.resolve(strict=False)
     if resolved != resolved_root and resolved_root not in resolved.parents:
@@ -81,8 +88,10 @@ def verify(scene_dir: Path, task_dir: Path, runtime_evidence_name: str, max_age_
             raise ValueError("scene directory name is unsafe")
         if not SAFE_NAME.fullmatch(task_dir.name):
             raise ValueError("task directory name is unsafe")
-        reject_symlink_components(scene_dir)
-        reject_symlink_components(task_dir)
+        if scene_dir.is_symlink():
+            raise ValueError("scene directory symlink is not allowed")
+        if task_dir.is_symlink():
+            raise ValueError("task directory symlink is not allowed")
         manifest_path = safe_relative_file(scene_dir, "manifest.json", "manifest.json")
         source_manifest = parse_json(manifest_path)
         if not isinstance(source_manifest, dict):