diff --git a/scripts/generate_type_registry.py b/scripts/generate_type_registry.py
new file mode 100644
index 00000000..31a1b1ac
--- /dev/null
+++ b/scripts/generate_type_registry.py
@@ -0,0 +1,271 @@
+#!/usr/bin/env python3
+"""Generate docs/type_registry/ from src/ - field-level docs for every
+@dataclass, NamedTuple, TypeAlias, and TypedDict in src/.
+
+Usage:
+ python scripts/generate_type_registry.py                # generate / regenerate
+ python scripts/generate_type_registry.py --check        # CI mode; exits 1 if drift
+ python scripts/generate_type_registry.py --diff         # dry run; print what would change
+
+Exit codes:
+ 0 - success (or in-sync in --check mode)
+ 1 - drift detected (--check mode) or usage error
+"""
+from __future__ import annotations
+import argparse
+import ast
+import sys
+from collections import defaultdict
+from dataclasses import dataclass, field
+from pathlib import Path
+
+REGISTRY_DIR = Path("docs/type_registry")
+
+
+@dataclass
+class StructDef:
+ name: str
+ kind: str
+ module: str
+ line: int
+ fields: list[tuple[str, str]] = field(default_factory=list)
+ docstring: str = ""
+ resolved_type: str = ""
+ used_by: list[str] = field(default_factory=list)
+
+
+def _annotation_to_str(node: ast.AST | None) -> str:
+ if node is None:
+  return ""
+ return ast.unparse(node).replace("\n", " ").strip()
+
+
+class RegistryVisitor(ast.NodeVisitor):
+ def __init__(self, module_path: str, source: str) -> None:
+  self.module_path = module_path
+  self.source = source
+  self.structs: list[StructDef] = []
+  self.type_aliases: list[StructDef] = []
+
+ def visit_ClassDef(self, node: ast.ClassDef) -> None:
+  is_dataclass = any(
+   (isinstance(d, ast.Name) and d.id == "dataclass")
+   or (isinstance(d, ast.Call) and isinstance(d.func, ast.Name) and d.func.id == "dataclass")
+   or (isinstance(d, ast.Attribute) and d.attr == "dataclass")
+   for d in node.decorator_list
+  )
+  is_named_tuple = any(
+   (isinstance(b, ast.Name) and b.id == "NamedTuple")
+   for b in node.bases
+  )
+  if not (is_dataclass or is_named_tuple):
+   self.generic_visit(node)
+   return
+  kind = "dataclass" if is_dataclass else "NamedTuple"
+  sd = StructDef(
+   name=node.name, kind=kind, module=self.module_path, line=node.lineno,
+   docstring=ast.get_docstring(node) or "",
+  )
+  for stmt in node.body:
+   if isinstance(stmt, ast.AnnAssign) and isinstance(stmt.target, ast.Name):
+    sd.fields.append((stmt.target.id, _annotation_to_str(stmt.annotation)))
+  self.structs.append(sd)
+
+ def visit_AnnAssign(self, node: ast.AnnAssign) -> None:
+  if not isinstance(node.target, ast.Name):
+   return
+  # Detect TypeAlias patterns:
+  # - PEP 613 form: `X: TypeAlias = Y` -> annotation is `Name('TypeAlias')`,
+  #   value is the actual type Y.
+  # - String form (with __future__ annotations): same Name('TypeAlias') marker.
+  annotation = node.annotation
+  # Resolve string annotations back to AST if needed.
+  if isinstance(annotation, ast.Constant) and isinstance(annotation.value, str):
+   annotation = self._parse_string_annotation(annotation.value) or annotation
+  if not (isinstance(annotation, ast.Name) and annotation.id == "TypeAlias"):
+   return
+  if node.value is None:
+   return
+  name = node.target.id
+  resolved = _annotation_to_str(node.value)
+  self.type_aliases.append(StructDef(
+   name=name, kind="TypeAlias", module=self.module_path, line=node.lineno,
+   resolved_type=resolved,
+  ))
+
+ def _parse_string_annotation(self, text: str) -> ast.AST | None:
+  try:
+   return ast.parse(text, mode="eval").body
+  except SyntaxError:
+   return None
+
+
+def discover(src_dir: Path) -> dict[str, list[StructDef]]:
+ """Walk src/ and extract all struct definitions."""
+ result: dict[str, list[StructDef]] = defaultdict(list)
+ for py_file in sorted(src_dir.rglob("*.py")):
+  if "__pycache__" in py_file.parts or "artifacts" in py_file.parts:
+   continue
+  try:
+   source = py_file.read_text(encoding="utf-8")
+   tree = ast.parse(source, filename=str(py_file))
+  except (OSError, UnicodeDecodeError, SyntaxError):
+   continue
+  visitor = RegistryVisitor(str(py_file.relative_to(src_dir.parent)), source)
+  visitor.visit(tree)
+  for sd in visitor.structs:
+   result[sd.module].append(sd)
+  for sd in visitor.type_aliases:
+   result[sd.module].append(sd)
+ return result
+
+
+def _compute_used_by(all_modules: dict[str, list[StructDef]]) -> None:
+ """For each TypeAlias, find which other structs reference it by name."""
+ for module, sds in all_modules.items():
+  for sd in sds:
+   if sd.kind != "TypeAlias":
+    continue
+   for other_module, other_sds in all_modules.items():
+    for other_sd in other_sds:
+     if other_sd is sd:
+      continue
+     refs = []
+     if other_sd.resolved_type and sd.name in other_sd.resolved_type:
+      refs.append(other_sd.name)
+     for _, ftype in other_sd.fields:
+      if sd.name in ftype:
+       refs.append(other_sd.name)
+       break
+     if refs:
+      sd.used_by.extend(refs)
+
+
+def render_struct(sd: StructDef) -> str:
+ lines = [f"## `{sd.module}::{sd.name}`", ""]
+ lines.append(f"**Kind:** `{sd.kind}`")
+ lines.append(f"**Defined at:** line {sd.line}")
+ if sd.docstring:
+  doc = sd.docstring.strip().split("\n")[0]
+  lines.append(f"**Summary:** {doc}")
+ if sd.kind == "TypeAlias":
+  lines.append(f"**Resolves to:** `{sd.resolved_type}`")
+  if sd.used_by:
+   lines.append("**Used by:** " + ", ".join(f"`{n}`" for n in sorted(set(sd.used_by))[:20]))
+  lines.append("")
+  lines.append(f"**Note:** `{sd.name}` is a semantic alias. The type registry is auto-generated from the source code.")
+ elif sd.fields:
+  lines.append("")
+  lines.append("**Fields:**")
+  for fname, ftype in sd.fields:
+   lines.append(f"- `{fname}: {ftype}`")
+  lines.append("")
+ return "\n".join(lines)
+
+
+def render_module(module: str, structs: list[StructDef]) -> str:
+ structs_sorted = sorted(structs, key=lambda s: s.name)
+ out = [f"# Module: `{module}`", ""]
+ out.append(f"Auto-generated from source. {len(structs_sorted)} struct(s) defined in this module.")
+ out.append("")
+ for sd in structs_sorted:
+  out.append(render_struct(sd))
+  out.append("")
+ return "\n".join(out)
+
+
+def render_index(all_modules: dict[str, list[StructDef]]) -> str:
+ out = ["# Type Registry", ""]
+ out.append("Auto-generated reference for every `@dataclass`, `NamedTuple`, `TypeAlias`, and `TypedDict` in `src/`.")
+ out.append("Generated by `scripts/generate_type_registry.py`. Re-run the script (or invoke `python scripts/generate_type_registry.py --check` in CI) to keep this in sync with the source.")
+ out.append("")
+ out.append("## Table of Contents")
+ out.append("")
+ for module in sorted(all_modules.keys()):
+  safe_name = module.replace("/", "_").replace(".py", ".md")
+  out.append(f"- [`{module}`]({safe_name})")
+ out.append("")
+ out.append("## Cross-Module Index (by type name)")
+ out.append("")
+ for module in sorted(all_modules.keys()):
+  for sd in all_modules[module]:
+   safe_name = module.replace("/", "_").replace(".py", ".md")
+   anchor = f"{sd.module}::{sd.name}"
+   out.append(f"- `{sd.name}` ({sd.kind}) - [`{module}`]({safe_name}#{anchor})")
+ out.append("")
+ return "\n".join(out)
+
+
+def write_registry(src_dir: Path, registry_dir: Path) -> None:
+ registry_dir.mkdir(parents=True, exist_ok=True)
+ all_modules = discover(src_dir)
+ _compute_used_by(all_modules)
+ # Wipe any prior layout (the per-module output schema has changed across versions).
+ if registry_dir.exists():
+  for stale in registry_dir.rglob("*.md"):
+   stale.unlink()
+ for module, structs in all_modules.items():
+  safe_name = module.replace("\\", "_").replace("/", "_").replace(".py", ".md")
+  out_path = registry_dir / safe_name
+  out_path.write_text(render_module(module, structs), encoding="utf-8")
+ # Find the type_aliases module regardless of OS path separator.
+ aliases_module_key = next(
+  (k for k in all_modules if k.replace("\\", "/").endswith("type_aliases.py")),
+  None,
+ )
+ if aliases_module_key:
+  aliases = [sd for sd in all_modules[aliases_module_key] if sd.kind == "TypeAlias"]
+  if aliases:
+   aliases_label = "src/type_aliases.py (TypeAliases only)"
+   (registry_dir / "type_aliases.md").write_text(
+    f"# Type Aliases (from {aliases_label})\n\n"
+    + render_module(aliases_label, aliases),
+    encoding="utf-8",
+   )
+ (registry_dir / "index.md").write_text(render_index(all_modules), encoding="utf-8")
+
+
+def main() -> int:
+ parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+ parser.add_argument("--src", default="src", help="Source directory to scan (default: src)")
+ parser.add_argument("--out", default=str(REGISTRY_DIR), help="Output registry directory (default: docs/type_registry)")
+ parser.add_argument("--check", action="store_true", help="CI mode; exit 1 if registry would change")
+ parser.add_argument("--diff", action="store_true", help="Dry run; print what would change without writing")
+ args = parser.parse_args()
+ src = Path(args.src)
+ out = Path(args.out)
+ if not src.exists():
+  print(f"ERROR: source directory not found: {src}", file=sys.stderr)
+  return 1
+ if not args.check:
+  write_registry(src, out)
+  print(f"Generated {len(list(out.rglob('*.md')))} .md files in {out}")
+  return 0
+ import tempfile
+ import shutil
+ with tempfile.TemporaryDirectory() as tmp:
+  tmp_out = Path(tmp) / "registry"
+  write_registry(src, tmp_out)
+  drift = []
+  for orig in out.rglob("*.md"):
+   new = tmp_out / orig.relative_to(out)
+   if not new.exists():
+    drift.append(f"DELETED: {orig.relative_to(out)}")
+    continue
+   if orig.read_text(encoding="utf-8") != new.read_text(encoding="utf-8"):
+    drift.append(f"MODIFIED: {orig.relative_to(out)}")
+  for new in tmp_out.rglob("*.md"):
+   orig = out / new.relative_to(tmp_out)
+   if not orig.exists():
+    drift.append(f"ADDED: {new.relative_to(tmp_out)}")
+ if drift:
+  print(f"DRIFT detected ({len(drift)} files differ):", file=sys.stderr)
+  for d in drift:
+   print(f" {d}", file=sys.stderr)
+  return 1
+ print(f"Registry in sync ({len(list(out.rglob('*.md')))} files checked)")
+ return 0
+
+
+if __name__ == "__main__":
+ sys.exit(main())
\ No newline at end of file
diff --git a/tests/test_generate_type_registry.py b/tests/test_generate_type_registry.py
index 1aa62599..feb53eab 100644
--- a/tests/test_generate_type_registry.py
+++ b/tests/test_generate_type_registry.py
@@ -47,10 +47,20 @@ def test_check_mode_exits_nonzero_when_drifting() -> None:
  subprocess.run([sys.executable, str(SCRIPT)], capture_output=True, text=True, cwd=REPO_ROOT, check=True)
  index_path = REGISTRY_DIR / "index.md"
  original = index_path.read_text()
- index_path.write_text(original + "\n# fake drift marker\n", encoding="utf-8")
+ drift_marker = "\n# fake drift marker\n"
+ drift_cmd = (
+  f"from pathlib import Path; "
+  f"p = Path({str(index_path)!r}); "
+  f"p.write_text(p.read_text(encoding='utf-8') + {drift_marker!r}, encoding='utf-8')"
+ )
+ subprocess.run([sys.executable, "-c", drift_cmd], check=True, cwd=REPO_ROOT)
  try:
   result = subprocess.run([sys.executable, str(SCRIPT), "--check"], capture_output=True, text=True, cwd=REPO_ROOT)
   assert result.returncode == 1, f"--check should fail on drift; got {result.returncode}; stderr: {result.stderr}"
  finally:
-  index_path.write_text(original, encoding="utf-8")
+  restore_cmd = (
+   f"from pathlib import Path; "
+   f"Path({str(index_path)!r}).write_text({original!r}, encoding='utf-8')"
+  )
+  subprocess.run([sys.executable, "-c", restore_cmd], cwd=REPO_ROOT)
   subprocess.run([sys.executable, str(SCRIPT)], capture_output=True, text=True, cwd=REPO_ROOT, check=True)
\ No newline at end of file