@@ -131,7 +131,7 @@ jobs:
   # The round trip (not just cu13 -> cu12) is what catches the second half:
   # coming back to cu13 must not silently reuse the cu12 extension.
   #
-  # cuda_core only: cuda_bindings_13 cannot be source-built in its cu12
+  # cuda_core only: cuda-bindings cannot be source-built in its cu12
   # environment at all, for reasons unrelated to stale artifacts.
   build-identity-roundtrip:
     name: "cu13 -> cu12 -> cu13 round trip (linux-64, CPU)"
 
@@ -122,13 +122,13 @@ jobs:
   #
   # Dependency graph (verified in pyproject.toml files):
   #   cuda_pathfinder   -> (no internal deps)
-  #   cuda_bindings_13  -> cuda_pathfinder
-  #   cuda_core         -> cuda_pathfinder, cuda_bindings_13
-  #   cuda_python       -> cuda_pathfinder, cuda_bindings_13, cuda_core (meta package)
+  #   cuda_bindings     -> cuda_pathfinder
+  #   cuda_core         -> cuda_pathfinder, cuda_bindings
+  #   cuda_python       -> cuda_pathfinder, cuda_bindings, cuda_core (meta package)
   #
   # A change to cuda_pathfinder (or shared infra) forces a rebuild of every
-  # downstream module. A change to cuda_bindings_13 forces rebuild of cuda_core.
-  # A change to cuda_core alone skips rebuilding/retesting cuda_bindings_13 and
+  # downstream module. A change to cuda_bindings forces rebuild of cuda_core.
+  # A change to cuda_core alone skips rebuilding/retesting cuda_bindings and
   # cuda_pathfinder, but still retests the downstream cuda-python metapackage.
   # Shared build/orchestration changes run the full pipeline; test-only CI
   # infrastructure runs every test suite without rebuilding package wheels.
 
@@ -19,11 +19,12 @@ subpackage in the `cuda-python` monorepo.
 - **Platform internals**: `cuda/bindings/_internal/` contains
   platform-specific implementation files and support code.
 - **Build backend**: `build_hooks.py` drives extension configuration and
-  Cythonization. Logic shared with `cuda_core` (toolchain selection, the
-  compiler flag set, the Cython cache helpers and the rebuild stamps) lives
-  in `_build_shared.py`. `cuda_bindings_13/_build_shared.py` is canonical;
-  `cuda_core/_build_shared.py` must be either a symlink to it or a byte-for-byte
-  identical copy.
+  Cythonization. Logic shared with `cuda_core` and `cuda_pathfinder`
+  (toolchain selection, the compiler flag set, the Cython cache helpers, the
+  rebuild stamps and the release-tag pinning of README links) lives in
+  `_build_shared.py`. `cuda_bindings_13/_build_shared.py` is canonical;
+  the `cuda_core` and `cuda_pathfinder` copies must each be either a symlink to
+  it or a byte-for-byte identical copy.
 
 ## Generated-source workflow
 
 
@@ -1,10 +1,11 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-"""Build helpers shared by the cuda-bindings and cuda-core PEP 517 backends.
+"""Build helpers shared by the cuda-bindings, cuda-core and cuda-pathfinder builds.
 
 ``cuda_bindings_13/_build_shared.py`` is canonical. ``cuda_core/_build_shared.py``
-is either a symlink to it or a byte-for-byte identical copy. Both backends load
+and ``cuda_pathfinder/_build_shared.py`` are each either a symlink to it or a
+byte-for-byte identical copy. Both backends load
 the helper through their package-local path, keeping ``Path(__file__)``-relative
 state package-local.
 
@@ -30,6 +31,7 @@
 import functools
 import hashlib
 import os
+import re
 import shlex
 import shutil
 import sys
@@ -509,3 +511,29 @@ def record_build_key(stamp, key):
     """
     stamp.parent.mkdir(parents=True, exist_ok=True)
     stamp.write_text(key + "\n", encoding="utf-8")
+
+
+_SOURCE_LINK_RE = re.compile(r"(https://github\.com/NVIDIA/cuda-python/(?:blob|tree))/main/")
+
+
+def make_egg_info_command(tag_prefix):
+    """Return an ``egg_info`` command that pins README source links to the release tag.
+
+    DESCRIPTION.rst links to ``main`` so that it renders correctly in the
+    repository. In a release build, the command pins those links to the tag
+    that produced the wheel (``tag_prefix`` followed by the version), so that
+    they survive later repository changes. Development builds keep ``main``.
+    """
+    from packaging.version import Version
+    from setuptools.command.egg_info import egg_info
+
+    class ReleaseLinksEggInfo(egg_info):
+        def run(self):
+            metadata = self.distribution.metadata
+            version = Version(self.distribution.get_version())
+            if not (version.is_devrelease or version.local):
+                tag = f"{tag_prefix}{version}"
+                metadata.long_description = _SOURCE_LINK_RE.sub(rf"\1/{tag}/", metadata.long_description)
+            super().run()
+
+    return ReleaseLinksEggInfo
@@ -11,7 +11,7 @@
 from setuptools.command.build_ext import build_ext as _build_ext
 from setuptools.command.build_py import build_py as _build_py
 
-from _build_shared import drop_empty_venv_include_dir
+from _build_shared import drop_empty_venv_include_dir, make_egg_info_command
 
 if os.environ.get("PARALLEL_LEVEL") is not None:
     warn(
@@ -65,6 +65,7 @@ def finalize_options(self):
     cmdclass={
         "build_ext": build_ext,
         "build_py": build_py,
+        "egg_info": make_egg_info_command("v"),
     },
     zip_safe=False,
 )
@@ -23,7 +23,7 @@ This file describes `cuda_core`, the high-level Pythonic CUDA subpackage in the
   one `_cpp/<name>.cpp` or as a directory `_cpp/<name>/` whose sources all
   compile into the `_<name>` extension (`_cpp/rt/` for `_rt`).
 - **Build backend**: `build_hooks.py` handles Cython extension setup and build
-  dependency wiring. Logic shared with `cuda_bindings_13` (toolchain selection,
+  dependency wiring. Logic shared with `cuda-bindings` (toolchain selection,
   the compiler flag set, the Cython cache helpers and the rebuild stamps)
   lives in `_build_shared.py`. `cuda_bindings_13/_build_shared.py` is canonical;
   `cuda_core/_build_shared.py` must be either a symlink to it or a byte-for-byte
 
@@ -316,8 +316,7 @@ for a complete capture-and-replay example with a measured speedup.
 Beyond the stream capture shown here, ``cuda.core`` also provides an explicit
 graph interface (the :mod:`cuda.core.graph` module) for building, inspecting, and
 modifying graphs directly, including graphs first produced by capture. The
-project's `graph tests
-<https://github.com/NVIDIA/cuda-python/tree/main/cuda_core/tests/graph>`_ have
+project's :repo-tree:`graph tests <cuda_core/tests/graph>` have
 many usage examples.
 
 
 
@@ -148,6 +148,10 @@ def _html_baseurl():
 copybutton_exclude = ".linenos, .gp"
 
 extlinks = {
+    "repo-tree": (
+        f"https://github.com/NVIDIA/cuda-python/tree/{GITHUB_EXAMPLES_REF}/%s",
+        "%s",
+    ),
     # Historical release notes use this role to link into pre-migration tags.
     "cuda-core-example": (
         f"https://github.com/NVIDIA/cuda-python/blob/{GITHUB_EXAMPLES_REF}/cuda_core/examples/%s",
 
@@ -79,7 +79,7 @@ test = [
     "psutil==7.2.2; sys_platform != 'win32' or python_version < '3.15'",
     "packaging>=24",
     "docutils==0.23",
-    # Required by tests/test_graphics.py; pinned to match cuda_bindings_13.
+    # Required by tests/test_graphics.py; pinned to match cuda-bindings.
     "pyglet==2.1.14",
     "cffi==2.1.1",
 ]
@@ -103,7 +103,7 @@ conflicts = [
 [project.urls]
 homepage = "https://nvidia.github.io/cuda-python/"
 documentation = "https://nvidia.github.io/cuda-python/cuda-core/"
-repository = "https://github.com/NVIDIA/cuda-python/tree/main/cuda_core"
+repository = "https://github.com/NVIDIA/cuda-python"
 issues = "https://github.com/NVIDIA/cuda-python/issues/"
 
 [tool.setuptools.packages.find]
 
@@ -13,7 +13,7 @@
 from setuptools.command.build_ext import build_ext as _build_ext
 from setuptools.command.build_py import build_py as _build_py
 
-from _build_shared import drop_empty_venv_include_dir
+from _build_shared import drop_empty_venv_include_dir, make_egg_info_command
 
 nthreads = int(os.environ.get("CUDA_PYTHON_PARALLEL_LEVEL", os.cpu_count() // 2))
 coverage_mode = bool(int(os.environ.get("CUDA_PYTHON_COVERAGE", "0")))
@@ -160,6 +160,7 @@ def finalize_options(self):
         cmdclass={
             "build_ext": build_ext,
             "build_py": build_py,
+            "egg_info": make_egg_info_command("cuda-core-v"),
         },
         zip_safe=False,
     )
Original file line number	Diff line number	Diff line change
`@@ -131,7 +131,7 @@ jobs:`
`131`	`131`	`# The round trip (not just cu13 -> cu12) is what catches the second half:`
`132`	`132`	`# coming back to cu13 must not silently reuse the cu12 extension.`
`133`	`133`	`#`
`134`		`- # cuda_core only: cuda_bindings_13 cannot be source-built in its cu12`
	`134`	`+ # cuda_core only: cuda-bindings cannot be source-built in its cu12`
`135`	`135`	`# environment at all, for reasons unrelated to stale artifacts.`
`136`	`136`	`build-identity-roundtrip:`
`137`	`137`	`name: "cu13 -> cu12 -> cu13 round trip (linux-64, CPU)"`