v2.34.1: fix macOS wheel delocate repair (headerpad)

CHANGELOG.md

@@ -5,6 +5,16 @@ All notable changes to TurboLoader will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [2.34.1] - 2026-07-11
+
+### Fixed
+- **macOS wheels failed to build for 2.34.0** and never reached PyPI (Linux wheels +
+  sdist were unaffected): the video path's three extra `-framework` load commands
+  exhausted the Mach-O header slack that delocate needs to rewrite bundled-dylib
+  install names during wheel repair ("larger updated load commands do not fit").
+  macOS links now pass `-Wl,-headerpad_max_install_names`. If you are on macOS,
+  use 2.34.1 — 2.34.0 has no mac wheels.
+
 ## [2.34.0] - 2026-07-11
 
 GPU-everything sprint: resident (pre-processed) epochs and hardware video decode on

docs/GPU_ACCELERATION.md

@@ -115,9 +115,9 @@ consumer stressing the output ring) — every async batch matched a distinct ref
 ### Build (on a real CUDA box; needs `nvcc` + toolkit, gcc 10+ for C++20)
 
 > **Prebuilt CUDA wheel (no compile):** Linux x86_64 / Python 3.10 / CUDA 13.x runtime —
-> attached to the [GitHub Release](https://github.com/ALJainProjects/TurboLoader/releases/tag/v2.33.0):
+> attached to the [GitHub Release](https://github.com/ALJainProjects/TurboLoader/releases/tag/v2.34.1):
 > ```bash
-> pip install https://github.com/ALJainProjects/TurboLoader/releases/download/v2.33.0/turboloader-2.33.0+cu13-cp310-cp310-linux_x86_64.whl
+> pip install https://github.com/ALJainProjects/TurboLoader/releases/download/v2.34.1/turboloader-2.34.1+cu13-cp310-cp310-linux_x86_64.whl
 > pip install nvidia-nvimgcodec-cu12
 > ```
 > Other Pythons/CUDA versions: build from source below.

docs/installation.md

@@ -174,9 +174,9 @@ work without PyTorch installed.
 ### GPU image loader (NVIDIA CUDA) — build from source
 
 > **Prebuilt CUDA wheel (no compile):** Linux x86_64 / Python 3.10 / CUDA 13.x runtime —
-> attached to the [GitHub Release](https://github.com/ALJainProjects/TurboLoader/releases/tag/v2.33.0):
+> attached to the [GitHub Release](https://github.com/ALJainProjects/TurboLoader/releases/tag/v2.34.1):
 > ```bash
-> pip install https://github.com/ALJainProjects/TurboLoader/releases/download/v2.33.0/turboloader-2.33.0+cu13-cp310-cp310-linux_x86_64.whl
+> pip install https://github.com/ALJainProjects/TurboLoader/releases/download/v2.34.1/turboloader-2.34.1+cu13-cp310-cp310-linux_x86_64.whl
 > pip install nvidia-nvimgcodec-cu12
 > ```
 > Other Pythons/CUDA versions: build from source below.

setup.py

@@ -657,6 +657,12 @@ def _compile_with_arc(obj, src, ext, cc_args, extra_postargs, pp_opts):
             if ct == "unix":
                 # macOS-specific flags
                 if system == "darwin":
+                    # Leave Mach-O header padding for delocate's install-name rewrites
+                    # during wheel repair. Without this, adding load commands (the video
+                    # path links 3 extra frameworks) exhausts the header slack and
+                    # cibuildwheel fails with "larger updated load commands do not fit"
+                    # (broke the v2.34.0 macOS wheels).
+                    link_opts += ["-Wl,-headerpad_max_install_names"]
                     # Respect MACOSX_DEPLOYMENT_TARGET when the build sets it
                     # (cibuildwheel does, matched to the runner's Homebrew dylibs).
                     # Only fall back to an explicit minimum otherwise, and never force