diff --git a/pyproject.toml b/pyproject.toml
index 753c5a63d2..3bdb64f566 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -157,6 +157,13 @@ huggingface = [
     "pyyaml",
     "nest-asyncio",
 ]
+# "notorch" means this list declares no torch of its own, not that installing it
+# leaves you torch-free: extras are additive to [project].dependencies, which on
+# this branch deliberately carry the full runtime so a bare `pip install unsloth`
+# works, and accelerate/peft/sentence-transformers each require torch anyway. The
+# genuinely torch-free path is the installer's --no-deps pass over
+# studio/backend/requirements/no-torch-runtime.txt, not plain pip resolution.
+# Kept byte-identical to main so the two branches can be diffed for drift.
 huggingfacenotorch = [
     "unsloth_zoo>=2026.7.6",
     "wheel>=0.42.0",
@@ -177,6 +184,11 @@ huggingfacenotorch = [
     "trl>=0.18.2,!=0.19.0,<=0.24.0",
     "sentence-transformers",
 ]
+# What this extra does on this branch is raise the bitsandbytes floor; it does not
+# hand you a ROCm stack. The base dependencies still apply, so ROCm torch has to
+# come from the AMD index or already be installed (pip leaves an installed
+# +rocm torch alone, since it satisfies the base specifier). Use install.sh or
+# `unsloth studio install` for a torch-free AMD install.
 amd = [
     "unsloth[huggingfacenotorch]",
     # 4-bit decode is unreliable on ROCm before 0.50.0, the first PyPI release
@@ -545,15 +557,19 @@ flashattention = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 # torchcodec is the audio decode path; its releases track torch minor versions,
-# so each torch minor gets its own window (#7225).
+# so each torch minor gets its own window (#7225). It publishes no sdist and only
+# manylinux_2_28_x86_64, macosx_*_arm64 and win_amd64 wheels, so Linux aarch64,
+# Windows ARM64 and Intel Mac have nothing to resolve and pip would hard-fail the
+# whole install. Gate on the platforms that have a wheel, matching
+# PLATFORM_LACKS_TORCHCODEC_WHEEL in studio/install_python_stack.py.
 audio-torch280 = [
-    "torchcodec>=0.6.0,<0.8.0 ; python_version >= '3.9'",
+    "torchcodec>=0.6.0,<0.8.0 ; python_version >= '3.9' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 audio-torch290 = [
-    "torchcodec>=0.8.0,<0.10.0 ; python_version >= '3.10'",
+    "torchcodec>=0.8.0,<0.10.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 audio-torch210 = [
-    "torchcodec>=0.10.0,<0.11.0 ; python_version >= '3.10'",
+    "torchcodec>=0.10.0,<0.11.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 colab-ampere-torch211 = [
     "unsloth[huggingface]",