ARM toolchains

arm-none-eabi GCC 13.2.1 cc_toolchains, one variant per CPU core we ship firmware on, with the core's codegen flags baked in:

Variant	Flags	Used by
cortex_m0	-mcpu=cortex-m0 -mfloat-abi=soft	stm32f0
cortex_m4f	-mcpu=cortex-m4 -mfpu=fpv4-sp-d16 hard float	stm32f4/g4
cortex_m7f	-mcpu=cortex-m7 -mfpu=fpv5-d16 hard float	stm32h7

The upstream @arm_none_eabi toolchains carry no codegen flags — that's the point of this package. Baking -mcpu/-mfpu/-mfloat-abi into the toolchain (selected via the target platform's //platforms:mcu_core constraint) means every cc_library in the dependency graph compiles with the right float ABI, with zero per-target flag plumbing. The flags sit in both copts and linkopts: linkopts also drive newlib/libgcc multilib selection.

toolchain.bzl is a local mirror of the fork's arm_none_eabi_toolchain macro with one change: the generated cc_toolchain/config targets are tagged manual. Without that, bazel build //... matches the toolchains for all five host OS/arch combos and downloads every host's ~150 MB gcc archive. Toolchain resolution ignores tags, so the registered toolchain() declarations still work and only the selected variant fetches its repo.

Adding a variant

New CPU core (say cortex-m33): add a constraint_value in //platforms, an arm_none_eabi_toolchain(...) call in BUILD.bazel here with the core's flags, and use the constraint in the new family's platform. //toolchains:all is already registered in MODULE.bazel.