What kinds of mismatches in toolchain flags cause subtle build or runtime failures?

The most dangerous are ABI-level mismatches across object files and libraries. The float ABI is the classic one: compiling some objects with -mfloat-abi=soft (floats passed in integer registers, software emulation) and others with -mfloat-abi=hard (floats in FPU registers) produces incompatible calling conventions, so a function call passing a float corrupts the value or the link fails, and the C library you link must match the same float ABI. Mismatched -mcpu/-march/-mfpu similarly produce code targeting different instruction sets, causing illegal-instruction faults at runtime. Mixing Thumb and ARM state inappropriately, or different -mthumb settings, can break calls. Beyond ABI, library/specs choices matter: linking full newlib vs newlib-nano changes size dramatically and using printf with floating-point support pulls in large code, which can overflow flash. Toolchain version differences are subtler, different GCC versions optimize and warn differently, so a build that works with one version may expose a latent undefined-behavior bug under another's more aggressive optimization, which is why you pin the toolchain version in CI for reproducibility. The unifying rule is to use one consistent, target-matched flag set (-mcpu, -mthumb, -mfloat-abi, -mfpu) across every object and the library, because the linker generally can't catch a semantic ABI mismatch that still links, it just misbehaves at runtime.