LM Studio update breaks tool calling

This is the ugly part of local AI infrastructure: the model may be fine, but the wrapper around it can still take your workflow down overnight. If your agents depend on provider-specific parsing, version pinning and rollback paths stop being optional. Tool calling is brittle because small output-format changes can break downstream orchestration even when the model itself is unchanged. Local stacks amplify the pain since users often rely on desktop apps or binaries that do not make downgrades easy. The community pattern is clear: pin known-good versions, archive installers, and treat inference software like any other production dependency. Provider abstraction helps, but it does not remove the need to test each model family and runtime combination for tool-call compatibility. For agent builders, this is a reminder to add canaries and regression tests around tool invocation, not just model quality.