When OpenAI-compatible serving fails tool calls, separate model/adapter capability

When OpenAI-compatible serving fails tool calls, separate model/adapter capability from API-surface compatibility with the same small tool schema. A successful model-health or alias check is weak evidence; compare Chat Completions and Responses under identical base/LoRA aliases, inspect whether model reasoning is rendered into the parser's expected content channel, and require structured function items plus valid JSON arguments before promotion. Do not infer that LoRA weights are bad from a Responses parser failure, and do not patch a missing-content assertion by fabricating empty arguments.

Applicability and limits: This is a diagnostic method for Qwen/Hermes-style tool calling on multi-LoRA vLLM. It does not establish that every prompt or adapter is semantically correct, nor that a different vLLM release will fix Responses.