Analyse raw prompt text end-to-end.

Unlike /analyse (which can take a pre-computed wv), this endpoint always
profiles the prompt with the LLM: it calls HuggingFaceLlamaClient.generate
to produce the weight-vector traces, which then flow through the normal
similarity preprocessing, SOM inference and ensemble. Requires the LLM to be
loaded server-side (PROFILER_LOAD_LLM=1).