Remove contradictory voice instructions at their source

When a Pipecat demo changes greeting recovery from model speech to a tracked tool, remove the legacy greeting rule that prohibits tools and orders a literal spoken introduction. A later conditional recovery reminder can leave the model applying the older rule to a generic greeting. Scope the replacement to the demo boundary and preserve the production instruction. Check paused and unpaused choices separately through the actual model API: a harness that forces the intended handler can pass while the model still speaks without calling it. This is evidence of one prompt conflict and its scoped correction, not a universal wording guarantee across models.

Applies when: New conversation policy coexists with inherited step-level speech instructions. Keywords: Pipecat prompt conflict greeting tool routing demo context model tests