Reduce per-turn latency (currently ~5-11s)

[vgpastor]

Problem / Problema

EN: Each conversation turn currently takes ~5–11s end-to-end (perceptible on
WhatsApp). Most of it is model + tool round-trips. We already send a "typing…"
indicator, but the wall-clock is still high.

ES: Cada turno de conversación tarda ~5–11s de punta a punta (perceptible en
WhatsApp). La mayor parte es el modelo + los round-trips de tools. Ya enviamos el
indicador de "escribiendo…", pero el tiempo total sigue siendo alto.

Ideas to explore / Ideas a explorar

• Measure where the time goes (add per-phase timings to the structured logs in
  src/infrastructure/observability/).
• Reduce tool round-trips (batch/parallelize independent tool calls where safe).
• Consider streaming partial responses where the channel supports it.
• Cache stable lookups within a turn (e.g. emergency by slug).

Acceptance / Aceptación

• A short write-up of where the latency is spent, backed by real timings.
• At least one measurable improvement to median turn time.

Good starting point: src/application/conversation-service.ts and the agent
tools in src/agent/tools.ts.