Add AgentPond tracing for Genkit language flows

[marcusschiesser]

Summary

• initialize Firebase Admin and the AgentPond exporter before Genkit and Vertex AI load
• add privacy-safe OpenInference LLM spans around all six language-learning prompt flows
• cover Chinese and English explanation, image analysis, sentence generation, collocation analysis, and word-in-context explanation
• add default-deny Firebase Storage Rules for server-owned trace data
• update Firebase Admin and TypeScript to versions required by AgentPond
• retain the AgentPond skills installed by npx agentpond@0.5.2 init

The manual spans add operation, provider, model, status, latency, and error metadata without copying prompt text, images, or generated explanations into span attributes. Genkit's own child spans continue to use the same global OpenTelemetry provider.

Why AgentPond

AgentPond complements HanziGraph's Genkit evaluations with per-request production traces stored in its Firebase project. Maintainers can compare latency and failures across six language-learning flows, identify whether a problem occurred in authorization, a named prompt, Vertex Gemini, or structured output, and analyze behavior locally with DuckDB.

Verification

• npm run lint in functions/
• npm run build in functions/
• root npm run build
• loaded the compiled bootstrap on Node 22 and confirmed all six Functions exports
• verified the manual OpenInference LLM span shape with an in-memory exporter
• git diff --check

Credentialed Genkit evaluations and production Firebase Storage export were not run because the contributor environment does not have Vertex or bucket access.