Observability: structured logging, Analytics Engine, and OpenTelemetry hooks

[PAMulligan]

Summary

Replace ad-hoc console.log calls in the worker with structured logs, ship metrics to Cloudflare Analytics Engine, and expose OpenTelemetry hooks for users who want to forward telemetry to their own observability stack.

Motivation

The analytics dashboard answers product questions; observability answers operational questions. Without it, debugging production issues (a flaky tool call, a slow Anthropic response) is guesswork.

Acceptance Criteria

• Worker logs are JSON with consistent fields: level, event, tenant, conversation_id, latency_ms, model, plus an error chain for failures
• Counters and histograms pushed to Cloudflare Analytics Engine: chat.turn.latency, chat.tool.calls, chat.rate_limit.hits, chat.errors
• OpenTelemetry tracer hook (createTracer()) lets users forward spans to their own collector
• PII-safe by default: message bodies never logged unless LOG_MESSAGES=true (documented as off in prod)
• Runbook in docs/operations.md covering common alerts and remediation steps