feat: per-task cost & token dashboard

[kovtcharov]

Summary

`backend/src/usage-reporter.ts` already has scaffolding for token reporting, but it's gated behind the `USAGE_DASHBOARD_URL` env var and only forwards to a remote endpoint — there's no in-app dashboard. Build a local per-task and per-workspace cost view so users can see what each task is burning in real time.

Scope

Backend

• Capture token usage from:
  • Claude Code CLI subprocesses — parse usage from their stream-json output (`conversation-parser.ts` is the natural seam)
  • Direct Anthropic SDK calls in `supervisor-chat.ts`, `voice-supervisor.ts`, `llm-service.ts`
• Persist per-task aggregates (input/output/cache-read/cache-write tokens, dollars-spent estimate by model)
• Optional: emit OpenTelemetry GenAI semantic conventions for export to Langfuse / LangSmith / Braintrust

Frontend

• Per-task: small badge in the task header showing tokens + estimated $
• Per-workspace: aggregate panel
• Settings: optional budget cap per workspace with soft warning + hard stop

Why

• Today users have no visibility into what tasks cost
• Pairs with the prompt-caching issue — you need the dashboard to see if caching is actually working
• Natural extension point for an OTel exporter once OpenTelemetry GenAI conventions stabilize

Out of scope

• Per-user/multi-tenant budgets (single-user assumed)
• Historical cost trends / charting (start with current-state numbers)