feat(dashboard): surface cumulative token totals in StatsBar

[ninadgns]

Problem

The StatsBar shows Context, which reads like overall token usage but isn't. context_tokens is a snapshot of the last message's window (input + cache_creation + cache_read of that one message, stream.py:412) — it moves up and down as the conversation compacts, so it can't be a total. Sitting next to a settled Cost figure, it invites exactly the wrong reading.

Meanwhile the four cumulative counters are already streamed to the browser. The usage audit event carries the full breakdown:

"context_tokens":              ...,
"total_input_tokens":          ...,
"total_output_tokens":         ...,
"cache_creation_input_tokens": ...,
"cache_read_input_tokens":     ...,
"total_cost_usd":              ...,

computeLiveStats received that whole object and read two fields, discarding the rest. There was no way to see overall usage in the UI at all — you had to query the audit endpoint or Postgres directly.

Change

Add a Tokens stat showing the sum of all four counters, with a hover breakdown (Input 1.5k · Output 2.0M · Cache write 0 · Cache read 300) naming each component. Context stays as-is — it's a genuinely useful and different signal.

Render-only. No backend, DB, or event-shape changes; the data was already client-side.

Notes

• Summing all four matters: cache reads dominate agentic runs, so input+output alone under-reports by an order of magnitude. Covered by a regression test.
• Follows the existing live-then-settled fallback used by formatContextStat.
• Renders — rather than 0 when nothing was ever reported, per the repo's fail-fast rule that distinct failure modes must render distinctly — a pipeline that never emitted usage must not look like a free run.
• total_output_tokens includes thinking tokens; the breakdown label stays generic rather than implying otherwise.

Testing

• pnpm exec vitest run — 95 files, 719 tests pass (6 new)
• pnpm exec tsc --noEmit — clean

Python checks (pyright, ruff, pytest) were not run: this change touches no Python.

🤖 Generated with Claude Code