feat(desktop): live conversation token rate (tok/s)

[GISWLH]

Fixes #394

Summary

• Show a live sliding-window output rate (tok/s) on the transcript stream-health strip while a turn is running (working / run-activity indicators, plus a compact chip while answer tokens stream).
• Prefer provider outputTokens mid-stream; otherwise estimate from visible thinking+answer text (four code points ≈ one token) and label as approximate (≈).
• Context inspector Generation speed remains the completed-turn snapshot (D428).

Algorithm

• Cumulative samples on a stable ~250ms interval (useLiveTokenRate keeps content/thinking/outputTokens in refs so stream deltas do not tear down the timer).
• Rate = rounded Δtokens / Δt over a 2s sliding window (min duration 200ms).
• TTFT: until some positive output has been seen, rate stays undefined (chip hidden)—never a false 0 tok/s.
• Stall: after positive output, a drained window reports 0 so reconnect/stall is visible.
• Estimate → provider handoff: if the source flips to provider usage and the provider count is lower than the estimate, reset the sample window so monotonic clamping cannot freeze a hard 0 without ≈.

Review fixes (this push)

• a11y: ticking rate is aria-hidden / decorative (mirrors working-elapsed); streaming-only chip is not a polite live region.
• Guard mounts with tokensPerSecond !== undefined (not always-truthy {tokenRate ? …}).
• zh-CN mirrors for component-spec, interaction-patterns §3.2a, E2E plan, decisions-log / D428.
• FR live copy uses jetons/s.

Tests

• pnpm typecheck / pnpm lint in apps/desktop — green
• node --test test/streaming-token-rate.test.mjs test/active-turn-surface.test.mjs — green (TTFT undefined, estimate→provider reset, hook interval deps, a11y/guard contracts)