[xvcezrqpbxtpzblrt]

Problem

The thinking row above an answer had three faults:

• while the model thought it showed the newest reasoning line, with no time: a five-minute thought looked frozen;
• after the answer began it said "Thought for 5m 26s" but no token count;
• that duration ran from the first reasoning token to the first answer token, so a tool call in between was counted as thinking.

Result

• While thinking: Thinking… 1m 20s · ≈3,120 tokens, one tick a second, nothing else in the chat re-renders.
• When the answer starts: Thought for 5m 26s · ≈3,120 tokens, frozen.
• The time is the sum of thinking segments. A segment ends when the answer text starts, a tool call begins, an error arrives or the turn ends; a later thought opens a new segment instead of replacing the first. Tool runs are not counted.
• The token count is an estimate (a CJK character is one token, other text about four characters per token) and is marked ≈; the engine does not report a reasoning-only count.
• All nine languages have the new ≈{0} tokens label; the existing "Thinking…" and "Thought for {0}" translations are reused.

Validation

• node --test tests/live-chat.test.mjs → 17 pass, including: a running segment counts up, a closed message counts only closed time, two segments add instead of replacing, an idle message adds nothing, and the token estimate (CJK vs Latin vs empty).
• npm run typecheck clean; catalog tests (key sets, French typography, gaps) pass.
• Live in the app: the row appeared while streaming as 思考中… 0s · 约 12 词元, then settled to 推理 · 约 37 词元 (Swift thought for under a second, so the "Thought for" form is not used); the long earlier answer now reads 思考了 5m 26s · 约 40,403 词元.
• Not run live: a multi-minute thought with a tool call in the middle; that timing is the unit-tested part.

Limits

• Reasoning tokens are approximate; only the engine's total output tokens are exact.
• A thought shorter than one second still shows "Reasoning · ≈N tokens" rather than "Thought for 0s".
• The text is display-only; it is not sent back to the model and does not change the answer.

[xvcezrqpbxtpzblrt]

Superseded by #4606, which carries this change with the rest of the 0.2.7 branch. Closing so the review has one place.