fix(desktop): reflow narrow trace turn headers - #3612

Reviewed at exact head 06b901d246211974df8fa9800386f09f0fd5ee90. Nobody had reviewed this PR before, so this is a full review rather than an incremental one.

Coverage: the container-query reflow and its threshold, the overflow-wrap: anywhere → white-space: nowrap swap on the failure phrase, and whether the new story actually constrains what it claims. Not covered: visual/aesthetic judgement, and locales beyond the two in conversation-copy.ts.

No P0–P3.

The one thing worth checking here was the unconditional white-space: nowrap on .maka-inspector-turn-failure. Replacing overflow-wrap: anywhere with nowrap removes a containment fallback, and the container query only restores a safe layout at or below 32rem — so if the failure text were unbounded, widths above the threshold would overflow horizontally where they previously wrapped.

It isn't unbounded. turnFailure resolves against the closed EN_TURN_FAILURE / ZH_TURN_FAILURE maps with a short default for unknown codes, so the longest phrase either locale can produce is Model call failed / 模型调用失败. With SESSION_WORKBAR_MIN_WIDTH = 320 and SESSION_WORKBAR_MAX_WIDTH = 600, the band left to the one-line flex layout is roughly 512–600px of workbar width, and label + failure + meta fit that comfortably in both locales. The swap is safe because the content is enumerated, not because nowrap is generally safe here — worth remembering if a raw provider message is ever routed into this span.

Choosing container: maka-inspector / inline-size on the scroll area rather than a viewport media query is the right lever, and the comment gives the reason: the workbar resizes independently of the window and the same panel also has a bottom placement, so window width is not the quantity that decides whether the head line fits.

The new story is a real constraint rather than a mirror of the implementation. It asserts geometry — failure below the label, failure and meta sharing a row within 1px, and turn.scrollWidth <= turn.clientWidth — so deleting the container query would fail it on the containment assertion, not on a selector. It also drives the actual 320px floor via width={320} instead of a synthetic viewport.

Gate state: exact-head test is completed/success, and I confirmed the Storybook steps ran rather than being skipped in that job — Build Storybook and Storybook smoke are both completed/success at this head, so the new play assertion is hosted-verified and not only locally verified. mergeable is true; blocked is the missing review, which this resolves.

Approving.