[sun-970]

Summary

Fixes #413: switching employees in the org workbench no longer flashes both the center PositionCard and the right TurnPanel to skeletons, stacks empty-state on the loading thread, or crossfades two tree-row highlights.

What changed

• loadPosition keeps the previous card.data while loading is true; PositionCard only skeletons when there is no record yet.
• Employee switch no longer setTurns([]) before the next history arrives; a null session during attach no longer wipes the current thread.
• TurnThread does not mount the empty-state node while loading; CSS .owb-turn-thread--empty[hidden] { display: none } covers the display:flex vs [hidden] override.
• Org tree rows no longer transition background, so selected fill moves in one frame.

Tests

• packages/ui PositionCard stale-while-loading + org-tree CSS contract
• TurnPanel loading does not keep .ui-org-empty-state in the DOM
• App A→B with B's position() in flight keeps A's card heading and thread
• renderer-css-syntax asserts the [hidden] rule

npm run test:ui (40), App.test.tsx (86), TurnPanel.test.tsx (28), renderer-css-syntax.test.cjs (3).

Acceptance (#413)

• A→B does not full-panel skeleton if A was already on screen
• Loading skeleton and empty-state are not both in the DOM
• .owb-panel shell no longer collapses to a 4-row skeleton over stale content
• Selected tree fill is not a 120ms crossfade
• First selection with no record still skeletons

[waterbro-8]

REQUEST_CHANGES，head 4abc1705。方向对（stale-while-revalidate + 空态不跟 skeleton 叠），但现在不能批。

• CONFLICTING/DIRTY： base 停在 #414（6e26755a），#410 已经进 main（f5e9af18）。多半是 CHANGELOG 同一处。请 rebase 到最新 main。我推不了你的分支。
• sessionId === null 不再 setTurns([])： 切到还没有 session 的员工 B 时，A 的线程会一直留在 B 的面板上。stale-while-revalidate 只应持续到 B 的 history/空 session 有结果；确认 B 无 session 后要清掉 A 的 turns（按 selectedId 盖戳也行）。
• CI 还没跑。rebase 后等绿再叫我复审。非产品验收。

[sun-970]

Rebased onto current main 6e26755 (#414). CHANGELOG keeps both #412 and #413.

loadSessions now setTurns([]) once this employee is confirmed to have no session (or the sessions read fails). Stale-while-revalidate still holds while attach is in flight. Added #413 clears employee A's thread once B is confirmed to have no session.

#410 (f5e9af18) is merged to feat/402-multi-party-approval-policy, not main.

[waterbro-8]

Code review APPROVE (not product 验收), head 37b4b33d.

Rebase onto current main looks clean (CHANGELOG keeps #412 and #413). Confirmed-empty session now setTurns([]) after the sessions read (and on sessions fail / offline), so employee A's thread does not stick on B. Tests cover in-flight stale card/thread and the no-session clear. CI 11/11 green.

Not product 验收. 胡奕舟 can merge.

[PeterGuy326]

Approved at head 37b4b33d. Read the full diff and traced the state transitions in App.tsx at this head.

Verified in code:

• loadPosition keeps prev.data only while loading; every terminal path still overwrites it (404 / position_missing → notFound: true, non-200 → data: null, catch → data: null), so a failed fetch cannot pin employee A's record on B's panel.
• The three removed setTurns([]) calls are covered by loadSessions: the confirmed-empty session branch, the non-200 sessions branch, and the offline catch all clear turns and historyLoading, each guarded by selectionVersion + selectedIdRef. A late read for A therefore cannot wipe B, and loadTurnHistory returning early on sessionId === null without clearing is safe.
• selectPosition bumps selectionVersion before the stale window opens, which is what keeps A's in-flight reads inert.
• .owb-turn-thread--empty[hidden] { display: none } is load-bearing rather than dead CSS: the author rule sets display: flex on --empty, which beats the UA [hidden] { display: none }, and the <ol> still carries both the --empty class and hidden. The new selector wins on specificity (0,2,0 vs 0,1,0) and on source order.
• Dropping transition: background from .ui-org-tree__row makes the prefers-reduced-motion override redundant, so removing that block is consistent.

Non-blocking, worth a follow-up: during the stale window selectedPosition is still A's record while selectedId is already B, and the actions slot mixes the two — DismissPositionDialog renders selectedPosition.name (A) alongside selectedNode's descendant count (B) and calls dismissPosition(selectedId) (B). The window is a single local position() round trip, so it is not realistically clickable today, but the label/target split is the kind of thing that bites later. Fine to land as-is.

CI 11/11 green on this head. waterbro-8's approval covers the rebase and the no-session clear; my pass was a diff + source read, I did not re-run the renderer suites locally.