fix: keep agent work expansions anchored

[morgmart]

Category: fix
User Impact: Expanded agent work stays visually anchored, while long tool details remain compact and can be scrolled with a pointer or keyboard.
Problem: Long tool-call details made Agent Work difficult to scan, and opening the surrounding Previous steps disclosure could make the virtual transcript jump toward the bottom. The compact details region also needed an accessible keyboard-scrolling path.
Solution: Cap Agent Work tool details in a labeled, focusable scroll viewport, pin user-opened Agent Work disclosures to their transcript row, and only auto-follow measured height changes while the virtualizer still owns a bottom anchor.

File changes

src/features/chat/ui/AgentWorkPanel.tsx
Pins the virtual transcript anchor before either Agent Work disclosure changes height, preserving the reader’s position.

src/features/chat/ui/ToolCallAdapter.tsx
Places Agent Work tool details inside the shared compact scroll viewport and supplies its localized accessible label.

src/features/chat/ui/VirtualMessageTimeline.tsx
Prevents measured-height bottom-follow from overriding a row anchor established by user interaction.

src/features/chat/ui/tests/ToolCallAdapter.test.tsx
Covers the Agent Work-only height cap, scrolling behavior, region label, and keyboard focusability.

src/features/chat/ui/tests/VirtualMessageTimeline.test.tsx
Protects the regression where expanding and remeasuring an Agent Work row forced the transcript toward the bottom.

src/shared/i18n/locales/en/chat.json
Adds the English accessible label for tool details.

src/shared/i18n/locales/es/chat.json
Adds the Spanish accessible label for tool details.

src/shared/ui/ai-elements/tool.tsx
Adds a reusable focusable tool-details viewport with the shared focus treatment.

Screenshots

Before

Long tool output expands the Agent Work row beyond the viewport.

After

Tool details stay compact in a keyboard-accessible scroll region while the expanded Agent Work row remains anchored.