Humanize tool-call cards and fix empty expanders

[SmartAI]

Problem

Tool-call rows in the chat showed the raw tool name (run_build123d) and, for text-only tools (load_skill, search_docs, lookup_docs, the record_* family, classify_reference), expanding the chevron revealed an empty box - the card only knew how to render code, measurements, the verify gate, or images, so a text-only result hit none of those branches.

Two asks from this:

1. Stop surfacing the build123d implementation detail in the UI.
2. Fix the chevron that opens onto nothing.

Change

Reworked ToolCallCard into an icon activity feed:

• Human summaries instead of raw names. toolPresentation() builds a per-tool icon + plain-language one-liner from the call's args/details (run_build123d -> "Executing code", load_skill -> Loaded skill "gears", search_docs -> Searched docs "fillet edge"). The raw name still rides in the row's title tooltip for debugging.
• No empty expanders. A row is collapsible only when it has a real body. Payload-less tools render as a flat row with no chevron, so nothing opens empty.
• Doc lookups are readable. search_docs/lookup_docs now render their result text and start collapsed to keep the chat tidy.
• CAD evidence unchanged. run_build123d and inspect_evidence keep their code / verify-gate / view-sheet bodies open by default.

Tests

• npm run typecheck clean across all workspaces.
• Full client suite passes; ToolCallCard.test.tsx gains cases for the friendly label + raw-name tooltip, the flat payload-less row, and doc-lookup expand-to-text.

Purely presentational (client-only); no protocol or server changes.