fix(ui): safely render conversation display text

[xz-dev]

Summary

• add a late, display-only safety gate for externally derived ConversationViewer text
• replace binary-like display bodies with [binary content] and visibly escape terminal-unsafe code points without mutating session messages
• route message bodies, tool data, streaming activity, header metadata, model names, and invocation tags through the same gate
• preserve exact-width rows while bounding multiline header/composer rendering to the overlay height
• apply display limits by Unicode code point before classification and escaping

Scope

This is intentionally a separate follow-up to #153 rather than an amendment to that PR. It builds on the exact-width rendering fix already merged there, but the display-safety changes span the viewer, activity and invocation-tag rendering, dependency metadata, and regression coverage.

No agent behavior, AgentSession.messages, model context, persistence, resume/fork behavior, or get_subagent_result output is changed.

Verification

• npm run lint
• npm run typecheck
• npm run test — 40 passed, 1 skipped; 754 tests passed, 5 skipped
• npm run build
• git diff --check upstream/master...HEAD