feat(chat): readability polish + message actions, regenerate, feedback

[OmGuptaIND]

Summary

• Bumped markdown / message font sizes and tightened line-height for better readability; redesigned the composer Add menu with a larger surface, viewport clamping, and brand-chip stack.
• Added per-assistant-message actions (copy, 👍 / 👎, regenerate) and wired thumbs feedback to Braintrust via a new feedback protocol message that pairs a per-turn messageId with the trace span id.
• Added a regenerate protocol message + server handler: Pi-SDK sessions pop the last user → assistant pair and re-run; Codex harness sessions re-feed the prior user text (no thread/resume RPC); legacy Claude Code harness is unsupported.
• TurnProgress steps auto-collapse once the agent finishes writing to the main conversation, while respecting manual toggles for the rest of the turn.

Test plan

• Type a long markdown reply and verify font sizes / spacing read better than main.
• Open the composer Add menu near the right edge of the window and confirm it stays inside the viewport.
• Send a message, then on the assistant reply: copy, thumbs up, thumbs down, and regenerate; confirm Braintrust receives the feedback score and the regenerated turn replaces the old one.
• Watch a turn complete on the Pi-SDK provider — steps should auto-collapse when the final answer renders; manually toggling pins the panel state for that turn.
• Repeat regenerate on a Codex harness session.

🤖 Generated with Claude Code