fix(desktop): snap caret out of agent mention chips

[marcoleejr]

Summary

Fixes #2707.

With Keep addressed agents active (persistent auto-tag), the composer re-inserts @AgentName after each send. Agent mentions are decorations on plain text, not atom nodes, and the leading @ is rendered at width: 0. A click that looks like the start of the chip often lands inside the display name; the next keystrokes mutate that name, the decoration dies, and the agent is no longer tagged.

This PR treats the interior of an agent mention as non-editable caret space:

1. snapCaretOutOfAgentMention — if the empty caret is strictly inside @Name, move it just after the mention (consuming the trailing space hydration always inserts).
2. appendTransaction — apply that snap after selection changes (clicks, arrows).
3. handleTextInput — if input would land inside a chip, insert after the chip instead of corrupting the name.

Edge positions stay put (before @ for Backspace, after name so typing after a chip without a trailing space still works). Human-only mentions are unchanged.

How it was implemented

• Pure helpers in mentionHighlightExtension.ts (findAgentMentionRanges, snapCaretOutOfAgentMention) so the policy is unit-testable without the full editor.
• Wired into the existing TipTap plugin that already owns agent decoration ranges — single source of truth for “what is an agent chip”.

Test plan

• Unit tests for range finding + caret snap (interior → after trailing space; edges untouched; multi-agent; empty agent list)
• node --test src/features/messages/lib/mentionHighlightExtension.test.mjs — 29/29 pass
• Manual (desktop): enable keep-addressed-agents, @ an agent in a thread, send a few messages, click the chip / place caret mid-name, type — text should land after the chip and the tag should remain intact

Notes

Out of scope: turning mentions into real atom nodes (larger rewrite). This is the minimal fix for the auto-tag breakage path reported in #2707.