Limit agent review replies to 600 characters

[selfcontained]

Summary

• cap new agent-authored review thread replies at 600 characters in both the MCP schema and persistence service
• inject concise 1–2 sentence guidance that discourages restating feedback or narrating work
• preserve the existing 10,000-character human message limit and leave historical production rows untouched
• centralize the limit and guidance to keep enforcement and prompts aligned

No database migration or constraint is included because production already contains historical agent replies above the new application limit.

Validation

• pnpm run check
• pnpm run test — server: 2,389 passed / 8 skipped; web: 196 passed
• targeted review-feedback and MCP tests — 40 passed
• Playwright parallel project — 122 passed; one unrelated split-pane timing timeout
• split-pane Playwright spec rerun in isolation — 6 passed
• Playwright serial project — 46 passed / 11 expected live-terminal skips
• author-requested backend-security-review persona — approved with no findings