feat(agents): Markdown toggle and token breakdown in Compare versions drawer

[Bor4brn]

Summary

The Compare versions drawer showed agent answers as raw text, so a structured answer (headings, lists, tables) was hard to read and compare. Its stats row showed only a total token count. This PR renders answers as Markdown by default, adds a toggle to switch to plain text, and adds an Input / Output / Cache token breakdown directly beneath the Tokens stat.

Changes

• Markdown / Plain text toggle. Markdown is the default. One setting applies to both columns, so the two answers are always shown the same way, and it stays when you switch versions, press Reset, or close and reopen the drawer. User messages and error turns look the same as before.
• Two-row stats. The first row is Latency, Cost, Tool calls and Tokens. Input, Output and Cache sit directly beneath Tokens, grouped by a divider. The layout responds to the column's width rather than the screen's, so in a narrow column the token group drops to its own full-width row.
• Missing token data. A value that no answer reported shows — with a "Not reported by the model" tooltip, while a reported zero still shows 0. When an answer gives input and output but no total, the total is now calculated from them (before, the Tokens stat showed —). The Cache label's tooltip explains that cached tokens are already counted in Input.
• Shared Markdown renderer. The session transcript's Markdown renderer is now a shared AgentMarkdown component, so both views look the same. It also brings back list bullets and numbers, which Tailwind's CSS reset was removing. The transcript had the same missing-bullets problem, so this fixes it there too.
• Translations. All of the drawer's text, which was hard-coded English, now comes from the agents.compare translation keys in English and Turkish.
• Scrollbar. The conversation scrollbar has its own space instead of covering the right edge of the message cards.
• Tests. src/__tests__/unit/agent-compare-versions-stats.test.ts covers reading token usage from the response, "not reported" versus zero, the totals, the default display mode, and that Turkish has every English translation key.

Validation

• npm run lint: 0 errors. There are 208 warnings elsewhere in the repo; none are in the changed files.
• npx tsc --noEmit
• Tests: ran the new test file and the related agent studio and session tests (84 passing). The full npm run test was not run.
• npm run build
• npm run docs:build: not run. No docs were changed.
• Tested by hand at localhost against a real agent comparing v1 and v2: Markdown on and off, the stats layout, and the token breakdown.

Release Notes

• Docs updated when behavior changed
• Security-sensitive changes reviewed
• License or policy files updated if repo-facing behavior changed