feat(ui): render mermaid diagrams in chat (experimental, admin opt-in)

[albanm]

Add experimental Mermaid diagram rendering to the agent chat UI, behind an
admin-only, per-user opt-in toggle.

What changed

• New mermaid dependency (dynamically imported, kept out of the initial bundle).
• Markdown renderer emits a <pre class="mermaid"> container for ```mermaid fences
  when the flag is on; a post-render pass (utils/mermaid.ts) turns each into an SVG
  independently, so one bad diagram can't break the rest of a message.
• Mermaid theme variables are derived from the live Vuetify palette (incl. an
  xyChart block so xychart-beta plots/axes stay on-brand and legible).
• New MarkdownContent.vue wrapper replaces the inline v-html render in
  AgentChatMessages; re-renders on stream end and theme change.
• Inline error fallback for invalid diagrams with a "Fix this diagram" button that
  sends a correction turn (visible user message + hidden context built by
  formatMermaidFix).
• mermaid few-shot (xychart-beta) appended to the system prompt when enabled.
• Settings/debug dialog opened to all users (was admin-only); the mermaid toggle
  inside is gated behind isAdmin.
• Unit tests for markdown rendering, theme-variable mapping, and formatMermaidFix;
  dev chat mocks gain a valid and an intentionally-invalid diagram.

Why: experiment with letting the assistant visualize quantitative data as charts
directly in the chat, without committing it to all users yet.

Regression risks:

• Settings dialog opened to all users (AgentChatHeader lost v-if="isAdmin" on
  the cog). The dialog's Info tab is not gated — it now exposes the full system
  prompt and every tool's description + JSON input schema to any user, including
  anonymous/external users in embedded widgets. Previously this was reachable only by
  admins. Confirm this disclosure is intended; if only trace-consent settings were
  meant to be exposed, the Info tab (and tool-exploration toggle) should stay
  admin-gated.
• Markdown rendering for assistant messages now flows through the new
  MarkdownContent.vue wrapper instead of inline v-html. Behavior should be
  identical when mermaid is off, but it's a shared rendering path for every message.
• renderMarkdown/renderStreamingMarkdown gained an optional opts.mermaid arg
  backed by a module-level mermaidActive flag set synchronously before each parse.
  Safe today (only consumed by MarkdownContent, parsing is synchronous), but it is
  shared mutable state to keep in mind.

Cleanup (non-blocking):

• AgentChatHeader isAdmin prop is now dead (and still passed from AgentChat.vue).
• MarkdownContent.vue introduces one vue/no-v-html lint warning — the
  eslint-disable-next-line comment targets the <div> line, not the v-html
  attribute line, so it doesn't suppress.