BYO AI usage dashboard and org-key chat runtime

What's New

If you bring your own AI key, you can now see what it actually burned — tokens per model and a history table — without mixing that with Road To cloud generation credits. The assistant also talks through the org's configured provider instead of always hitting Gemini.

Features

• BYO AI usage: on the AI Usage page, cloud credits stay a generation count. A separate org key / BYO block shows token cards by model (share of BYO total, cost labeled BYO) and a history table (date, type, model, tokens). Empty states link to AI settings. Platform Gemini calls are not listed there.

Improvements

• Assistant uses the org AI slot: OpenRouter or another BYO provider is the runtime for chat, not the env Gemini key. That removes the 502 when Gemini is not the active provider.
• AI settings: the API key is masked after save and the selected model stays selected.

Corrections

• Changelog MDX no longer uses HTML comments that broke prerender of /changelog on Vercel (#124).
• Update-check TypeScript: GitHub not-modified without a cached release is handled so production typecheck passes (#123).

Administration

• New append-only AiByoUsageEvent (org, provider, model, tokens, surface). Logged only when runtime.source === "org" (self-host included). Test-connection is silent. Telemetry persist is fail-open (#126, #127).

Tooling

• Prisma migration for AiByoUsageEvent — run pnpm prisma migrate deploy on existing instances.