Live Codex usage (fix 0%), OpenCode Go provider, multi-account Gemini + UI fixes

[luispriet]

This adds the improvements I built locally on top of v1.1.0 (Fedora 44, KDE Plasma 6). All changes are backward-compatible; metadata bumped to 1.2.0.

Fix — Codex usage showed 0%

The Codex provider read the passive ~/.codex/sessions logs. Codex 0.139.0 emits a token_count whose rate_limits.limit_id == "premium" with null primary/secondary; the old fallback picked it, so used_percent fell back to 0. Now it queries the same live endpoint /status uses — GET https://chatgpt.com/backend-api/codex/usage (ChatGPT OAuth from ~/.codex/auth.json, ChatGPT-Account-Id header). It's a plain GET (no token cost), read-only (never writes auth.json), and falls back to the local history (hardened to skip the null premium bucket) if the request fails.

Feature — OpenCode Go provider

Dollar-value limit windows ($12 / 5h, $30 / week, $60 / month, all configurable). OpenCode has no public usage API, so spend is a local estimate from ~/.local/share/opencode/opencode.db (read-only SQLite) and is clearly labeled "this machine (est.)". The monthly window is anchored to a configurable renewal day.

Feature — multiple Gemini accounts

Reads the primary ~/.gemini/oauth_creds.json plus any ~/.gemini/accounts/*.json → one card per account (labeled by email from the id_token). It refreshes expired tokens itself (via oauth2.googleapis.com/token) using gemini-cli's public OAuth client (installed-app credentials shipped in the gemini-cli package — not a user secret), so an account the CLI isn't actively using doesn't die when its token expires. JSON stays backward-compatible (top-level = first account, plus an accounts array).

Fixes

• Stuck loading (grey refresh button, … instead of values): isLoading was a manual +1/-1 counter, but Plasma's executable engine collapses identical in-flight sources, so the counter never returned to 0. Now derived from runner.connectedSources.length.
• Popup clipped with many providers/accounts: content wrapped in a Flickable with a height cap → it scrolls instead of being cut off.

Happy to adjust naming, split into smaller PRs, or rework anything you prefer.