Render all usage windows dynamically from the new limits[] API (per-model / Fable, spend)

[amalakhovsky]

Summary

Anthropic's usage endpoint (GET /api/oauth/usage) has moved its per-window data out of the flat top-level keys (five_hour, seven_day, seven_day_<model> — now returned as null) and into a self-describing limits[] array. Each entry carries its own kind (session / weekly_all / weekly_scoped), group, percent, severity (normal / warning / critical), resets_at, is_active, and an optional scope.model.display_name naming a per-model window (e.g. a weekly Fable limit). Money likewise moved to a structured spend object.

The current code only knows the flat keys, so per-model windows — and any future window type — are silently dropped. In practice a maxed-out scoped limit (e.g. Fable at 100% / critical) is completely invisible.

This PR reads the new shape and renders every window the API reports, dynamically, while keeping the legacy keys as a fallback for older responses.

What changed

• New src/lib/usageModel.js — a pure, GI-free, unit-testable module that turns the payload into an ordered list of windows (normalizeWindows, preferring limits[] and falling back to the legacy keys) and normalises the extra-usage money block (normalizeSpend).
• Dynamic meters — the popup now renders one meter per reported window, reusing meters by a stable key and tearing them down only when the API stops reporting that window. Per-model windows like Fable now appear automatically.
• Severity floored at the API's — the existing burn-consequence coloring is preserved, but a window the API flags warning/critical never reads calmer than that.
• Structured spend — the extra-usage line now uses spend (authoritative amount / limit / percent / severity), falling back to extra_usage and scaling by the API's own decimal_places instead of assuming cents. It tints amber/red with the spend severity.
• New "Worst active limit" panel option — so the top-bar gauge can surface whichever limit is most severe right now; a maxed-out per-model window reaches the panel even when the 5-hour and 7-day totals are calm.
• Docs: refreshed the AGENTS.md data-source notes and the changelog; tools/poll.js now prints the normalised windows + spend.

Testing

• node tools/test-usageModel.mjs — new pure-node unit tests (10, all passing) covering the new limits[] shape, the legacy fallback, unknown/future scoped models, and the spend/extra_usage normalisation.
• Validated normalizeWindows / normalizeSpend against a real live payload: renders 5-hour 8%, 7-day (all models) 64%, 7-day Fable 100% (critical, active), and Extra usage: USD 416.54 / USD 500.00 (83%).
• glib-compile-schemas and ./build.sh both succeed; the bundle includes the new module.

Note: I couldn't get a live gjs -m tools/poll.js run on my machine (libsoup throws Gio.TlsError reaching api.anthropic.com, though curl works), so the model was validated against a captured real payload via node rather than end-to-end through the shell. A second pair of eyes on the live popup would be welcome.