Add opt-in live quota windows, budgets, pace projections, and alerts

[ekalb81]

Outcome

Show current subscription/API allowance windows, reset timing, balances, budget headroom, and projected run-out without conflating provider limits with transcript-derived token or API-equivalent cost.

Current state

• Sessions may contain plan_type, credits_balance, and unlimited-plan metadata, but these are point-in-time transcript fields.
• Odometer has no quota polling, reset countdowns, pace projection, configurable provider/project budgets, or usage alerts.

Scope

• Define a provider-agnostic QuotaSnapshot with window kind, used/remaining values, unit, limit/reset times, provenance, fetch time, confidence, and stale/error state.
• Implement quota sources as capabilities separate from transcript parsers. Reuse an existing provider session only with explicit consent; never copy passwords/tokens into Odometer config or logs.
• Support multiple simultaneous windows (for example burst/session, weekly/monthly, credits) without coercing unlike units.
• Add pace, reserve/deficit, projected exhaustion, and reset-aware budget calculations from one backend service.
• Add per-provider and per-project soft budgets with local notifications and quiet/stale/offline behavior.
• Surface quota status in the dashboard and existing tray. Hard enforcement and config hooks are explicitly deferred.

DRY boundary

• One quota service feeds dashboard, tray, future widgets, CLI/statusline, and alerts.
• Historical usage comes from the ledger; quota snapshots remain provenance-bearing external state.
• Pricing/budget math consumes the shared pricing/plan authority.

Acceptance criteria

• Provider outages, authentication expiry, rate limits, clock skew, sleep/resume, offline startup, and stale caches produce honest states rather than zero usage.
• Credentials and response bodies never enter diagnostics, performance logs, exports, or fixtures.
• Notifications are opt-in, deduplicated, reset-aware, and configurable per window.
• Forecast math is deterministic and suppresses projections when evidence is insufficient.
• Transcript-derived balances and live-provider snapshots can coexist and identify their source.

Dependency and ordering

• Phase: 2 — product intelligence
• Priority: P0
• Depends on: Unify pricing, model aliases, cache creation, currencies, and subscription plans #42, Add provider detection, parser health, and data-quality diagnostics #39
• Parent roadmap: Roadmap: close competitive analytics gaps on shared local-first foundations #36