- Add _get_explicit_cost() with priority chain: usage.cost.total > usage.cost flat > cost/cost_usd
- aggregate_per_model() and _top_sessions() now use _get_explicit_cost()
- Fix message limit prediction to use elapsed time instead of total window
- Add 224 lines of test coverage for typed limits, explicit cost priority,
pricing override shapes, and prediction logic
- Add typed limit fields to RuntimeUsageCurrent: output_token_limit,
total_token_limit, message_limit with matching pct and source
- Add total_output_tokens and output_tokens_per_minute to burn rate
- _build_current() now computes each pct from matching units only
- Legacy token_limit backfilled from typed limits for backwards compat
- Frontend aggregateRuntimeUsage() tracks typed limits separately
- limit_kind field on predictions indicates which limit drove time-to-limit
- Add source and confidence fields to RuntimeUsageWindow, ModelUsageEntry,
TopSession, RuntimeUsageCurrent, and ProviderUsageScrapeResult schemas
- _build_window() assigns source based on data origin:
provider_native > provider_api_rate_limit > local_jsonl_estimate
- _build_current() tags token_limit_source and cost_limit_source
- Frontend relabels 'Current session'/'All models' to 'API rate limit'
- Shows source label and confidence in usage strip
- Changes 'did not return active usage windows' to 'did not return
API rate-limit windows for percent + reset diagnostics'
null