[mindtheflo]

Summary

Replace start-of-window/per-licence extrapolation with a global enabled-fleet forecast, separately for Codex and Claude. Individual licences now show measured quota and renewal times only.

• Persist minimal local quota observations; use actual usage deltas, not a quota-window lifetime average.
• Combine a last-24-comparable-hour baseline (6h half-life), a recent burst estimate (30m half-life), and learned local-hour weekday/weekend profiles. Carry Friday into Monday without weekend-zero dilution.
• Simulate session and weekly constraints on the same account, future renewals, idle cycle starts and cooldowns. An idle weekly allowance gets a full-week horizon.
• Show blue/green/orange only when conservative demand scenarios support the classification. Orange requires a deficit against an optimistic cumulative supply envelope, not a failed routing heuristic. Neutral on insufficient history, missing/stale readings, ambiguous availability, mixed/unknown plan capacity, or colour-boundary disagreement.
• In OpenAI + Claude mode, bar heights use pooled weekly capacity only (86/500 = 17.2%); session limits still affect forecast colours, not fill. Provider tooltips explicitly say Weekly. Keep provider capacities separate even in combined session/weekly icon mode. Add global forecast lines and neutral reasons in English/French; remove per-licence pace text, colours and diagnostic fields.

Full equations, assumptions and confidence limits: docs/forecasting.md. Sensitivity bands are not claimed as calibrated statistical confidence intervals.

Validation

• swift test: 63 tests passed, including bursts, weekday/weekend boundaries, inactive weekends, Monday continuity, staggered renewals, repeating sessions, coupled constraints, provider isolation and abstention.
• bash scripts/install.sh with the existing Developer ID: release app, archive and plugin built successfully; installed app signature verified. Installed build identity is 60d6394, with no local source modifications.
• Live read-only quota check succeeded. Both provider forecasts correctly stayed neutral while recent history is being collected. After the weekly-fill update, the latest recorded Claude weekly pool is 88/500 points, or 17.6%.
• Native demo menu visually checked: global forecast lines, neutral measured licence bars, no individual forecasts. No real account data used in the visual check.

Test: swift run cpamc-menubar --demo --show-menu — this is a native macOS menu-bar app; there is no HTTP preview URL. Production forecasts need at least six observed comparable hours plus recent coverage after the new app starts collecting history.

Draft only; not merged or released. Installed locally at the user's request, preserving configuration, signing identity and collected forecast history. Proxy routing and enabled-account state were not changed.