v1.6.1 - per-model weekly limits, CI fix

Per-model weekly limits return, CI fixed

• Model-specific weekly limits (e.g. Fable week 0%) are shown again when API top-up is on. They appear in the card hover and the copy-for-agents summary only, never as extra bars, so cards stay compact.
• npm test now works on Node 22 (the Windows build workflow was failing on it).
• Includes everything from v1.6.0: all Claude accounts watched at once via local capture, opt-in rate-budgeted API top-up, and the security hardening.

Builds are unsigned, so Windows may warn about an unknown publisher; verify the zip against the .sha256.

v1.6.0 - multi-account Claude usage

Every Claude account, always current, without API bans

• All Claude accounts watched at once via local Claude Code status-line capture. The single "active account" switch and the Paused state are gone.
• Opt-in API top-up (Settings, per account, off by default, confirmed once): only when an account's newest reading is over 25 minutes old. Read-only token use; credentials are never refreshed or written.
• Rate budget: at least 10 min between calls per account, one call started per minute across accounts, oldest reading served first, 429 backoff that never shortens (1h/2h/4h, honours Retry-After), and fails closed if its state file is unreadable.
• Session and weekly windows each take their own newest observation; late responses never overwrite newer readings; cards show source and age inline (local · live, api · 12 min ago).
• Security hardening from adversarial review: approval bound to the credentials path, settings IPC can't repoint profiles, symlinked/outside credentials refused, token validation, no transport-error leakage, atomic writes, profile IDs never reused.

Upgrade notes: re-tick API top-up once per account. Model-specific weekly limits are not shown; only 5-hour and weekly. Builds are unsigned, so Windows may warn about an unknown publisher; verify the zip against the .sha256.