feat: quota tracking and load rotation across subscriptions

[eliahilse]

Never launch an engine that's out of quota, and spread load across the subs that still have headroom:

• Cooldown tracking: every engine run records its outcome in ~/.local/state/kyora-review/usage.json. A run that hits a usage limit (detected from vendor error signatures on failed runs only — a finding about rate limiting can't false-positive) puts that engine on cooldown: vendor "try again in N hours/minutes" hints are parsed when present, otherwise cooldownMinutes (default 60). Cooled-down engines are skipped with a log line; --ignore-quota forces them; if the whole panel is cooling, the run no-ops CI-safe instead of burning attempts.
• Load rotation: --max-engines <n> / maxEngines config runs only the n least-recently-used healthy engines per review — with a 6-engine panel and max-engines: 3, consecutive PRs rotate through different subscription triples.
• kyora-review usage shows per-engine runs, last outcome, and remaining cooldown; doctor annotates cooldowns inline.
• CI persistence: the action's cache now carries the state dir alongside ~/.codex, so cooldowns survive between workflow runs. Also fixes the action's auto-detection gap for the new engines: ZAI_API_KEY → glm, QWEN_API_KEY → qwen (and claude-binary install covers both); dropped the dead qwen-code install.
• 7 new unit tests (26 total).

Stacked on #6 — merge that first and this collapses to the quota changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CQy3ZJ5MyUxo4qjZwA93Na

[eliahilse]

Extended with live usage probes — the selector now asks vendors for remaining quota before launching, falling back to cooldowns wherever the probe can't answer. Live-validated on this machine: claude reports 41% remaining (5h/7d windows, via its own login token from file or macOS Keychain against the OAuth usage API) and glm reports 99% (Z.ai's official monitor endpoint, same key as inference). kimi wired per its documented coding/v1/usages endpoint (validates when a key is added), qwen needs a pasted console cookie (Alibaba gates usage behind console auth — confirmed empirically that the inference key can't read it). codex/grok have no subscription-usage endpoint at all (even CodexBar doesn't poll ChatGPT limits) and stay on the cooldown fallback. Parsers are shape-tolerant with fixtures from the real payloads (34 tests).