v0.3.3

v0.3.3 — Mid-stream fallback + gpt-5.6 cache keep-warm

• Reroute when an account hits its limit mid-generation (WebSocket). If the active account's quota is exhausted partway through a response, the request now fails over to a healthy fallback account on the same turn, instead of failing on the exhausted one. This applies before any output has streamed; once a response has started, the turn completes and the account is avoided on the next turn (so nothing is billed or run twice).
• gpt-5.6 prompt-cache keep-warm. Keep-warm now matches gpt-5.6's ~30-minute prompt cache (vs ~5 minutes on older models), with a bounded warm cap for short-lived subagent sessions and an optional /openai-cachekeep HH-HH clock window to limit warming to set hours.