deadair 0.70.2

• A new setting, "How hard the production model thinks" (llm.productionReasoning), sets the reasoning effort for writing and re-drafting each turn of a production, the way llm.breakReasoning already does for breaks. It defaults to A little, which is what every turn asked for until now. On a model that thinks at length even then, every phone-in and co-host banter failed as "came back empty twice" because one turn ran out of time or room before it said anything. Not at all fixes that.