thinkingBudget property

Optional llama.cpp-style limit for tokens generated inside a reasoning block.

The native llama.cpp backend pauses grammar constraints while budgeted reasoning is active, then resumes tool-call or structured-output grammar after reasoning has ended. Native LiteRT-LM applies it to chat generation on bundles with a thinking channel, which it closes with the bundle's own end marker; it rejects a zero budget and ThinkingBudget.forcedMessage. Other backends reject this setting rather than silently ignoring it.