[mattt]

Several README statements and doc comments don't match the code ahead of 1.2.0. This change corrects what stopped responses and streams return, token usage, tool calling on Core ML, the requirements for tvOS, watchOS, WASI, and Swift 6.4, the models that ignore maximumToolRounds or skip tool callbacks, transcriptErrorHandlingPolicy, and Anthropic thinking. It also documents that some session property misuse stops the process where Foundation Models 27 ignores it, how images in tool outputs behave on each provider, that a profile should create its model outside body, and how edits to a replayed turn take precedence over stored text, with structured responses as the limit. The replay text depends on #401, and the tvOS statement depends on #400.