fix(core): fix setting ChatGeneration.text

[ccurme (ccurme)]

We previously had a hack where we'd take the first text block.

This isn't suited to model calls nowadays, where we generally expect multiple text blocks in a single invocation. For example when using Opus 4.6 with adaptive thinking, we can get output of the form:

[
  {
    "text": "\n\n",
    "type": "text"
  },
  {
    "signature": "...",
    "thinking": 'The user said "Hello"...',
    "type": "thinking"
  },
  {
    "text": '{"response":"Hello! How can I help you today?"}',
    "type": "text"
  }
]

(Note: I verified that the raw response from Anthropic in the above case genuinely contains TextBlock(citations=None, text='\n\n', type='text') as the first content element.)

This was mentioned in a comment here.

The impact here is that output parsers do not access the full text content, only the first block. So structured output can fail.

[Mason Daugherty (mdrxy)]

the old code accepted dict blocks without a type key

the message.text property requires block.get("type") == "text" explicitly, so blocks missing a type key are now excluded

is this a problem?