This branch makes `stream_gen_ai_spans` opt-out (default `True`) via `should_truncate_gen_ai_input` in `tracing_utils.py`. With the new default, `should_truncate_gen_ai_input` returns `False`, so integrations use the raw normalized messages instead of routing them through `truncate_and_annotate_messages`. Because `redact_blob_message_parts` is only invoked inside `truncate_and_annotate_messages`/`truncate_and_annotate_embedding_inputs`, the unconditional blob redaction (replacing base64 image/blob content with `[Filtered]`) is now bypassed on the default code path. There is no equivalent redaction in the streaming (`traces.py`) path, so multi-modal prompt blobs are sent unredacted to Sentry whenever prompt capture is enabled (`send_default_pii=True` and `include_prompts=True`). This redaction was previously unconditional (applied even with PII enabled), so coupling it to the now-disabled truncation appears to be an unintended side effect.