Streamed Anthropic generations stack output tokens when a stream emits more than one message_delta #1099

Version

7.75.0 (current master, includes #1074)

Steps to Reproduce

1. Stream an Anthropic response whose stream emits more than one message_delta event. This happens with server tools such as web search or code execution, which pause and resume the turn inside one stream.
2. Each message_delta carries cumulative usage, but the accumulator in posthog/ai/anthropic/_anthropic_stream.py merges output_tokens incrementally, so successive deltas stack.

Minimal repro:

from types import SimpleNamespace as NS
from posthog.ai.anthropic._anthropic_stream import _AnthropicStreamAccumulator

acc = _AnthropicStreamAccumulator()
acc.consume(NS(type="message_start", message=NS(usage=NS(input_tokens=100, output_tokens=1, cache_creation_input_tokens=0, cache_read_input_tokens=0))))
acc.consume(NS(type="message_delta", usage=NS(output_tokens=50)))
acc.consume(NS(type="message_delta", usage=NS(output_tokens=110)))
acc.consume(NS(type="message_stop"))
print(acc.usage_stats["output_tokens"])

Ran it twice, identical both times: 160

Expected Result

110, the final cumulative total. The Anthropic SDK accumulates stream usage by overwriting on each message_delta ('Usage counts on a message_delta are cumulative totals, so they overwrite rather than add', anthropic-sdk-python lib/streaming/_messages.py). #1074 fixed the same stacking for input and cache counts; output_tokens still stacks.

Actual Result

160 (50 + 110 stacked). Streamed generations over-report output tokens whenever a stream contains more than one message_delta.