fix(ai): take cumulative Anthropic stream usage from message_delta

[pcbeingused333]

💡 Motivation and Context

Anthropic's message_delta usage is cumulative. With server tools (web search, web fetch, code execution), its input_tokens and cache counts grow past the message_start values during the response. The Anthropic stream accumulator only recorded input and cache counts from message_start (it reads output_tokens from message_delta), so streamed generations under-reported $ai_input_tokens and $ai_cache_creation_input_tokens.

Example from a recorded web-search stream: message_start has input 2247 and cache creation 0, while message_delta has input 2254 and cache creation 17269. Before this change the event reported 2247 and 0.

_AnthropicStreamAccumulator.consume, shared by the sync and async wrappers, now applies the message_delta input and cache counts after the per-event merge. It keeps the larger of the two values, since the totals never shrink, and leaves a field the event omits at the message_start value.

💚 How did you test it?

• New test_streaming_takes_cumulative_message_delta_usage and its _async twin. They use real anthropic.types stream events with the values above and check the captured $ai_input_tokens, $ai_output_tokens and cache properties. Both fail on main.
• posthog/test/ai: 951 passed, 15 skipped; on main, only the two new tests fail.
• ruff format --check passes. ruff check reports the same findings as on main for the touched files. mypy … | mypy-baseline filter reports nothing new.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed. (n/a)
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file (added .sampo/changesets/cumulative-anthropic-stream-usage.md in the same format)

🤖 Agent context

Autonomy: Fully autonomous

• The bug was found and the fix, tests and description were written by Claude Code; I reviewed the diff and the explanation before submitting.

• Written with Claude Code (Claude Opus 5.5).

• The fix keeps the larger of message_start and message_delta for each count, rather than switching the merge to "cumulative" mode. That mode would replace the merged raw_usage and lose message_start's fields, and max also tolerates a delta that omits or under-reports a field.

🤖 Generated with Claude Code