fix(ai): preserve Anthropic cache-write TTL breakdowns

[gouveags]

Note

This PR depends on PostHog/posthog#71323 and will remain a draft until the server-side event contract is merged.

💡 Motivation and Context

PostHog/posthog#71323 adds separate costing for Anthropic's 5-minute and 1-hour cache writes. The Python SDK already preserves that breakdown in its native Anthropic wrappers, but two integrations drop it before ingestion.

Fixes #745. Related to PostHog/posthog#71306. Depends on PostHog/posthog#71323.

This change:

• Makes the LangChain callback emit the aggregate cache-write count and a complete 5-minute/1-hour property pair. LangGraph uses the same callback path.
• Preserves raw usage in Claude Agent stream events and result-message fallback, matching the native Anthropic integration.
• Treats manually supplied TTL token properties as passthrough token data.
• Adds regression coverage for native Anthropic sync, async, streaming, and the shared Anthropic Bedrock wrapper path.

The PostHog Bedrock and Vertex clients reuse the native Anthropic message adapters. Aggregate-only events remain unchanged; invalid or empty TTL details fall back to the aggregate, and a valid breakdown replaces rather than adds to it.

Direct boto3 calls are outside this PR because posthog-python has no boto3 Bedrock wrapper. The OTel integration forwards spans through a separate endpoint and is not changed.

💚 How did you test it?

• Full suite: 1629 passed, 59 skipped.
• Focused AI/token suite: 135 passed, 9 skipped.
• Ruff format and lint.
• Mypy across 172 source files.
• Public API snapshot and import-warning checks.
• Package build and Twine validation.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed. The server properties are documented in fix(aio): price Anthropic cache writes by TTL posthog#71323.
• No breaking change or entry added to the changelog.

If releasing new changes

• Added a Sampo patch changeset.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

I used Codex to inspect the server contract, audit the Python SDK integration paths, implement the focused changes, and run independent test, code, and writing reviews. No public session link is available.

The first pass focused on the LangChain callback. The broader audit showed that native Anthropic wrappers already preserve raw usage, while Claude Agent needed the same raw-usage path. The final diff reuses those existing mechanisms instead of adding provider pricing or a new integration layer.

[gouveags]

@greptile review

[greptile-apps]

_{Reviews (1): Last reviewed commit: "fix(ai): preserve Anthropic cache-write ..." | Re-trigger Greptile}