ENG-3027 Scrub OpenCode AI SDK payloads before Langfuse export

[zigelboim-misha]

Summary

• Add a span processor wrapper around LangfuseSpanProcessor that strips OpenCode/Vercel AI SDK prompt and response payload attributes before export.
• Preserve model, usage, timing, finish reason, and gen_ai attributes so cost and performance analytics still work.
• Add a focused test proving payload attributes are removed and usage/model attributes remain.

Duplicated Data Addressed

OpenCode's AI SDK spans were exporting these large payload attributes into Langfuse:

• ai.prompt
• ai.prompt.messages
• ai.prompt.tools
• ai.prompt.toolChoice
• ai.response.text
• ai.response.toolCalls
• ai.response.object

These duplicate snakur/OpenCode request and response content and were a major contributor to Langfuse ClickHouse/MinIO growth.

Related PRs

• snakur wrapper payload removal: https://github.com/dalton-hq/snakur-2/pull/1165
• discovery-code plugin pin: https://github.com/dalton-hq/discovery-code/pull/120

Verification

• bun test
• bun run typecheck
• bun run lint
• bun run format:check

Linear: https://linear.app/dalton-ai/issue/ENG-3027

[zigelboim-misha]

Closing per scope change: we are not changing the Langfuse plugin for ENG-3027. The active fix will be limited to snakur-2 wrapper payload capture.