fix(ai): preserve Anthropic usage metadata - #38751

Problem

Anthropic usage events were decoded with a closed Effect Schema.Struct, so every provider field outside four token counters was silently stripped before usage mapping. This dropped current fields such as server-tool usage and output_tokens_details.thinking_tokens, and made the documented raw provider-metadata escape hatch lossy for future Anthropic additions.

Solution

• Decode Anthropic usage with Schema.StructWithRest plus Record<string, unknown> so unknown top-level fields survive untyped.
• Apply the same open shape to known nested server_tool_use and output_tokens_details objects so future nested counters also survive.
• Keep known fields typed:
  • server_tool_use.web_search_requests
  • output_tokens_details.thinking_tokens
• Map thinking_tokens to canonical Usage.reasoningTokens while keeping outputTokens inclusive.
• Preserve reasoning tokens and raw provider fields when merging usage split across message_start and message_delta.

Tests

The Anthropic stream fixture verifies:

• canonical input, output, total, and reasoning token counts;
• known server-tool usage;
• unknown top-level usage fields;
• unknown nested server-tool and output-detail fields;
• merging raw fields from initial and terminal usage events.

Verification

• bun test from packages/ai: 452 passed, 28 skipped
• focused Anthropic suite: 44 passed
• bun typecheck from packages/ai: clean
• git diff --check: clean