Record Anthropic thinking tokens from the raw response (litellm reports 0)

litellm 1.88.1 estimates Anthropic reasoning tokens by token-counting the visible thinking text instead of reading the API's usage.output_tokens_details.thinking_tokens; Claude Opus 5 / Fable 5 / Sonnet 5 default to display: "omitted" (empty thinking text), so completion_tokens_details.reasoning_tokens is structurally 0 for them. Filed upstream with a proposed fix: BerriAI/litellm#36290.

Until that lands, our per-row usage recording (and the published usage_summary.csv) shows zero reasoning tokens for every Claude row regardless of actual thinking. Options:

• Prefer the raw provider usage when available: read output_tokens_details.thinking_tokens off the response and record it in our reasoning column directly, falling back to litellm's field for other providers.
• Regenerate the affected analysis CSVs for runs where thinking actually occurred (the tool_choice-auto sensitivity run is the first).

The board's canonical runs are unaffected in substance (Claude genuinely produced no thinking under the forced tool — see sensitivity/claude-thinking-2026-08.md), but the column should stop being wrong-by-construction before any v2 run where Claude thinking engages (#139).

🤖 Generated with Claude Code

Start by locating the per-row usage recording and the code that writes reasoning tokens to usage_summary.csv, then inspect how the raw response usage is represented. Verify the change against the affected analysis CSVs, especially the tool_choice-auto sensitivity run, and confirm that provider responses with thinking report the expected reasoning count while other providers retain their existing fallback.