fix(claude): surface redacted thinking and reasoning tokens

[zawakin]

Summary

First code fix from the v2.1 log-format audit (#57). Addresses the original symptom — thinking became invisible — plus reasoning-token accounting.

Adaptive thinking (Opus 4.7+) no longer persists plaintext thinking: blocks arrive as {"type":"thinking","thinking":"","signature":"…"}. The text is unrecoverable, so:

• Redacted-thinking marker: empty thinking with a signature now emits a Reasoning event with [thinking redacted] instead of a blank line. Blocks with neither text nor signature are dropped.
• Reasoning tokens: parse usage.output_tokens_details.thinking_tokens into the reasoning bucket (previously hardcoded to 0). Since output_tokens already includes thinking, generated = output_tokens - reasoning, keeping total() unchanged. Pre-v2.1 logs lack the field, so they behave exactly as before.

Changes

• claude/schema.rs: add OutputTokensDetail { thinking_tokens } to TokenUsage.
• claude/parser.rs: redacted-thinking marker; reasoning-token split; 3 new unit tests.
• Snapshots updated (additive output_tokens_details: null in raw output; no event-level change).

Notes

• Plaintext thinking cannot be restored — it is not in the logs. This makes the reasoning step visible, not its content.
• New record types and system subtypes are handled in a follow-up PR.

🤖 Generated with Claude Code