fix: use last streaming entry for output token count

[GuyMoses]

Summary

• The transcript requestId deduplication was taking the first streaming entry per API call, which carries a partial output_tokens count (1–7 tokens from the streaming start). The last entry has the final count.
• Verified against a real transcript: 132 out of 847 requestIds had differing usage across streaming entries. The first entry consistently had output_tokens: 1 while the last had the real value (100–1280).
• Impact: ~15% undercount of output tokens (224,818 reported vs 263,699 actual across one session). Input tokens and cache tokens were unaffected.
• Fix: changed from skip-if-seen to last-write-wins — store the last entry per requestId in a map, then sum across all requestIds after reading the transcript.

Test plan

• TestReadTurnUsageDeduplicatesRequestID updated to use different output_tokens values across streaming entries (1 vs 150) and assert the final value is used
• All existing transcript tests pass (single assistant, multi-iteration aggregation, turn reset, tool_result handling, pretty-printed JSON, partial fields)
• Full test suite passes (go test ./...)
• Deploy and compare output token totals against native claude_code.token.usage counter

[GuyMoses]

Plugin vs Native Token Gap — Investigation Summary

After fixing the streaming dedup bug in this PR, I investigated the remaining token gap between the plugin and native Claude Code telemetry.

Root cause

The transcript file does not record token usage for background API calls. Two types are missing:

Background call	Frequency	Model	Token impact	Cost impact
Title generation (ai-title entries)	~35/session	Haiku	~14K input, ~350 output	~$0.02 (negligible)
Context compaction (away_summary entries)	~10/session	Main model (Opus)	~1M input (mostly cache reads), ~3K output	~$0.50–1.50

Native tracing captures both as claude_code.llm_request spans with full token counts. The plugin only reads assistant transcript entries, which don't include these calls.

Magnitude

On a real 29-turn Opus session ($19.30 total):

• Background gap: ~$0.50–1.50 (~3–8% of session cost)
• Compaction dominates; title gen is <0.1%
• Gap grows with session length (more compaction events)