feat: add cc.output_tokens trace-level aggregate for Sankey visualization

[jverre]

Details

Adds cc.output_tokens to the trace-level metadata, enabling the spend Sankey diagram to use sum(metadata.cc.output_tokens.by_category.*) directly without span-level aggregation.

Previously the output breakdown (thinking / assistant_text / builtin_tool_use / mcp_tool_use / skill_invocations) only existed on individual spans via cc.llm_call.attributed_output_tokens + cc.llm_call.block_kind. This is fine for span queries but blocks trace-level aggregations.

New field shape:
```json
"cc": {
"output_tokens": {
"summary": { "total_tokens": 1312 },
"by_category": {
"thinking": 364,
"assistant_text": 671,
"builtin_tool_use": 240,
"mcp_tool_use": 25,
"skill_invocations": 12
}
}
}
```

Classification rules:

• thinking — ContentType == "thinking"
• assistant_text — ContentType == "text"
• mcp_tool_use — ContentType == "tool_use" + name starts with mcp__
• skill_invocations — ContentType == "tool_use" + name == "Skill"
• builtin_tool_use — all other tool_use blocks

total_tokens equals sum(by_category.*) and matches the LLM call's actual output_tokens via the existing attribution logic (DeduplicateUsage).

Change checklist

• extractOutputTokensSnapshot added to extractors.go
• Wired into domainSnapshotsFromEntries in main.go (reuses the already-computed parsedTurn slice, no double-parse)
• TestDryRunOnTestThread domain list updated
• Unit tests added: happy path + nil-on-empty guard

Issues

N/A

AI-WATERMARK

AI assisted

Testing

go test ./... passes.

Documentation

N/A — internal metadata field; consumers are the Sankey FE component.