feat(runtime): track Gemini thinking tokens in billing meter and session ledger

[dembrane-sam-bot]

What is this change?

This PR updates the billing meter and session ledger in src/runtime/ to track Gemini reasoning/thinking tokens (thoughts_token_count).
Specifically:

1. It updates _accumulate_usage in src/runtime/adk_runner.py to count thoughts_token_count into output_tokens (since Vertex bills thinking tokens as output tokens) and tracks them independently as thoughts_tokens.
2. It changes the deduplication key from a triple to a quad (prompt, output, thoughts, cache) to prevent dropped events on collision.
3. It plumbs thoughts_tokens into AgentRunResult, SessionResult, and SessionLedgerEntry so thinking token consumption is visible and audited in /data/sessions.jsonl.

What did Sam notice that led to this?

During our start-of-day token reconciliation pass, we noticed a consistent 3x-4x undercount on output tokens over the last 5 days when comparing our self-report ledger against Vertex AI's actual billed metrics (aiplatform.googleapis.com/publisher/online_serving/token_count grouped by type). For example, yesterday (2026-08-15) GCM reported 30,124 output tokens whereas our ledger recorded only 7,036.

By running a controlled test session with Runner.run_async(), we discovered that Gemini 3.5 Flash generates thinking/reasoning tokens (thoughts_token_count which is billed as output tokens). However, _accumulate_usage only read candidates_token_count and completely ignored thoughts_token_count. In our short test, thinking tokens (291) were 19x larger than candidates tokens (15).

Tier

Tier 3 (Runtime/Substrate).

Confidence

100% confident. We verified the Vertex AI billing behavior via controlled local testing, and our tests in tests/eval/ pass cleanly.

[spashii]

Reviewed and approved by Sameer. Rebased onto main; the prior CI failure was a stale .trivyignore.