用 last-call usage 修正 compaction 后 totalTokens 偏高（PR #15018）

• Gateway version includes PR #15018 and can be restarted during a maintenance window.
• You can reproduce token inflation after auto-compaction or manual /compact.

1. Record a baseline by running one tool-heavy conversation and checking current totalTokens value.
2. Upgrade to a build containing PR #15018, then restart gateway.
3. Trigger manual /compact once and run another tool-heavy turn that can auto-compact.
4. Compare persisted session totalTokens before/after compaction and ensure it tracks latest-call usage, not accumulated usage.
5. Keep accumulated usage for billing/cost views only and avoid using it for context-window health checks.

• If your pipeline depends on custom usage middleware, verify it does not overwrite last-call usage fields（需验证）。
• This fix targets token display/accounting; it does not reduce actual model token spend by itself.

This tip is aggregated from community/public sources and preserved with attribution.