The gateway's update_session() used += for token counts, but the cached
agent's session_prompt_tokens / session_completion_tokens are cumulative
totals that grow across messages. Each update_session call re-added the
running total, inflating usage stats with every message (1.7x after 3
messages, worse over longer conversations).
Fix: change += to = for in-memory entry fields, add set_token_counts()
to SessionDB that uses direct assignment instead of SQL increment, and
switch the gateway to call it.
CLI mode continues using update_token_counts() (increment) since it
tracks per-API-call deltas — that path is unchanged.
Based on analysis from PR #3222 by @zaycruz (closed).
Co-authored-by: zaycruz <zay@users.noreply.github.com>
22cfad157b