fix: gateway token double-counting — use absolute set instead of increment

[teknium1]

Summary

The gateway's token usage stats inflate with every message because update_session() uses += for cumulative values the agent already tracks.

Root cause

The cached agent's session_prompt_tokens / session_completion_tokens are running totals that accumulate across run_conversation() calls. The gateway reads these cumulative values and passes them to update_session(), which does entry.input_tokens += cumulative. Each message re-adds the full running total.

Live test result (before fix):

After msg 1: input=1000 → stored 1000 ✓
After msg 2: input=2500 → stored 3500 (should be 2500) ✗
After msg 3: input=5000 → stored 8500 (should be 5000) ✗
Token inflation: 1.7x after 3 messages

Fix

• gateway/session.py: Change += to = for input_tokens, output_tokens, cache_read/write_tokens, estimated_cost_usd
• hermes_state.py: Add set_token_counts() — uses SQL direct assignment (input_tokens = ?) instead of increment (input_tokens = input_tokens + ?)
• gateway/session.py: Switch DB call from update_token_counts to set_token_counts

CLI mode continues using update_token_counts() (increment) since it tracks per-API-call deltas.

Live test result (after fix):

After msg 1: input=1000 → stored 1000 ✓
After msg 2: input=2500 → stored 2500 ✓
After msg 3: input=5000 → stored 5000 ✓
Token inflation: 1.0x

Validation

• python -m pytest tests/test_hermes_state.py tests/gateway/test_session.py -n0 -q → 186 passed

Based on analysis from PR #3222 by @zaycruz.
Co-authored-by: zaycruz zay@users.noreply.github.com