fix: count actual tool calls instead of tool-related messages

[0xbyt4]

Summary

tool_call_count in the session database was inaccurate due to two bugs in append_message():

Bug 1 — Under-counting parallel tool calls: An assistant message with N parallel tool calls (e.g. "kill the light and shut off the fan" = 2 ha_call_service calls) only incremented tool_call_count by 1 instead of N.

Bug 2 — Over-counting tool responses: Tool response messages (role=tool) also incremented tool_call_count, double-counting every tool interaction.

Combined effect: 2 parallel tool calls produced tool_call_count=3 (1 from assistant + 2 from tool responses) instead of the correct value of 2. This makes /insights and /usage report inflated tool call numbers.

Fix: Only count from assistant messages that carry tool_calls, incrementing by len(tool_calls). Tool response messages no longer affect the counter.

Test plan

• test_tool_call_count_matches_actual_calls — 2 parallel tool calls + 2 tool responses = count should be 2 (fails on old code with 3)
• test_tool_response_does_not_increment_tool_count — tool response alone should not increment count
• test_assistant_tool_calls_increment_by_count — single tool call increments by 1

uv run pytest tests/test_hermes_state.py tests/test_insights.py -q  # 108 passed

[teknium1]

Merged to main — thanks @0xbyt4! Good catch on the double-counting from tool responses and under-counting of parallel calls.