fix(bot): keep memory-commit side effects out of the LLM-visible answer loop

[r266-tech]

Fixes #3082.

Summary

• Treat openviking_memory_commit as a side-effect tool in the agent loop.
• Execute and keep it in tools_used/logs, but do not add its tool result to the LLM-visible reflection transcript.
• For memory-only turns, run a no-tool finalization turn so the model answers from conversation context instead of reflecting on commit status.
• For mixed turns, reflect only visible information-tool results and suppress memory-commit tool/result events from the outbound bus.

Notes

• tools_used still retains the memory commit call/result for traceability, matching the issue's expectation that side effects remain recorded outside the LLM transcript.
• Once a side-effect tool has executed in a request, the loop keeps it disabled for later iterations to avoid duplicate commits in the same user turn.
• A forced no-tool finalization response that still contains tool calls is ignored with a warning and falls back to the previous assistant content when available.

Verification

• /tmp/ov-3082-venv/bin/python -m py_compile bot/vikingbot/agent/loop.py bot/tests/test_agent_loop_outcome.py
• git diff --check
• Direct control-flow harness covering memory-only, empty-content, failed structured status, mixed success, mixed failure, stop-tool order, and side-effect event suppression.

/tmp/ov-3082-venv/bin/python -m pytest bot/tests/test_agent_loop_outcome.py -q did not reach test collection in this sparse local checkout because bot/tests/conftest.py imports the full OpenViking service stack and the local venv is missing the declared volcengine SDK dependency.