fix: preserve user focus after todo compression

[ykq]

Summary

• Prevents context compression from appending the preserved todo snapshot as a standalone latest user message.
• Labels the todo snapshot as an internal state note and prepends it into the latest real user turn instead.
• Preserves assistant messages unchanged, including signed-thinking/reasoning-bearing assistant content.

Test plan

• ./venv/bin/python -m pytest -o addopts='' tests/agent/test_compression_todo_injection.py tests/tools/test_todo_tool.py -q → 17 passed
• ./venv/bin/python -m pytest -o addopts='' tests/agent/test_context_compressor.py tests/agent/test_context_compressor_summary_continuity.py tests/run_agent/test_compression_boundary.py tests/agent/test_compression_todo_injection.py tests/tools/test_todo_tool.py -q → 108 passed

Notes

This fixes a compression-focus regression where internal continuity state could be interpreted as the newest user request after compaction.

[alt-glitch]

Related: This competes with #26981 and #18883, all addressing #26979 (todo snapshot injected as synthetic user message after compression). This PR takes the approach of prepending the snapshot into the latest real user turn rather than removing the injection entirely (#26981) or stripping subjects (#18883). Maintainers should pick one approach.