fix: persist TUI session undo rewinds

[0xBlocky]

Summary

• make the TUI session.undo RPC persist rewinds through SessionDB.rewind_to_message
• reload the active-only transcript after undo so live context and durable resume state match
• preserve the previous in-memory fallback for transient sessions without a persisted DB row

Why

The TUI undo path could report that messages were undone while only trimming the in-memory session history. After exiting and resuming the session, the supposedly undone turn reappeared because the rows were still active in state.db.

Tests

• HERMES_PYTHON=/usr/bin/python3 scripts/run_tests.sh tests/tui_gateway/test_undo_command.py
• git diff --check

[teknium1]

Thanks for this — you correctly identified that the TUI session.undo RPC only mutated in-memory history, so a restart/resume brought soft-undone turns back.

This landed on main via PR #93784 (the composite-carrier retry/undo salvage, merged today): session.undo now routes through _rewind_active_session_history(), which archives the target row and tail durably in SessionDB in the same transaction and reinstalls the active view. I verified by running this PR's own test file — including test_session_undo_persists_rewind_for_resume — against current main without your code change: 3/3 pass.

Closing as implemented on main. Appreciate the report and the tests — sorry the broader carrier rework got there first.