fix(compression): flush messages before session rotation to prevent data loss

[x7peeps]

Problem

compress_context() in conversation_compression.py and new_session() in cli.py call end_session() without first flushing in-memory messages to the session DB. When auto-compression triggers mid-turn (context exceeds threshold), messages generated during the current turn that haven't yet been persisted are permanently lost:

• The old session ends with message_count from its last clean turn
• All messages from the turn that triggered compression are gone
• session_search, /resume, and offline backup all show a gap

This is especially painful for long tool-heavy turns (vision analysis, code execution, multi-step workflows) where 50+ messages can be generated in a single turn.

Fix

Add _flush_messages_to_session_db(messages) before end_session() in both sites:

1. agent/conversation_compression.py:compress_context() — flush before end_session when auto-compression rotates the session
2. cli.py:new_session() — flush before end_session when the /new command rotates the session

Both fixes use the existing _flush_messages_to_session_db() method which is already idempotent (tracks _last_flushed_db_idx) and has proper error handling.

Closes #47202

[alt-glitch]

Related: this is the flush-before-rotation fix for the mid-turn compression data-loss bug (#47202, #46567). Competing/complementary open PRs target the same root cause via different mechanisms and file-sets:

• fix(agent): reset flush cursor atomically during compression session rotation #41057 (earliest) — atomic flush-cursor reset in conversation_compression.py only
• fix(agent): flush messages before compression session rotation (#46567) #46587 — flush before rotation across conversation_compression.py + conversation_loop.py + run_agent.py

This PR (#47204) is the only one that also covers the /new command path (cli.py:new_session()) in addition to compress_context(). Maintainer to pick the canonical mechanism and consolidate.

[teknium1]

Closing as a duplicate of #47215, which fixed the same flush-before-rotate root cause and was merged via #48584. Both PRs landed the identical guard; #47215 was salvaged with its author's commit preserved. Thanks for the fix — same bug, credited to the merged commit.