fix(gateway): prevent session hygiene compression from overwriting original transcript

[annguyenNous]

Fixes #39704

Problem

When Gateway Session Hygiene triggers compression and the temporary AIAgent has _session_db=None, the original session's messages are permanently overwritten with compressed ones. This is a data loss bug — the uncompressed transcript is gone forever.

Root cause: Two code paths interact incorrectly:

1. agent/conversation_compression.py:501 — _compress_context gates the entire session-rotation block on if agent._session_db:. When it's None, no new session is created and session_id remains unchanged.

2. gateway/run.py:9277 — rewrite_transcript() executes unconditionally, outside the if _hyg_new_sid != session_entry.session_id block. When no rotation occurred, it overwrites the original session's messages.

Fix

gateway/run.py: Move rewrite_transcript() inside the session-rotation if block so it only runs when a new session was actually created. Add a log message in the else branch for the no-rotation case.

agent/conversation_compression.py: Add an else branch after the if agent._session_db: block that still rotates session_id even without a session DB. This ensures the gateway's existing if _hyg_new_sid != session_entry.session_id check works correctly, and the context engine notification fires with the correct old_session_id.

Verification

• With _session_db available: behavior unchanged — old session preserved, new session gets compressed messages
• With _session_db=None: session_id rotates, rewrite_transcript is NOT called, original transcript preserved, compressed messages used in-memory only

[teknium1]

Thanks @annguyenNous — this is a correct diagnosis of the gateway session-hygiene transcript-overwrite data-loss bug (same site you pointed at, gateway/run.py _hyg_* guard).

The same bug was fixed on main via #52658 (salvaging #50098), which reads the compaction result flag _last_compaction_in_place in the guard instead of gating on the id change. That approach also covers the /compress site and the in-place-success case (where the id is unchanged for a success reason, not a failure), which is why it was the version that landed alongside the compression.in_place default flip (#38763).

Your fix and the merged one solve the identical data-loss class. Closing as superseded — credit to you for the independent diagnosis. The transcript-overwrite path is now guarded on main.