fix(gateway): preserve transcript on /compress and hygiene auto-compress

[SandroHub013]

Closes #21301.

Problem

Manual /compress and Gateway Session Hygiene auto-compress both physically destroy the original transcript. Reproduction (per the issue):

1. Telegram/Discord gateway, 60+ message session.
2. Send /compress.
3. state.db shows the original session's message_count collapsed to the compressed count, end_reason is NULL, parent_session_id is NULL, no continuation session was created.

Documented behavior (sessions.md "Auto-Lineage on Compression") says the original should be closed with end_reason='compression' and a new continuation session should carry parent_session_id pointing back.

Root cause

Both code paths build a temporary AIAgent for the compression run:

• _handle_compress_command at gateway/run.py:9445 (tmp_agent = AIAgent(...))
• Session Hygiene at gateway/run.py:6218 (_hyg_agent = AIAgent(...))

Neither passes session_db=. That makes tmp_agent._session_db = None, and _compress_context()'s session-rotate block is guarded by if self._session_db: (run_agent.py:9048), so it silently no-ops. tmp_agent.session_id never rotates.

The downstream safeguard added in 1544638 / cd2e180:

new_session_id = tmp_agent.session_id
if new_session_id != session_entry.session_id:
    session_entry.session_id = new_session_id
    self.session_store._save()
self.session_store.rewrite_transcript(new_session_id, compressed)

assumes the rotate already happened. With _session_db is None it didn't, so new_session_id == session_entry.session_id and rewrite_transcript overwrites the original.

Fix

Add session_db=self._session_db to both constructors:

 _hyg_agent = AIAgent(
     **_hyg_runtime,
     model=_hyg_model,
     max_iterations=4,
     quiet_mode=True,
     skip_memory=True,
     enabled_toolsets=["memory"],
     session_id=session_entry.session_id,
+    session_db=self._session_db,
 )

 tmp_agent = AIAgent(
     **runtime_kwargs,
     model=model,
     max_iterations=4,
     quiet_mode=True,
     skip_memory=True,
     enabled_toolsets=["memory"],
     session_id=session_entry.session_id,
+    session_db=self._session_db,
 )

Same shape as PR #20021 (ACP adapter) and PR #4802 (API server adapter), which fixed identical omissions in their respective tmp-agent constructors.

Tests

New: tests/gateway/test_compress_preserves_session_db.py — two static-source assertions that both constructor call sites carry session_db=self._session_db. Avoids the SQLite + full AIAgent wiring a true integration test would need, which would be heavier than the bug. The reporter has independently validated end-to-end behavior in production (Telegram /compress on a 112-message session — original preserved, new session linked).

python -m pytest -o addopts= tests/gateway/test_compress_preserves_session_db.py -q
2 passed

Out of scope

• The ACP / API server fixes referenced above. Already shipped.
• Refactoring _compress_context() to no longer require session_db. The if self._session_db: guard exists for non-gateway callers (one-shot CLI, etc.) and removing it would change behavior beyond this fix.

[teknium1]

Merged via #49925 — your authorship is preserved as a co-author trailer on the merge. This fix incorporates the hygiene/compress transcript-preservation guard from your PR. Thanks for the contribution! #49925