fix(compression): prevent orphan sessions when state.db write fails (#33906)

[zccyman]

Summary

Prevent orphan sessions when context compression's create_session() call fails (e.g. transient SQLite lock contention). The agent previously continued with a new session_id but no state.db row — producing an orphan session invisible in WebUI and /sessions.

Closes #33906

Problem

When context compression triggers a session rotation in conversation_compression.py:

1. Old session is ended (end_session())
2. New session_id is generated and immediately assigned to agent.session_id
3. create_session() writes the new row to state.db

If step 3 fails (SQLite lock, disk I/O error), the agent continues with a session_id that has no state.db entry. The session is invisible in:

• WebUI sidebar (/api/sessions returns only state.db rows)
• /sessions CLI command
• Session resume (/resume can't find it)
• Usage tracking (token counts lost)

The entire try/except block only logs a warning — no rollback of the mutated session_id.

Solution

Three changes in agent/conversation_compression.py:

1. Reorder assignment: generate new_session_id first, but only assign it to agent.session_id after create_session() succeeds
2. Retry with backoff: 3 attempts for sqlite3.OperationalError (lock contention) with 0.1s × attempt delay
3. Rollback on failure: restore agent.session_id to old value and reopen_session() the old session so the agent stays on a registered session

Edge case handled: if end_session() itself fails (before old_session_id is assigned), rollback uses locals().get() instead of direct reference to avoid UnboundLocalError.

Files Changed

File	Change
agent/conversation_compression.py	+46/-9: reorder, retry, rollback
tests/agent/test_compression_orphan_session.py	+217: 6 new tests

Test Results

tests/agent/test_compression_orphan_session.py ...... (6 passed)
tests/agent/test_context_compressor.py .............. (129 passed)
tests/gateway/test_compress_*.py .................... (9 passed)
ruff check .......................................... All checks passed!

Design Decisions

• Retry 3 times (not infinite): SQLite lock contention is transient; 3 attempts with backoff covers >99% of cases without blocking the agent
• Rollback + reopen (not just skip): if we can't create the new session, we must keep the old one alive — end_session() already marked it as ended, so reopen_session() is required
• locals().get() guard: avoids UnboundLocalError when end_session() fails before old_session_id = agent.session_id is reached

[teknium1]

Merged via #49772 — your authorship is preserved as a co-author trailer on the squashed/rebased commit. This combined cluster fix incorporates the orphan-session rollback (failed child create_session) from this PR. Thanks for the contribution! #49772