Summary

• Capture last_update_time and _storage_update_marker in DatabaseSessionService.append_event before commit(), instead of reading storage_session.update_time afterward
• Add regression test ensuring storage revision fields are not read after commit completes

Fixes #5761

Problem

With postgresql+asyncpg:// and default pool_pre_ping=True, post-commit ORM access to storage_session.update_time can lazy-load an expired column (due to onupdate=func.now()), triggering sqlalchemy.exc.MissingGreenlet during asyncpg pool pre-ping.

Fix approach

Follows maintainer guidance on #5761: read get_update_timestamp() / get_update_marker() before commit, then assign locals to the in-memory session.

Test plan

• uv run pytest tests/unittests/sessions/test_session_service.py::test_append_event_reads_storage_revision_before_commit -q
• uv run pytest tests/unittests/sessions/test_session_service.py -k "append_event or last_update_time" -q (26 passed)

Manual verification (asyncpg + Postgres)

• Env: postgresql+asyncpg://..., default pool_pre_ping=True, Postgres 16 (local docker compose)
• Fix branch: append_event × 100 → PASS (0 MissingGreenlet)
• main repro: post-commit ORM read of update_time after session.expire(..., ["update_time"]) → reproduces MissingGreenlet (lazy load → pool_pre_ping → asyncpg.ping())
• Fix pattern: reading revision fields before commit survives forced attribute expiry
• Note: a simple append_event loop alone did not naturally trigger attribute expiry on main; the repro requires the post-commit lazy-load path described above