[codex] Add on-prem graph memory backend

[heikog]

What changed

• Adds a graph-memory adapter layer with Zep Cloud as the default backend.
• Adds a Graphiti bridge adapter for on-premise graph memory.
• Adds a separate Graphiti Bridge service backed by FalkorDB so Graphiti dependencies do not conflict with OASIS.
• Adds an optional docker compose profile for graphiti-falkordb and graphiti-bridge.
• Mirrors Reddit OASIS trace rows into reddit/actions.jsonl so run status, timelines, and graph-memory updates receive real simulation activity.
• Makes report status lookup work consistently by task_id, report_id, or simulation_id, including GET query parameters.
• Makes interview history sorting robust across mixed SQLite timestamp formats emitted by different simulation platforms.
• Documents the on-premise graph-memory setup.

Why

Zep Cloud can be expensive for self-hosted deployments. This keeps existing Zep behavior as the default while adding an on-premise path that can be enabled with GRAPH_MEMORY_BACKEND=graphiti_bridge.

The API fixes are included because the full on-premise end-to-end validation exposed two integration issues outside the adapter itself: Reddit simulation activity was not visible to graph-memory updates before trace mirroring, and completed report/interview endpoints had edge cases that made non-GUI verification unreliable.

Validation

• python3 -m compileall -q backend/app graphiti_bridge backend/scripts/run_reddit_simulation.py
• python3 -m py_compile backend/app/services/simulation_runner.py backend/app/api/report.py
• git diff --check
• Direct adapter smoke test on Paul using GraphitiBridgeGraphMemoryAdapter against graphiti-bridge/FalkorDB.
• Manual full API end-to-end test on Paul with GRAPH_MEMORY_BACKEND=graphiti_bridge:
  • project proj_40134302015f, graph mirofish_11dcdf61aa4c4eb7, simulation sim_d2b03920d048, report report_33bf2737c50d
  • ontology generation produced 10 entity types and 9 relationship types
  • initial graph build produced 24 nodes and 6 Graphiti/Falkor edges
  • profile preparation produced 16 profiles
  • parallel Reddit + Twitter simulation completed 2/2 rounds with 16 platform actions
  • Reddit and Twitter action logs were both written
  • graph-memory update increased Graphiti/Falkor edges from 6 to 11 after simulation activity
  • interview batch completed and wrote a response file
  • report generation completed with title 2030 Market Debate: How AI Reshapes the Software Developer Market
  • report chat, report retrieval, report sections, report checks, report download, and simulation config download returned successfully
• Targeted regression checks after the API fixes:
  • POST /api/simulation/interview/history returns HTTP 200 with 20 interview records from the AI developer market simulation
  • GET /api/report/generate/status?report_id=report_33bf2737c50d returns HTTP 200, completed, progress 100
  • POST /api/report/generate/status with report_id returns HTTP 200, completed, progress 100
  • POST /api/report/generate/status with simulation_id returns HTTP 200, completed, progress 100