✨ WP2: update-integrity benchmark — supersession scenario suite

[Wuesteon]

Summary

Executes the WP2 packet (lane B, ungated): a small, honest, reproducible test of the core claim — when a fact changes, the engine returns the current truth and keeps the old one queryable.

• bench/update_integrity.py: ten scripted scenarios through the public API only (Memory.add → Memory.search) — employer change, identity change (mem0#4896's class of case), city move, preference flip, additive "also" EXTENDS, replacement-after-additive retire-all, multi-valued preservation, as-of-before-everything, restart persistence, verbatim-restatement no-dup (WP11 pin). Per scenario: top-1 is the new fact, superseded fact has is_latest=False + superseded_by, as_of=<t before update> returns the old fact.
• --markdown renders the results table from a single command (26/26 assertions PASS, exit 0); table pasted into the new appendix in docs/competitive-landscape.md.
• Same scenarios pinned offline as tests/test_update_integrity_scenarios.py (bench-import via the test_phase2_* sys.path precedent; excluded from the sdist like its peers).
• Optional mem0 comparison arm is designed in the plan but deliberately not implemented (opt-in per packet).
• Zero src/ changes (git diff --stat main -- src/ is empty).

Test plan

• Core suite 301 passed (289 + 12 new); console suite 153 passed.
• PYTHONPATH=src .venv/bin/python bench/update_integrity.py --markdown → ALL PASS, exit 0.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YRU73aDWiYUqgaL5UETUMp