feat(mem0): eval sets by Bob subagents + local reranker run

[ana-lan]

Bob B12 wrote 51 eval cases for 3 mem0 features with 3 parallel subagents. Local reranker run on Qwen tiers: 7b 71%, 3b 62%, 1.5b 50%, 0.5b 50%, so no safe downgrade (3b keeps 88% < 95%); failures cluster in the partial-relevance band. Analysis cost $0.29. Report fix: clear sentence when there is no payback. Bob B13 updated the case-study doc and web (one invented claim corrected by hand); a README edit was undone with Bob's rollback.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
downshift	Ready	Preview	Sep 27, 2026 1:52am UTC

[github-actions]

Downshift cost diff

No LLM call site cost changes in working tree vs origin/main (0 added, 0 changed, 0 removed, 7 unchanged). Projected monthly cost stays at $7,339.50.

Projection, not a bill: tokens estimated from source (words in resolved prompt text; output = max_tokens, or 256 if unset) x illustrative prices x configured calls/day x 30 days.

[ana-lan]

Bob built-in code review (/review, 34 tools, 1.98 coins) found 3 issues, all fixed by hand: rr-11 and rr-19 had wrong expected bands (documents that directly answer the query), and the report printed 'No payback' when payback was never computed. Only the 2 relabeled cases were re-run. Final: 7b 17/24, 3b 13/24, 1.5b 13/24, 0.5b 11/24; best candidates keep 76% of baseline, no safe downgrade.