v0.4: eval suite for Cognee graph extraction quality (per (route × model))

[thinmintdev]

Deferred from v0.3 per ADR-0014 §4 and audit gap G2 in docs/internal/audit-2026-05-22-phase8-skill-review.md.

Scope

Build the eval suite that measures graph-extraction quality for each (route × model) pair so the v0.3 "configurable model" gate can graduate from "configurable + caveat" to "configurable + measured" — and so future versions can auto-flip the default route when a local model demonstrably passes the bar.

• Standard corpus (held-out documents covering technical + casual + multi-entity passages).
• Metrics: entity recall, relation precision, schema-violation rate, structured-output reliability.
• Per-route runner (upstream / primary / agent slot).
• Output report consumable by docs (docs/memory/graph.md) + dashboard.

When this lands

• Once landed, ADR-0014 §4 caveat copy ("Graph quality varies by model. We don't currently measure it for you — your results may vary.") can be replaced with a per-model quality readout.
• Could enable an auto-default-on path for proven model families (additive; no schema migration).

Tag

v0.4 (no v0.3 label).