test(memory): benchmark current private-memory recall and define adoption gates

[robinbraemer]

Outcome

Create a reproducible, privacy-safe retrieval evaluation and baseline the existing LLM topic selector before changing memory recall.

Scope

• Build synthetic and sanitized topic corpora with explicit relevant/irrelevant judgments, contradictions, recency, aliases, sparse index hooks, and multilingual cases.
• Measure selector recall@k, precision@k, attachment bytes/tokens, latency, provider calls/cost, and downstream task success.
• Exercise pinned topics, attachment limits, malformed topics, pause, correction, forget, and selector failure.
• Define minimum improvement and maximum resource-regression gates for lexical and hybrid candidates.

Acceptance criteria

• Fixtures contain no real Mate memory, credentials, transcripts, or hidden reasoning.
• Results are reproducible across the pinned model/configuration or explicitly report variability.
• The baseline exposes false-positive, false-negative, latency, cost, and failure distributions.
• Adoption gates are numerical and cannot be satisfied merely by reducing prompt tokens.
• The suite can compare current, lexical, vector, hybrid, and fallback paths through one public contract.