recency_boost_newer_memory_ranks_higher test panic — tie-break at equal scores

[Liorrr]

Bug

Unit test echo::tests::recency_boost_newer_memory_ranks_higher panics with:

Newer memory (Meta, score=0.771923) should rank higher than older (Google, score=0.771923)

Two memories with different ages produce identical final_score values, causing the assertion to fail. The recency boost is not creating sufficient separation between memories of different ages.

Context

Found by QA during KS74 micro-benchmark baseline run. Pre-existing — not introduced by KS74.

The KS74 LME temporal-reasoning benchmark scored 0/5 due to recency boost miscalibration (recent distractors override semantic similarity). This tie-break bug may share the same root cause — the ACT-R activation / recency weight may be saturating or not scaling correctly.

Related

• KS74 temporal benchmark: 0/5 (recency boost dominates all queries)
• KS74 KU-strict: 3/5 failures from cross-entity recency contamination (e.g., "moved to SF" ranked KS67: Schema-driven fact extraction + Greptile AI review #1 for job queries)

Reproduction

cargo test -p shrimpk-memory recency_boost_newer_memory_ranks_higher

Expected

Newer memory should have a measurably higher final_score than an older memory with similar embedding similarity.