Release v2.7.0: harden retrieval broker

[RodCor]

What changed

• move whole-bundle abstention onto absolute cosine evidence, with per-model auto defaults for new projects
• arbitrate the evidence band with the cross-encoder through one shared CLI, MCP, and daemon path
• preserve learned usefulness and supersession policy after reranking instead of letting relevance-only rescoring erase it
• recognize correction-qualified rewrites while preserving historical queries and same-batch co-writes
• report reranker fallback and honor HF_HOME for user-defined Hugging Face rerankers
• bump the workspace and path dependencies to 2.7.0 and add release notes

Why

The normalized composite gave the best candidate a strong score even when the entire corpus was unrelated, so the old gate could not abstain. Cross-encoder reranking then recomputed scores from query relevance alone, erasing citations, regret, freshness, and pre-rerank supersession policy.

The new pipeline separates admission evidence from ranking policy: raw cosine and raw reranker scores decide whether evidence exists; policy is carried on capsules and applied after reranking.

User impact

New brains use the calibrated auto abstention floor. Existing project configs remain opt-in because an omitted key still deserializes to the disabled value. Lean and cross-model retrieval remain exempt from the cosine-only gate.

Validation

• comprehensive BrainBench, 264 scenarios: 91.3% -> 92.4%
• importance: 92.1% -> 97.4%
• retrieval: 56.6% -> 97.0%
• workflow slice, 150 scenarios: 79.0%
• update stream, 100 scenarios: 89.3%; newer fact wins 33/33 applicable cases
• cargo test -p kimetsu-brain --features embeddings --lib: 658 passed
• cargo check --workspace --all-targets
• cargo build --release -p kimetsu-cli --features embeddings
• release binary: kimetsu 2.7.0 (embeddings)

Two Docker bind-mount paraphrase cases regress because TinyBERT scores them below the ordinary topical retention floor. The release intentionally does not bypass that relevance floor for cited memories, which would reintroduce false injection.