v0.2.8: Bounded Recall Quality

v0.2.8 adds a deterministic, pluggable quality layer that limits weak recall evidence before it enters Agent context.

Added

Recall now supports strict-v1, balanced-v1, and exhaustive-v1 policies plus a registry for custom deterministic policies.
Providers explicitly declare whether their scores are normalized relevance values. Unknown or undeclared score scales remain supplemental evidence instead of being interpreted through unsafe absolute thresholds.
Recall sources expose structured quality statistics, including fetched, retained, selected, dropped, unscored, and unknown-scale counts.
Candidate expansion runs before filtering so strict selection can choose stronger evidence from a wider bounded pool.

strict-v1 is the default policy. Scores at or below zero and normalized scores below 0.25 are discarded before serialization to the Agent.
All high-confidence results (score >= 0.6) are eligible up to the requested limit. Medium-confidence results are capped at four and unscored or unknown-scale results are capped at two.
Recall no longer fills the requested limit with weaker evidence merely because capacity remains.
Thresholds, candidate expansion, medium-result budget, unknown-result budget, and policy selection are configurable under recallQuality.

Persisted memory data and Memory Space layouts are unchanged; no migration is required.
This release intentionally changes default recall selection. Set recallQuality.policy to balanced-v1 for broader low-score fallback or exhaustive-v1 to preserve the previous top-N behavior.
Custom policies are validated for bounded candidate limits, eligible selections, duplicates, and result limits; invalid custom behavior falls back to strict-v1.

The complete suite passes with 343 tests; one Windows-only smoke test is skipped only on non-Windows local runs. Ubuntu and Windows CI pass.
In an isolated 28-memory benchmark, answerable-query context decreased by 48.3% and absent-query context decreased by 60.4%.
The measured tradeoff was Recall@10 decreasing from 100% to 85.4%, while nDCG@10 changed from 0.948 to 0.923. Deployments that prioritize maximum coverage should use or fall back to a broader policy.

Update dsh-mnemon and restart the affected DSH profile. Review recallQuality if the deployment prioritizes maximum recall over bounded context.

See the v0.2.7 release notes for the previous patch release.