v0.6.1

Fixes to the mandatory-first-call path and identifier retrieval.

Fixed

• memory_orient is bounded in every detail mode (#254, PR #260). It is the mandatory first call of every agent session, and compact — the mode documented as token-sensitive — had no default dashboard cap, so a mature estate could overflow MCP client tool-output limits. Three of four independent test agents hit the failure on their first call. Independently re-measured on a synthetic 460-namespace estate: compact 8,997 B → 2,717 B, standard 36,552 B → 9,406 B. Nothing is hidden silently — dashboard_meta.truncated_groups and namespaces_meta.truncated still describe the full estate, and both caps remain caller-overridable.
• A uniquely-named entry is no longer buried by structural ranking bonuses (#252, PR #261). An identifier lookup keeps its answer first, gated tightly: a non-orientation query of at most three distinctive terms where exactly one candidate contains all of them. Replaying 258 real production queries showed the gate firing zero times, so ordinary topical ranking is unchanged. The broader issue — that structural score can outweigh relevance on multi-term queries — remains open in #252.
• memory_insights.positive_outcome_rate is a rate again (#255). It divided positive-outcome count by retrieval events, exceeding 1.0 (3.457 observed in production).
• memory_history can filter for patch (#255). Partial writes are audited as patch, but the action enum offered only update.
• memory_query reports when it clamps an over-limit request (#255). A limit above 50 was silently clamped, making a truncated page indistinguishable from an exhausted result set.

Every change in this release was independently reviewed with executed verification — regressions mutation-tested to confirm they fail without the fix.