v0.2.0

First release since 0.1.0 (February). Highlights:

Fixed

• Processes embedding Kairn now exit cleanly. The server closed no database connection on shutdown, so any process that ran a tool hung at interpreter exit on a non-daemon aiosqlite worker thread. A server lifespan now closes the store, and a failed store initialization cleans up after itself (#11).
• CLI paths with ~ resolve correctly in all 24 commands (#8).
• Day-granular as-of filtering in the flag-gated bi-temporal search path (#8).

Added

• kn_judge (5-verb relationship judgments), kn_doctor (read-only health diagnostics), kn_preference (utterance-time preference capture), kn_promote_pending - 22 MCP tools total.
• preference experience type with documented decay half-life.
• Namespace support and per-workspace isolation hardening.
• Bi-temporal schema (valid_from / valid_to) and rule-based cross-session entity keys, shipped as additive infrastructure (#9).

Docs

• Radical-honesty benchmark scorecard: the README leads with the full LongMemEval-S per-category numbers, worst cells included, each linking its diagnosis (#12). This scorecard stays current with every release that touches recall; this release does not touch the recall path.

Full test suite: 498 passed, exits cleanly.