v0.6.0

Added

• Search results are now returned as flat objects ({"score": 77.9, "path": "...", ...}) instead of [score, entry] tuples in MCP tool responses, improving agent ergonomics. Internal APIs retain tuple format for backward compatibility
• memory_search_global now accepts memory_dir as a convenience fallback — if provided, it is appended to memory_dirs, allowing single-directory search without wrapping in a list
• memory_state_from_note now includes a warnings field in JSON responses for cap-exceeded, auto-prune, auto-improve, and stale-item notifications

Changed

• Breaking (internal API): enforce_cap() now returns a (kept, dropped) tuple instead of just the kept list. All internal callers updated
• _auto_prune() now returns a list of dropped items instead of printing to stderr
• memory_state_from_note no longer emits warnings to stderr; all diagnostics (cap exceeded, auto-prune, auto-improve candidates, stale items) are included in the structured JSON response
• Quick mode (mode="quick") now strips None and empty-string ("") fields from each search result entry, further reducing context consumption
• Fallback search (triggered when index returns 0 results) now uses original query terms instead of expanded terms, reducing noise from CJK n-gram fragment matches