v0.11.0

Changed

• xstep step-recording JSON is slimmed (~30% smaller, no information loss): per-step location/function are dropped (read them from stack[0]), the shared breakpoint is emitted once at the top level instead of on every step, and variables is omitted on diff frames (recoverable from diff) (#86).
• xprofile time_ms/memory_mb are now real measurements parsed from the cachegrind summary: line (via the events: header) instead of heuristic estimates (#87).
• xprofile bottlenecks is now a structured array [{function, percentage, time_ms}] instead of preformatted strings, so it can be filtered/sorted programmatically (#87).
• Output format details now live in the JSON schemas; the xdebug skill and MCP tool descriptions defer to them instead of restating the shape (#86, #87).

Fixed

• xprofile per-function cost summed the line number instead of the Time column, making bottleneck percentages meaningless; costs are now time-based and the bottleneck time_ms values sum to the reported total (#87).

Removed

• Dead XdebugProfiler class and unused ProfileStatistics/ProfileAnalysis DTOs — never instantiated; both the CLI and MCP paths go through bin/xprofile (#88).