v1.5.0 — MCP sampling for semantic-rename suggestions

Added

• MCP sampling for suggest_semantic_names -- when the connected client advertises the sampling capability, the server now calls back through the host LLM via ctx.sample() to pre-fill rename suggestions for cryptic identifiers, returning a suggestions dict alongside the cryptic-name lists in a single tool call.
  • Gated on a new ENABLE_SAMPLING env flag (default true). Set to false to force the legacy manual-review path everywhere.
  • Clients without sampling support (e.g. Claude Desktop) silently fall back to the legacy response shape — no breaking change.
  • Sampling requests, results, and failures are logged with elapsed time and item counts for observability.

Fixed

• MCP session crash on client disconnect during suggest_semantic_names -- a notification (ctx.info) write that hit anyio.ClosedResourceError because the client had already closed the streamable-HTTP session was caught by the handler's outer except and triggered a second doomed write, bringing down the entire FastMCP TaskGroup. Notifications are now sent through safe_ctx_info (failures swallowed at debug level), and ClosedResourceError-class disconnects re-raise cleanly so the framework tears the session down instead of writing into a dead transport.
• Sampling response parsing -- replaced pydantic-ai's result_type=Dict[str, str] (which forces the model to call an injected final_response tool, fragile on large responses) with explicit JSON parsing that accepts bare JSON, ```json fences, and prose-embedded JSON.