Replopencode v1.17.14-17

A sandbox MCP server that fails to connect now logs why, not just that it failed. connectRemote and connectLocal already captured the reason into status.error; the "server unavailable" log site dropped it, so a connect failure reached Loki as status=failed with no exception attached. That is the common path — across 15 days of Replo production logs it accounted for 109 of 111 occurrences, against 2 for the tools/list path that does log its error. Observability only, no behaviour change. PR #51.