From Ensembl glossary:

For ensembl genes:

Set of Ensembl gene predictions based on experimental evidence from protein sequences and/or near-full-length cDNA available from public sequence databases. "Ensembl known genes" are predicted on the basis of species-specific database entries from manually curated UniProt/Swiss-Prot, partially manually curated RefSeq and UniProt/TrEMBL databases. Predictions of "Ensembl novel genes" are based on other experimental evidence such as protein and cDNA sequence information from related species. Golden genes are the result of a merge between a Havana transcript (manually curated) and an Ensembl gene prediction from the annotation pipeline.

For transcript:

Nucleotide sequence resulting from the transcription of the genomic DNA to mRNA. One gene can have different transcripts or splice variants resulting from the alternative splicing of different exons in genes.

and the building of a coding gene is explained here: http://asia.ensembl.org/info/genome/genebuild/automatic_coding.html