Each mRNA is probably present in the sample in multiple copies.  The fragmentation is pseudo-random, so the fragments that came from different copies of the same mRNA will probably be broken at different positions.  

The principle at work is that it may be possible to identify overlaps, where the 5' end of one fragment (from copy A of the mRNA in question) is the same as the 3' end of another fragment (from copy B of the mRNA), and so a 'contig' or contiguous sequence of the mRNA can be built up by 'tiling' these pieces together.

Each mRNA is probably present in the sample in multiple copies.  The fragmentation is pseudo-random, so the fragments that came from different copies of the same mRNA will probably be broken at different positions.  

The principle at work is that it may be possible to identify overlaps, where the 5' end of one fragment (from copy A of the mRNA in question) is the same as the 3' end of another fragment (from copy B of the mRNA), and so a 'contig' or contiguous sequence of the mRNA can be built up by 'tiling' these pieces together.

A transcript is expressed if it has been assembled from our data set, but some transcripts may have artificially low expression values simply because they are incompletely assembled and do not recruit both pairs of PE reads in order to be properly accounted for during abundance estimation. If we assume that most relevant transcripts are well assembled and well quantified by the abundance estimation method used, we might infer the approximate number of expressed genes or transcripts as the number that are expressed above some minimum expression threshold.

I would recommend to use a genetically engineered reverse transcriptase, or an RT kit having RNase H activity. The RNase H will destroy the original RNA template at the end stage or the reverse transcription, where the RNA as a template and the DNA as the newly formed product are still hybridized. 

The yield of cDNA is less with these kind of enzymes, but the produced cDNA population is going to be the exact match of the mRNA population.

For qPCR studies I regularly use the Transcriptor First Strand cDNA Synthesis Kit from Roche Life Sciences. It has RNA-directed DNA polymerase activity, unwinding activity, and RNase H activity that degrades RNA in RNA:DNA hybrids.  

Good luck.

It depends a bit on your library prep.  The introduction of PCR into the process, whether during the library or emulsion steps (on some platforms), imparts an inherent sequence bias.  This is why it's a tough prospect to show RNA-Seq as a quantitative method for sequence prevalence. 

"If we lost some fragment in the transcript" - not sure what you're asking here.  Are you talking about detecting alternative splicing events?