v1.2.0 - Sub-sample verification of the latency correction

Integer alignment removes whole samples of renderer transport latency;
what it leaves is the fractional part of each renderer's group delay.
This release adds the code and the tables that establish what that
residual does to the scores, and what it is.

Five scripts under pipeline/src measure BINAQUAL's response to known
fractional offsets, build and re-score a stimulus set corrected by each
file's own residual, compare that re-score with the baseline, and check
both the phase-ramp operator and whether a constant delay describes the
real pairs at all. Section 12 of the analysis reads the four resulting
tables in data/ and emits every value quoted from them as a macro, with
assertions on the qualitative claims.

Fig 3 is relabelled so that every point is identifiable in the congested
low-score region, and Fig 8 is renamed from AnchorFamilyBias to
ReferenceAnchorBias, following the manuscript's move from "anchor family"
to "reference anchor".