Two independent review rounds of coherent/incoherent propagation, and
their remedies. Every finding of both rounds is applied.
Behaviour changes a caller can see:
* A cross-polarization observable is REFUSED, not answered, when the
default coherence mode meets a flagged stack. FRONT_BLOCK's
intensities are exact totals over every incoherent echo, but its p-s
cross channel comes from the first coherent block alone, so Delta,
DOP, S2/S3, the raw cross terms and the retardance would be a ratio
of two different stacks. Pass COHERENCY_MATRIX, which cascades the
cross channel with the echoes. Nothing flagged: all modes agree bit
for bit and nothing is refused.
* Optical gain (Im(n) < 0) is refused at every door. It was silently
mangled before, and differently on each path: the coherent one
conjugated the phase back to decay, returning a POSITIVE absorptance
for a medium that amplifies.
* Synthesis refuses a `coherent: false` film instead of silently
solving the design as one coherent block. Multiblock synthesis
remains a phase of its own.
* A coherence flag is canonicalized rather than trusted. Any non-zero
value used to partition the stack while only 1 applied the
attenuation element, so a flag of 2 made a slab's absorption vanish
from the energy books.
* A flag on a half-space, or on a layer too thin to decohere, now
warns. Both remain warnings: the engine honours the array either
way, and the thin incoherent limit is a legitimate thing to model.
Verification, which is the point of the series: the incoherent cascade
had never been checked against anything but a port of itself. It is now
checked against a closed form, against bit-exact thickness independence
over three decades, and against the definition -- a flagged answer IS the
coherent answer averaged over one round-trip phase period. The same
average on the Stokes vector is what turned the cross-channel finding
from an inference into a measurement.
Also in this tag: the README documents the coherence contract, the
sixteen dispersion models and six EMA mixing rules, multi-environment
design and the guided-mode surface; and a sixth CI guard keeps a review
harness from joining the battery undocumented.
Full detail in CHANGELOG.md, sections 0.7.8 through 0.7.14.