Add predictions-vs-measured ΔE report mode #3

OLE-Toolset today measures a display and reports its fidelity against ideal
targets (PQ EOTF tracking, white-point stability, dI/chromatic error). The
color-wrangler
pipeline needs a second report mode: measured versus predicted, where the
predictions come from a generated OCIO config rather than a standard.

Inputs

Both artifacts already exist upstream:

• Predictions —
  ocio-display-gen
  writes a .predictions.yaml (schema ociodisplaygen/predictions/1) beside
  every generated config, plus the probe patch imagery it describes. For each
  patch: the scene-linear content, the code values the config emits, and the
  predicted CIE XYZ in cd/m². That work has shipped (§spec:verification there
  is complete).
• Measurements — produced by color-wrangler verify sessions, which drive
  the probe patches and read the instrument.

Output

A threshold-compliance report on the residuals. Thresholds per color-wrangler
§spec:verification-policy: ΔE2000 ≤ 2 average, ≤ 5 maximum, unity exponent.

Why here and not upstream

Neither generator nor session tool should grade its own output — a generator
that scored its own config would be marking its own homework, and a
disagreement between two prediction implementations would be
indistinguishable from a wall fault. This layer reports; humans judge from the
report.

Related

Three smaller items are currently tracked in color-wrangler's roadmap
(§road:validation-integration) for want of a roadmap here — SDR-gamma
measurement path (the analysis is PQ-native today), colorimeter support for
drift checks, and a SPEC.md/ROADMAP.md naming color-wrangler as upstream
context. Worth landing that governance alongside this, so the remaining work
lives in this repo.