## Validate interval widths relatively when checking significant digits

What happened: the T618 readback checker initially compared a Sage MPFR
endpoint/diameter with an interval-field threshold, which raises a coercion
error. After matching the threshold's type, a single absolute width cutoff
still rejected valid 100-significant-digit companion values of order 10^3.
The checker now checks positive finite intervals and bounds diameter/lower
endpoint, using a scalar threshold of the same real-field type. A preflight
on every stored companion row and both cross-table identities passed.
Neither failure indicated a wrong stored number.

What the skill says now: verify finiteness and that the radius is small enough
to mean something before trusting interval overlap.

What it should say: for a significant-digit claim across varying magnitudes,
use a relative width check away from zero; reserve absolute widths for a
stated absolute-error goal. Compare endpoints and scalar thresholds in a
compatible ring. Near zero needs a separate absolute-error criterion.

Evidence: T618's readback_checks.py, preflight-readback.log, and the preserved
failed readback.log. This is proposed guidance for a person to evaluate with
a test, not an edit to the canonical skill or an independent critique.