[galafis]

Sample and approximate entropy subtract template vectors in the original input dtype. Unsigned subtraction wraps and signed subtraction/abs can overflow, so integer sensor signals produce different template-match counts and entropy from their mathematical values. Boolean signals raise on subtraction.

Promote integer inputs before computing template distances. Use float64 when all integer values are within +/-2**52, so their pairwise differences are exact and the normal fixed-width path stays fast. Use Python integer arithmetic outside that range to preserve exact signed differences at int64/uint64 limits; the floating-input path and the existing tolerance calculation are unchanged.

Eleven regression cases cover int8/uint8/int16/int64/uint64 arrays and Series plus booleans. A separate Python-loop oracle counts template matches using arbitrary-precision integer subtraction. All eleven fail before the fix. With the fix, the new tests and entire existing feature-calculation module pass: 96 passed, 3 skipped. Black 22.12.0, isort 5.12.0 (Black profile) and git diff --check pass.

The change is separate from pending PRs1133/1134: those cover energy and time-adjacent differences, whereas these estimators compare all embedded template pairs. No benchmark or performance claim is made. Only 64-bit values outside +/-2**52 take the object-integer path; it favors exact comparisons and may cost more than the fixed-width path.