Preserve sub-second precision when subtracting a DateTime from a Time - #58213

Summary

Time - DateTime (ActiveSupport layers this on top of Time#-) converted both sides via Float (to_f). A Float cannot represent every microsecond exactly, so the resulting difference was wrong for roughly half of all sub-second DateTime values.

Before / After

t  = Time.utc(2000, 1, 1, 0, 0, 1)
dt = DateTime.civil(2000, 1, 1, 0, 0, Rational(1, 1_000_000), "+0") # .000001s
t - dt
# before => 0.9999990463256836
# after  => (999999/1000000)

dt = DateTime.civil(2000, 1, 1, 0, 0, Rational(123_457, 1_000_000), "+0")
t - dt
# before => 0.8765430450439453
# after  => (876543/1000000)

Whole-second differences are unchanged (Time.utc(2000, 1, 2) - DateTime.civil(2000, 1, 1) is still 86400).

Why this is correct

• It brings Time - DateTime in line with the sibling paths that already use exact rational timestamps: Time.at(DateTime) (Preserve sub-second precision when coercing a DateTime in Time.at #58085) and ActiveSupport::TimeWithZone - DateTime (via getutc).
• The exact fractional second on a DateTime is datetime.to_i + datetime.sec_fraction; subtracting that from time.to_r preserves every microsecond.
• Existing whole-second coverage continues to pass; new assertions lock the sub-second contract.