Read missing, infinite and fractional-millisecond datetimes and durations

[Thenewmanator15]

A datetime array holding NaT, or a duration holding NaN or a fractional number of milliseconds, is not converted at all today: the conversion throws — DateTime of NaN, Millisecond(1.5), Millisecond(NaN) — and the variable comes back as its raw uint32 MCOS reference (UInt32[0xdd000000, 0x2, …]) instead of values. Fractional milliseconds are common in real durations (milliseconds(1.5), seconds(0.0025)), so any such array was affected, as was every datetime array with a gap.

With this change:

• NaT and NaN durations become missing, so an array holding one is Union{Missing,DateTime} / Union{Missing,Millisecond}; arrays without one keep their concrete element type;
• ±Inf datetimes (which DateTime cannot hold) are missing too;
• durations are rounded to whole milliseconds, as datetimes already are.

The comment asking whether a stored datetime's imaginary part is the sub-millisecond part now answers it: it is the low half of a double-double, below half an ulp of the real part, so real + imag == real in Float64 and nothing is lost at DateTime resolution.

Tests, and the same request as #253. test/missing_gen.m writes test/v7/missing_times.mat and test/v7.3/missing_times.mat: NaT in an array, alone, and with a time zone; ±Inf datetimes; NaN durations, in an array and alone; 1.5, 2.25 and 2.75 ms. I can't contribute the .mat files, so they are not included — could someone with MATLAB run the script in test/ and commit its two files? Until then the testset skips. Against the files the script writes (MATLAB R2025a), on master all seven variables come back as raw uint32 references in both formats; with this change the full suite passes (599). Without the files: 581 pass, 2 skipped.

Not in scope: writing arrays with missing datetimes or durations back to a file.

🤖 Generated with Claude Code