@rolfdegen
The problem with using a 1.5v alkaline cell (or even a cr2032 with a fixed potential divider using M ohm resistors) is their own tempco. I know this because these were two of the first things I'd tried when checking the excellent tempco specification of the SDM3065 I'd purchased way back in October 2021 after allowing it to clock up a couple of thousand hours to age its internal reference.
In the end, I purchased a cheap AD584LH, fully aware that it too had a lousy (by voltnut standards) tempco. This can be stable enough given a constant room temperature. I never did get around to placing it in another constant temperature room whilst using my hobby room as a "variable temperature environmental test chamber"

so this testing regime was rather flawed.
However, the one thing I did discover was that this volt reference's tempco virtually cancelled the tempco of the cheap Owon XDM1041 I'd purchased a few months later as a convenient alternative to a pair of cheap (but surprisingly good) 9999 counts Mestek DM91A hand held DMMs.
The Owon bench DMM offered an improvement over the hand helds' tempco performance, effectively sitting between that of the hand helds and the SDM3065. Not a bad trade off considering its order of magnitude reduction in power consumption over that of the SDM3065's 13W (and, incidentally, that of an SDL1020X-E). As a consequence, I've had it powered up continuously over the past couple of years, monitoring the EFC voltage of my MK II (G3RUH inspired design) GPSDO.
With all that being said. I started using the 4.97844v, from a well aged 7805 fixed to the heat spreader inside my temperature stabilised (to within +/- 5mK of a 36.05 deg C set point) LPRO 101 based RFS used solely to feed a 5K 10 turn calibration pot, as a long term temperature stabilised reference.
Checking my notes, I see I got readings of 4.98005 on 2022-03-27 followed by 4.97897v on 2022-09-16. To be honest, the 4.97844v reading was taken just now. However, ISTR my more recent, unrecorded, checks all showing a value of 4.978xx volts over the past 12 months or so suggesting a long term stability between my temperature stabilised 4.97844 voltage reference and the SDM3065 of just 0.01%. The only thing I can be reasonably sure of is that the SDM3065 more than meets its claimed tempco performance over its specified ambient temperature range of 18 to 28 deg C which was the sole reason in the first place for my purchasing it over its cheaper 5 1/2 digit cousin. I'd thought at that time that the extra digit was "overkill" for my perceived needs.
In hindsight, I was proved ever so wrong in this assumption when I discovered that I could now detect lamp volt variations in tens of micro-volts due to variations of ambient temperature altering the internal thermal gradients within my well insulated LPRO 101. If I had decided "to cheap out" on the SDM3055, I wouldn't have been able to detect this effect or, at best, be left guessing about this small deficiency in the efficacy of my temperature stabilisation measures.
As for the fixed speed cooling fan in the SDM3065, that, in view of its excellent tempco performance, seems an obvious measure to curtail the "FY6600 effect" (ie wild internal temperature excursions due to a virtual absence of any cooling airflow) and flatten any internal temperature gradients.
The innards are going to track some 2 or 3 degrees above ambient reducing demands on the internal temperature compensation requirements (as well as minimising the variation between a cold power up and a stabilised calibration reached within just 5 minutes or so).
[EDIT 2024-05-02] I just noticed at around 1pm when I transferred the meter connection to the Vref/calibrate pot wiper volt stereo jack socket that the Vref was still showing yesterday's 4.97844v readings. This was after the meter had been powered up for just over an hour (I normally monitor the lamp voltage via the lamp/Xtal stereo jack - it's less boring than monitoring the Vref).
Out of idle curiosity, I switched the meter off for about twenty minutes so I could check the cold to warmed up voltage reading drift. The initial reading was 4.97839v rapidly increasing to 4.97842v about half a minute later taking another 3 minutes to show 4.97843v / 4.97844v finally settling back on the original 4.97844v some 4 minutes after that.
This test merely confirmed my gut feeling that it only required a mercifully short 10 to 15 minutes of warm up time from cold power up to produce stable readings. Indeed, if the equivalent of 5.5 digit accuracy is all you initially require, then it easily exceeds this specification within half a minute of power up.

A quieter replacement fan
is likely to might prove to have a lower CFM rating but you can compensate for this if you're willing to modify the case to remove the restriction of the colander of small holes on each side with a suitably fan sized hole saw and fitting wire finger guards. This measure will reduce turbulence noise created by the colander of ventilation holes, increasing flow rate enough to allow a dropper resistor to be added with little to no penalty on the original fan's cooling efficacy.
Obviously, this is a modification best left to the end of the three year warranty period (if you can hold out that long) so the best option whilst there's still some years of warranty cover at stake is simply replace the fan with a quieter, more efficient one. You can still opt to modify the case once the warranty has finally expired if so desired in order to maximise the benefit of your investment in the original replacement fan upgrade.