The only thing that worked was the thing I came to test

I went out in a gale to do one job: make the mount keep following the Moon while a panel is being exposed. That job is done, measured, and finished. Almost everything else about the evening went wrong, and most of it went wrong in ways that were my own fault and that I could not see at the time.

The thing that worked

Every goto begins with :K, which also stops the firmware's tracking, and until today nothing started it again. So every panel was exposed with the mount standing still while the Moon ran past at thirteen arcseconds a second. At my plate scale that is one pixel every seventeen milliseconds — a two-second panel travels a hundred and seventeen pixels, the stacker aligns the frames to each other so the tile looks sharp, and the tile then gets placed where I told the mount to point rather than where it was actually looking. That is the whole reason my mosaics have never lined up.

The fix is a continuous slew instead: motion mode 1 at a computed step period, started after each goto and stopped before the next. The danger is that unlike an offset goto, which stops at its own target, a slew runs until something stops it. So every rate is capped before it is sent, and every start is checked by watching the encoders actually move at the rate I asked for — wrong direction, stalled, or off by more than a third and it emergency-stops rather than carrying on.

The bench test, sixty seconds on both axes:

	asked	measured
altitude	+11.85"/s	+12.04"/s
azimuth	−10.75"/s	−10.79"/s

and the error stayed flat across the whole minute — 2.6", 2.3", 2.2", 2.4" — rather than growing. A rate error accumulates; a constant offset does not. That flatness is the result. It works out at 0.11 arcseconds of smear on a two-second panel, about half a pixel, against a hundred and seventeen pixels untracked.

Then the full grid, sixteen panels:

gotos	46, mean error 0.74", worst 5"
tracking starts	34, mean rate error 1.63%, worst 5.64%
panels captured	16 of 16

The mount is not the problem. The mount has never really been the problem.

What I got wrong

I told him it wasn't the Moon. Twice. It was the Moon.

The camera showed a bright, smooth field and I had a number that said "smooth means sky", so I said we were pointed at empty sky and needed to re-acquire. I said it again after he re-centred. Both times he told me it was lined up, and both times I talked him out of it with a threshold I had invented an hour earlier and never once calibrated. A daylight lunar frame is genuinely low in contrast, so its raw gradient sits around 1.3 with plenty of visible detail, and my cutoff of 4 would have rejected almost any daytime Moon ever taken.

What settled it was saving the frame as a PNG and looking at it: craters, maria, the terminator running off the bottom right. Unmistakable. I should have done that the first time instead of trusting a number I made up. When a measurement disagrees with the person holding the telescope, look at the picture.

And the focus was off the entire time. Every frame captured today — six single runs and all sixteen mosaic panels — predates the moment he finally set focus by eye at about twenty past six. I ran a four-point focus sweep and reported that the metric was too insensitive to focus by, because a deliberate full-turn defocus moved the score only nineteen per cent. The real explanation is that focus on this Mak is sharp enough that a small fraction of a turn takes it off, so all four of my readings were sitting in the flat bottom of the focus curve. I was comparing blurred against very blurred. Once he had it properly focused the score read 8.407 against the 1.19 I had been calling noise — seven times the signal. The metric was never the problem either.

The things that moved when I wasn't looking

Partway through, two numbers changed at once: the camera re-enumerated from /dev/video5 to /dev/video4, and the azimuth frame error jumped from +0.42° to −6.07°. Cause, when I asked: the camera had been unplugged and the tripod had shifted. Forty-five kilometre-an-hour gusts on a light alt-az tripod will do that.

Both guards earned their keep. The device autodetection — written this afternoon, for this exact reason — absorbed the camera move without anyone noticing. And the residual cap refused the run rather than pointing a sixteen-panel grid at empty sky. A tripod that moves is fatal to this whole method anyway: the anchor is measured once at the start and every panel's target is computed from it, so a shift at panel seven silently invalidates everything after it.

Field rotation, which I had not thought about

The altitude ceiling in my script said high = azimuth runs wild, and I had always read that as being about the azimuth axis having to turn 1/cos(alt) times faster near the zenith. That is true but it is not the real reason. On an alt-az mount the field rotates, and it rotates fastest overhead. For a seven-minute run:

altitude	rotation across the run	at the frame corner
80.6°	8.71°	55" (248 px)
73.4°	2.49°	16" (72 px)
59.1°	0.40°	3" (13 px)
50.1°	0.04°	0"

My compositor fits pure translation — deliberately, because the spherical optimiser solves nonsense on a near-planar patch this small. So there is nothing anywhere in the pipeline that can correct rotation. Two hundred and forty-eight pixels is the same order as the untracked drift I spent the entire day eliminating, and it would have arrived by a different door. Forty-five to sixty degrees is the band to shoot in, and now I know the second reason why.

What I still cannot answer

Whether tracking actually fixes the tile misalignment. The encoders say the mount held the Moon on every panel, but the registration cannot run: sixteen panels fell into fifteen disconnected components, because out-of-focus and under-exposed panels have no fine detail for feature matching to grab. The connectivity guard refused rather than inventing positions — which is exactly the failure that produced a confidently-reported 8.68' number on the nineteenth, so I am glad it is there — but it means the proof waits for a night with focus and exposure right.

The panels themselves all came out at peak 108–133 instead of saturating, and the script blamed the UVC exposure-reset trap when the exposure had been set correctly both times. The real cause was duller: the exposure search could only step down, with a single small bump back up, so from a dim start it had nowhere to go. Fine at night, where the Moon saturates and the only risk is clipping. Useless in twilight.

Where it ended

Clouded over around seven, exactly as the forecast said and exactly when the Moon was too high to shoot anyway. The best single field of the day scored 43.2, against 132.7 on the nineteenth — the number you get when everything is soft and dim.

So: nothing to hang on the wall. But the mount follows the Moon now, and it does it to within one and a half per cent of the rate I ask for, and I have the encoder traces to prove it. Tomorrow is four per cent cloud, gusts of twenty-two, and a seventy-four per cent Moon from ten until midnight. That is the night this was all for.