HDMA line phase slips on one frame in four (SimCity scenario view)

[blackerking]

Repo: mstan/snesrecomp — observed at 251a966 (main) merged into a
downstream fork, i.e. with b8ef573 runtime: execute HDMA on LLE beam timeline.

Game: SimCity (U), scenario view screen (the isometric city on a desk).

Symptom

Reported from play as the map "jittering a little bit every couple of frames".

What it actually is

Frame-by-frame capture at SC_DUMP_INTERVAL=1 shows a strict period-4 cycle:

frame 62  ->  63   3671 px change
frame 63  ->  64   3671 px change
frame 64  ->  65      0
frame 65  ->  66      0

and frame 64 is pixel-identical to frame 62. So one frame in four renders
differently and the next returns exactly. It is not a moving object.

It is also not a shift. Testing frame 62 against 63 over the map region at
every whole-pixel offset, dx=0, dy=0 is the closest match (3529 differing px)
and every offset is worse:

dx=-1 dy=0  15050      dx=+1 dy=0  11360
dx= 0 dy=-1 17986      dx= 0 dy=+1 16718

The structure is the tell. Differing scanlines are:

43, 47, 51, 55, 59, 63, 67, 71, ... 199, 203, 207

every 4th line, 42 of them, evenly spaced across the map, never two
adjacent. A tear would be a contiguous band; a re-render would be dense. A
strict 4-line comb is an HDMA cadence.

hdmaActive on this screen is 0xe0 — channels 5, 6 and 7 live on every
frame, which is how the skewed map is drawn.

So: one frame in four, the per-line HDMA writes land on the wrong phase, and
every 4th scanline keeps the previous frame's value.

Not widescreen

Identical with the widescreen extension disabled (256 px output), so it is not
the margin path.

What I could not establish

Whether this predates b8ef573. I have no A/B, because the save-state format
diverged in the same range: states written by the newer build are refused by
the older one, and states written by the older build are refused by the newer
(RTL_SAV_VERSION_MIN raised to 6). A cold boot does not reach this screen
without menu input I could not script reliably.

Worth flagging separately: the OLD runner does not refuse a NEW-format state,
it silently mis-loads it — logic_changes=0, nmi_serviced=0, wedged at
cb:0d33. The version gate protects the new build from old files but not the
reverse, which is the same silent-corruption failure the gate was added to
prevent.

Reproduction

A save state on this screen reproduces it on current main (it is in the new
format, so it loads there). Available on request — this repo is ROM-free and
does not carry states, so it is not attached here.

SC_DUMP_INTERVAL=1 SC_DUMP_START=60 <runner> --load-state <state> --qualify 120

then diff consecutive frames; the period-4 cycle and the every-4th-scanline
comb are immediate.