Same binary (md5 identical), different run: Eng1/Eng2/Comm read
17981/17981/15656 on some runs and 19011/19011/16909 on others. The shipped
binary is STABLE across three runs, so this is ours. Eleven consecutive
clean runs is what let it hide, and the dirty numbers are exactly the
pre-gate numbers from 5.3.32 -- the same artifact recurring.
Read off the raw words rather than inferred: in the artifact rect the shipped
binary writes ONE masked plane bit per pixel (0x4000 = Heat, 0x002D = sec),
while we write arbitrary 16-bit words (0x8BD6 / 0x8AD6 / 0x8AC0) that light
sec+overlay+Mfd1+Eng1+Eng2+Comm at once. Raw values going into a
plane-packed buffer -- which is why one bad draw appears as extra pixels on
six heads and as 483 MISSING on Heat at the same time.
Intermittent because these widgets are change-driven: on runs where the
coolant value never moves after the first paint, the bad draw never happens.
The six fixes of 5.3.33-39 stand (each moved specific attributable pixels).
What must be restated is the claim that four heads are exact -- true only
when this defect does not fire. The rig now needs a run-variance check
before any per-head number is believed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CydandClaude Fable 5
c8350408fc