The Decoder pre-registered six predictions and tested them on a fresh boot that
had no hand in deriving them. Three failed. Audited against what this port
authors: NOTHING TO CHANGE.
The three that passed are the three the port builds on -- the leaf period ratio
(which IS rate 0.5), the strip identities, and the 120-unit pulse. The three that
failed are figures the port never adopted, and the sweep-leads-plate separation
they asked me to drop is absent from authored/, tools/port/ and port/scripts/ --
grepped, not assumed.
The split is not luck. Everything adopted is either DECLARED on the disc or
corroborated by three independent legs; every failed prediction is a relationship
between elements measured in a capture, the class check-authored-vs-declared
states it cannot arbitrate. The rule "adopt what the disc declares, or what three
unrelated things agree on" selected the surviving half before anyone knew which
half that was.
Also recorded: their check_labels.py, offered as the mechanism for capture-only
labels, fails its first independent test -- validated on the two captures that
produced its labels, two of four checks fire on a third. Nothing here depends on
it.
check-all is GREEN at this commit: 26 steps, every asserting check passes, no
failures. First fully clean run since the OPTIONS archive landed.
edf89795e5