2026-08-27 - hardening the probe for hardware: the extent blind spot was half-closed
Asked to harden the active probe before the hardware run. The two
axes named - the size record (binary-search a size argument
to derive exact structure sizes) and the "wrote-0 vs untouched" extent
fix - were already built by concurrent work: the
poison-fill dump, obs_report_written with
changed/untouched runs,
obs_report_size, and check_query_size_ladder
all present, and both records documented in OUTPUT.md.
The residual: the dump was moved onto the poison basis and the
verdict counter beside it was left on the old non-zero
one. Observed live on the host - the same call reporting
extent 256 in its dump and pass|0x11 (17) in
its verdict. On hardware that mismatch fails the other way: a call
writing a real structure with a zeroed trailing field gets "the call
succeeded and wrote nothing", a false failure contradicting its own
dump. Fixed to count changed-from-poison; the call now reports
pass|0x100, matching. verify: ok, 24 gates.
(D204)
Surprise worth noting: the decision index had drifted (203 body
entries, stale index) from the concurrent selfish split and
D199-D203 - regenerating brought it current before D204 landed.