ben-manes / caffeine / #5728

100%

master: 100%

push

github

Record the depressed-window study of the norank sentinels

The two cells the audit sweep rowed as sentinels were worked one
question at a time; nothing changes in the machine. Recorded in
hill-climber.md section 7 (the dated entry), section 5 (the arms
killed) and section 8 (items 1 and 4), and in the gate skill's rows.

On flood_j100 the marginal signal is right (tail-band hits are zero at
every static window) but the parked arms cannot use it: marggate's
gate is on 47 of 48 steering samples through the SLRU geometry alone,
since a rotation larger than protected demotes every promoted key
before its reuse, and margrest's band is entered only at 57-76%.

A census of 600 trajectories (battery and corpus, seeds 1-8) finds a
park on a starved position 3 times in 642 audit confirms, so the row's
proposed rule fires there and loses one of them; the upper corner's
probe cadence costs 3.3pp of rep_r6's steady state and 0.5-3.8pp on
constructed cliff terrain while earning +0.2 on the two real cells it
runs on, so it stays. The census also measured three ways a found
position is lost (an audit's own crash releasing its park, a veto
return's arrival read as a shift, the fixed 5pp shift trigger
discarding the anchor on 10-34% of noisy cells' samples); the arms that
address them are trades with a mechanism on each side and are recorded
as such, as is the next-sample half of the wedge pricing.

census.py and corner.py join the skill as the instruments behind those
numbers, and harness.py's starvwrite edit is re-anchored on the new
starvationEnding (38 of 38 wired, strip round-trips to pristine).

4488 of 4500 branches covered (99.73%)

8922 of 8933 relevant lines covered (99.88%)

1.0 hits per line