E58: on a 7-joint arm (Pusher), the curves miss most faults, the step ledger catches every one, and no fault the curves miss leaves a worse policy

[tactino]

E54's 21 client faults and 2 log faults on Pusher-v5 at doses one, 0.01 and 1.0, 409,600 steps, seeds 0-2 traced, ledgered and evaluated on a clean environment, seeds 10-19 fault-free; 226 runs, every check and prediction holds (results/verdicts.txt).

dose	status rule	strict	byte identity	record	step ledger	harmed (clean)
one value	0	0	63	18	63	0
0.01	3	1	63	18	63	3
1.0	21	31	63	15	63	27

• Byte identity caught all 189 client-fault runs, final-obs included: every Pusher episode ends at the time limit, where SB3 reads the final observation.
• The ledger caught and located 189/189 with no false alarm in 27; ledger or record 207/207.
• Clean evaluation (E57's rule): 0 of 155 runs that passed both curve checks were harmed; every one of the 30 harmed runs had been caught by a curve check.
• With every reward zeroed the log read 0, Pusher's best possible return, while the policies scored -99 to -129 against a band of -31.8 ± 2.1.

The protocol and verdicts.py were committed (67db23d) before the registered runs.