Consensus outliers, and the runtime census

Reading the 24 outliers rather than only generating them found a defect in the table that generated them (D156). Ten were host alone says fail, which is a PC being a PC. Twelve were 900-surface/* with PS5PCEM standing alone, which is the stub-versus-honest split restated once per library. Two mattered, and one was that 900-surface/control fails on shadPS4 and fpPS4 - the check whose own message reads "every count in this section is meaningless" - while the compatibility table printed their census counts beside PS5PCEM's with nothing to separate them. Marked (void) now, not hidden: the 35,337 is itself the evidence of stub-everything behaviour, and the marker says only that it measures a loader rather than a platform.

The runtime census exists (D157), and one run settled the argument D149 could only make:

instrument, same shadPS4 run	reported
address census	373 of 373 symbols present
sceKernelLoadStartModule	0x80020002 for all five paths

0x8002_0002 is 0x8002_0000 \| errno, so ENOENT - no such module, five times, specifically. The loader says it has nothing while the census on the same run says it has everything, and only one of those came from the platform being asked.

Getting there needed two exclusions to reach the section at all, both long-known shadPS4 crashes - which is the exclusion walk working exactly as WORKFLOW.md describes it.