The bisection lever had never been connected
Went to use the stub policy on sceSysmoduleLoadModule -
the classic one-bit oracle, answer ok and see whether the guest proceeds
- and it did nothing. So ran the control:
default_return = "ok", everything answers success. Also
nothing. Identical imports, identical calls, identical fault, which is
impossible if every stub had changed its answer.
Two gaps (D166). The config file did not carry a
[policy] section at all, and - worse - the stub-return
table handed to the dispatcher was built only from the knowledge file's
declared return kind. The policy's default and overrides were consulted
nowhere between the file and the guest. Every override ever written was
silently ignored.
That is D082 one layer up, and it hid for the same reason: the
mechanism looks present from every angle except the one that matters.
orbistoun-cli policy prints an editable file,
StubPolicy has overrides, the service holds one - and none
of it reached the call path.
Caught by the control experiment rather than by reading the code, which is the lesson worth keeping: before believing a setting works, set it to something that must visibly break, and check that it does. Same method that found the entry-convention bug this morning.
Precedence now: explicit override, then the knowledge file's declared kind, then the policy default. The middle one beats the default deliberately - a blanket "answer ok" must not undo D125 and start returning error codes in pointer registers.
Measured working: default_return = "ok" moves PPSA28061
from 47/933 to 48/935. Small, but real where there was provably none.
And it says something about the current wall - blanket success does not
get past image+0x43c4, so that fault is not a stub
value.