The run now says what to do about itself
Reframed by a good question: orbistoun's end goal is to be loud and clever about what it dumps, and eventually to propose its own fixes. That consumer is not a person reading a terminal, so the diagnostics built this session were right in spirit and wrong in shape - each built ad hoc for one bug, each rendered as prose (D179).
A run now produces findings: a routable kind, a subject, evidence from the trace, a suggested action, a weight. Ranked by confidence before weight, because a heavy guess must never outrank a light certainty.
Confidence is the field that matters. A confidently wrong suggestion
is worse than none - it gets acted on - and this project has produced
three of those already: an entry convention that looked right, a stub
policy that looked wired, a name sweep whose vocabulary could not hold
the answer. Nothing reports Certain unless the trace shows
it.
Unimplemented needed one new fact - whether a called
import has a handler behind it - because without it a trace cannot tell
"used this and it worked" from "used this and got a placeholder".
It reproduced a day of manual reading in one pass.
Pointed at the corpus it classified all three failure shapes found by
hand today: the spin on sceKernelDirectMemoryQuery, the
placeholder dereferenced by PPSA21564, and the deliberate abort in two
titles - with the error-reporting call before the abort surfaced as
evidence.
Also caught a formatting bug worth noting: cargo fmt
collapsed a line-continued string literal and baked the indentation into
it, so the rendered action had runs of spaces in the middle. Output
meant to be consumed cannot afford that.