v0.32.0

The seed writes a program, and the fuzzer learns to refuse

Three things.

The closed toolchain had one line in it that was not closed: everything after
the seed needed nothing installed, and the seed itself emitted intermediate
text. It writes octets now - the same back end again, in the language this
file is written in - and it was identical to the compiler written in GSL-2 on
the first run, for a tiny program, for the front end, and for the compiler
itself. So the report can say something it could not before: the seed's own
output is already the fixpoint, two implementations of one back end agreeing
about every octet of a compiler. The check moves to the job that installs
nothing.

The two front ends had only ever been compared on programs they accept. Now
one token of a generated program is broken and both are asked; a thousand of
those across five seeds, and they draw the edge of the language in the same
place. The one disagreement it found was in the test: a machine faulting
while running a program is not a front end refusing it.

And a finding is now as small as it can be made rather than as large as the
generator happened to make it, and is kept somewhere that outlives the run
that found it.

Full Changelog: v0.31.1...v0.32.0

LOC: 19124