Result parser / error signature matching

Work out pass/fail from test output and check it's the same failure
across bisection steps.

This matters because you can hit a different bug mid-bisect and end
up chasing the wrong thing. Logspec already generates SHA1 error
signatures — we should use those to confirm we're still seeing the
original regression at each step.

Also need to handle non-monotonic pass/fail along the history, which
is more common than you'd hope.

Start by locating the result parser and the code that processes bisection steps, then trace how Logspec's SHA1 error signatures are produced or exposed. Done means pass/fail is derived from test output, the original failure signature is checked at each step, and non-monotonic histories are handled without chasing a different bug.

Written by the indexing model from the issue text.