Only new spawns were held back after a failure, so every action already running
was allowed to finish even though its output would be discarded. The scheduler
now sends SIGTERM to the remaining children on the first failure, and reports
them separately rather than counting a process it killed as a failure of its
own. Partial outputs are removed and their cache entries dropped, so the next
run starts clean.
Under -k the existing behaviour is kept, since keeping going is the point.
On a graph with three six-second generators alongside one translation unit that
fails immediately, -j4 now returns in 0.12 s rather than 6.03 s.
A header included by fourteen translation units was hashed once per dependent,
and every built action hashed its inputs twice: once to decide it was stale and
again to store the new key. Digests are now memoised for the run, and the
scheduler drops the entries for an action's outputs and depfile once it
finishes, so a freshly written file is never read from the memo.
Checks each node's key at the moment it becomes ready rather than up front,
because an input may have been produced by a node that ran seconds earlier.
This is what gives early cutoff.
milner