Three times in one night a file that looked idle from outside was load-bearing to a process that had not announced itself: a compressor gzipping logs the corpus save was between enumerating and opening, a sweep within minutes of gzipping a live log into a valid truncated one, and a target/release deleted under a running benchmark. Care did not prevent any of them; every one used a reasonable proxy for idle. Freeze what the result depends on and fail; rebuild what can be reconstructed and survive. A guard that covers the input somebody thought of and not the one somebody changed is worse than none, because it gets cited as coverage. Also records the two coordination errors either side of a line: a correct git cherry count about the wrong branch, which only domain knowledge catches, and a conflict report that was a claim about patches rather than content, which a file-set intersection settles in seconds. Change-Id: I1ce160560a9ca464dfcffb8ec0f6a96cb7cde23b